A pdf version of this note is available.

Contents

1 Introduction

The Table Query Language (TaQL, rhymes with bagel (though some people pronounce it as tackle)) is a language for querying and manipulating data in Casacore tables. It makes it possible to get results or select rows from an arbitrary table based on the contents of its columns and keywords. It supports arbitrary complex expressions including units, extended regular expressions, and many functions. User defined functions written in C++ or Python are supported. TaQL also makes sorting and column selection possible. Furthermore TaQL has commands to update, add or delete rows and columns in a table and to create a new table.

The first sections of this document explain the syntax and show the options. The last sections give several examples and show the interface to TaQL using Python or C++. The Python interface makes it possible to embed Python variables and expressions in a TaQL command.

1.1 TaQL vs SQL

TaQL is modeled after SQL and contains a subset of SQL’s functionality. Some familiarity with SQL makes it easier to understand the TaQL syntax. The most important features of TaQL different from SQL are:

• The result of a SELECT is another table (either temporary or persistent). Usually this is a so-called reference table, but it is also possible to make a deep copy and create a plain table.
  A reference table is a table that can be used as any other table, but does not contain data. Instead it contains references to the rows and columns in the original table. Thus modifying data in a reference table means that effectively the data in the original table are modified.
• A very rich set of mathematical and other functions.
• Any operand can be a scalar or an N-dimensional array. Many reduce functions can be applied to arrays.
• Arrays can optionally be masked.
• Full support of units and automatic conversion of units.
• Support of various types of patterns/regular expressions and support of maximum string distance (Levensthein (aka Edit) distance).
• Specific operators and functions for cone searching (i.e., spatial searching with a search radius).
• An advanced way of specifying intervals.
• No support of indices, thus a linear table search is done. Because data are stored column-wise, a linear search is usually very fast, even for very large tables.
• Limited support for joins (only implicit joins on row number).
• Many aggregate functions.
• The COUNT command exists to count the occurrences of column values. Although it can still be used, this command is obsolete now GROUPBY is fully supported.
• The CALC command exists to calculate an arbitrary expression (including subqueries) on a table. This can be useful to derive values from a table (e.g., the number of flags set in a measurement set). It can even be used as a desk calculator.
• TaQL can be used from languages with different conventions, for example the order of array axes. Therefore it is possible to set the language style to be used.
• The language can be extended by means of User Defined Functions, possibly implemented in Python. Some standard UDFs exist to deal with MeasurementSets and to do measure conversions (for directions, epochs, positions, and stokes).

TaQL has a keyword that makes it possible to time the various parts of a TaQL command.

2 TaQL Commands

2.1 Command summary

TaQL contains several commands. In the commands shown below the square brackets are not part of the syntax, but indicate the optional parts of the commands.

• show (or help)
    SHOW [type ...]

  can be used to give some TaQL explanation or to show table information. A sole show command shows the possible options. HELP is a synonym for SHOW.

• selection
    SELECT [[DISTINCT] expression_list]  
      [INTO table [AS options]]  
      [FROM table_list]  
      [WHERE expression]  
      [GROUPBY expression_list]  
      [HAVING expression]  
      [ORDERBY [DISTINCT] sort_list]  
      [LIMIT expression] [OFFSET expression]  
      [GIVING table [AS options] | set]  
      [DMINFO datamanagers]

  It can be used to get an optionally sorted subset from a table. It can also be used to do a subquery (see section 4.11 for more information on subqueries).

• updating
    UPDATE table_list SET update_list [FROM table_list]  
      [WHERE ...] [ORDERBY ...] [LIMIT ...] [OFFSET ...]

  It can be used to update data in (a subset of) the first table in the first table list.

• addition
    INSERT INTO table_list SET column=expr, column=expr, ...  
  or  
    INSERT INTO table_list [(column_list)] VALUES (expr_list)  
  or  
    INSERT INTO table_list [(column_list)] SELECT_command

  It can be used to add and fill new rows in the first table in the table list.

• deletion
    DELETE FROM table_list  
      [WHERE ...] [ORDERBY ...] [LIMIT ...] [OFFSET ...]

  It can be used to delete some or all rows from the first table in the table list.

• counting
    COUNT [column_list] FROM table_list [WHERE ...]

  It can be used to count occurrences of column values. Although the command can still be used, it is basically obsolete because the same (and more) can be achieved with the GROUPBY clause and aggregate functions in the SELECT command.
  Furthermore, usually GROUPBY is faster.

• calculation
    CALC expression [FROM table_list]

  It can be used to calculate an expression, in which columns in a table can be used.

• table creation
    CREATE TABLE table [AS options]  
      [column_spec]  
      [LIMIT ...]  
      [DMINFO datamanagers]

  It can be used to create a new table with the given columns and number of rows. Optionally specific table and data manager info can be given.

• table structure modification
    ALTER TABLE table  
      [ADD COLUMN [column_spec] [DMINFO datamanagers]  
      [RENAME COLUMN column_pair_list]  
      [DROP COLUMN column_list]  
      [SET KEYWORD key=value, key=value, ...]  
      [COPY KEYWORD key=other, key=other, ...]  
      [RENAME KEYWORD keyword_pair_list]  
      [DROP KEYWORD keyword_list]  
      [ADD ROW nrow]

  It can be used to add, rename, and remove columns and keywords and to add rows. Multiple such subcommands can be given, separated by white space.

The commands and verbs in the commands are case-insensitive, but case is important in string values and in names of columns and keywords. Whitespace (blanks and tabs) can be used at will.
Section 7 (Modifying a table) explains the UPDATE, INSERT, and DELETE commands in more detail. The CREATE TABLE command is explained in section 8 (Creating a table). The ALTER TABLE command is explained in section 9 (Modifying the table structure).
Section 10 (Counting in a table) explains the COUNT command in more detail.
Section 11 (Calculations on a table) explains the CALC command in more detail.

2.2 Using a style

TaQL can be used from different languages, in particular Python and Glish. Each has its own conventions breaking down into three important categories:

• 0-based or 1-based indexing.
• Fortran-order or C-order of arrays.
• Inclusive or exclusive end in start:end ranges.

The user can set the style (convention) to be used by preceding a TaQL statement with

The possible (case-independent) values are:

• BASE0 or BASE1 telling the indexing style.
• ENDEXCL or ENDINCL telling the range style.
• CORDER or FORTRANORDER telling the array style.
• PYTHON which is equivalent to BASE0,ENDEXCL,CORDER
• GLISH which is equivalent to BASE1,ENDINCL,FORTRANORDER

The following values are also possible and are described in the next subsections.

• synonym=libname to define a synonym for a user defined library.
• TRACE or NOTRACE to (un)set tracing.
• TIME or NOTIME to (un)set timing.

If multiple values are given for a category, the last one will be used. The default style used is GLISH, which is the way TaQL always worked before this feature was introduced.

It is important to note that the interpretation of the axes numbers depends on the style being used. e.g., when using glish style, axes numbers are 1-based and in Fortran order, thus axis 1 is the most rapidly varying axis. When using python style, axis 0 is the most slowly varying axis.
Casacore arrays are in Fortran order, but TaQL maps it to the style being used. Thus when using python style, the axes will be reversed (data will not be transposed). Note: unless said differently, all examples in this document are done using the Python style.

The style feature has to be used with care. A given TaQL statement will behave differently if used with another style.

2.2.1 UDF library synonyms

The style clause can also be used to define synonyms for the library names of user defined functions. For example:

defines the synonym mscal. Synonyms make it easier (i.e., less typing) to specify user defined functions.
Note that the synonym in the example above is automatically defined by TaQL as well as the synonym py for pytaql.

2.2.2 Tracing

It is possible to get tracing output during the execution of a TaQL command by using the case-insensitive value TRACE in the using style command.

2.2.3 Timing

It is possible to time a TaQL command by using the case-insensitive value TIME in the using style command. For historical reasons it is also possible to use the the case-insensitive keyword TIME before or after the optional style command.

Timing shows the total execution time and the times needed for various parts of the TaQL command on stdout. For example:

shows the time to do the where part (i.e., row selection on FLAG), projection (selection of columns), and distinct (unique column values).

2.3 Reserved words

TaQL uses the following words as part of its language.

These words are reserved. Note that the words in the TaQL vocabulary are case insensitive, thus the lowercase (or any mixed case) versions are also reserved.

The reserved words cannot directly be used as column name, keyword name, or unit. However, a reserved word can be used that way by escaping it with a backslash like \AS. When reading further, the meaning of

might become clear. It means: use unit in (inch) for column IN and test if it is in the given set.
Note this is unlike SQL where quotes have to be used to use a reserved word as a column name.

3 Selection from a table

The SELECT is the main TaQL command. It can be used to select a subset of rows and/or columns from a table and to generate new columns based on expressions.

As explained above, the result of a selection is usually a reference table. This table can be used as any other table, thus it is possible to do another selection on it or to update it (which updates the underlying original table). It is, however, not possible to insert rows in a reference table or to delete rows from it.

If the select column list contains expressions, it is not possible to generate a reference table. Instead a normal plain table is generated (which can take some time if it contains large data arrays). It should be clear that updating such a table does not update the original table.

The FROM clause can be omitted from the select. In that case no columns can be used in the selection, but functions like rand and rowid make variable output possible. Clauses like ORDERBY can be given. The GIVING (or INTO) might be useful to store the result in a table.

There is no explicit JOIN clause, but it is possible to equi-join tables on row number. Such tables must have the same number of rows. One can also join , for example, the main table of a MeasurementSet with a subtable like the ANTENNA table using a subquery. Joins are explained further in section 6.1.

3.1 SELECT command overview

The SELECT command consists of various clauses of which most are optional. The full command looks as follows where the optional parts are shown in square brackets.

The clauses are executed in a somewhat different order.

1. FROM to define the tables to be used.
2. WHERE to select the rows.
3. GROUPBY to group selected rows.
4. SELECT to fill select columns used in HAVING or ORDERBY
5. HAVING to select groups.
6. SELECT to fill the remaining select columns.
7. ORDERBY to sort the result.
8. LIMIT (or TOP) and OFFSET to ignore entries in the sorted result.
9. DMINFO to define the data managers to be used if the result is
10. GIVING/INTO to store the final result. stored in a plain table. See section 8.2 how to specify them.

All clauses are explained in full detail in the subsequent sections.

3.1.1 Column/keyword lookup

Expressions in the various clauses will normally use column names to select, sort, or group a table. It is also possible to use table keywords or column keywords by giving their names. Furthermore, it is possible to use a column, created in the SELECT clause, in the HAVING and ORDERBY clauses. This can save time in both specifying and executing the command, because a possibly complicated expression can be used to create such a column. If such columns are used, that part of the SELECT is executed before HAVING.

TaQL uses the following lookup scheme for column/keyword names.

1. If preceded by a shorthand (like in t0.DATA), the name is looked up in the corresponding table.
2. If not preceded by a shorthand, a name is first looked up in the select columns. If not found, the name is looked up in the first table given in FROM.
3. A name is first looked up as a column in a table. If not found, it is looked up as a table keyword.

See the discussion of column names and keyword names for more details.

3.2 SELECT column_list

Columns to be selected can be given as a comma-separated list with names of columns that have to be selected from the tables in the table_list (see below). If no column_list is given, all columns of the first table will be selected. It results in a so-called reference table. Optionally a selected column can be given another name in the reference table using AS name (where AS is optional). For example:

It is possible to precede a column name with a table shorthand indicating with table in the FROM clause has to be used. If not given, a column will be looked up in the first table. Note that if equally named columns from different tables are used, one has to get a new name, otherwise a ’duplicate name’ error will occur. For example:

Apart from giving exact column names, it is also possible to use wildcards by means of a UNIX filename-like pattern (like p/pattern/) or a regular expression (like f/regex/ for a full match or m/regex/ for a partial match). They can be suffixed with an i indicating case-insensitive matching. See section 4.3.6 for a discussion of these constants. The operator ~ needs to be given before the pattern or regex to indicate that columns have to be included. Thereafter operator !~ can be used with another pattern or regex to remove columns. Such an excluding pattern or regex only removes columns from the wildcarded columns before it until the latest non-wildcarded column.
A special pattern is * (which is the same as  p/*/). For example:

selects all columns except the ones ending in _DATA.

selects columns with a name containing DATA except the ones ending in _DATA.

does select the CORRECTED_DATA column.
Note it is not possible to change the name or data type of wildcarded columns.

It is also possible to use expressions in the column list to create new columns based on the contents of other columns. When doing this, the resulting table is a plain table (because a reference table cannot contain expressions). The new column can be given a name by giving AS name after the expression (where AS is optional). If no name is given, a unique name like Col_1 is constructed. After the name a data type string can be given for the new column. If no data type is given, the expression data type is used.

Note that unit conversion can be (part of) an expression. For example:

to store the time in unit d (days). Units are discussed in section 4.9.

It is possible to change the data type of a column by specifying a data type (see below) after the new column name. Giving a data type (even if the same as the existing one) counts as an expression, thus results in the generation of a plain table. For example:

Note that for subqueries the GIVING clause offers a better (faster) way of specifying a result expression. It also makes it possible to use intervals.

Special aggregate functions (e.g., gmin) exist to calculate an aggregated value (minimum in this example) per group of rows where the grouping is defined by the GROUPBY clause. The entire column is a single group if no GROUPBY is given. Aggregation is discussed in more detail in section 5.

If a column_list is given and if all columns (and/or expressions) are scalars, the column_list can be preceded by the word DISTINCT. It means that the result is made unique by removing the rows with duplicate values in the columns of the column_list. Instead of DISTINCT the synonym NODUPLICATES or UNIQUE can also be used. To find duplicate values, some temporary sorting is done, but the original order of the remaining rows is not changed.
Note that support of this keyword is mainly done for SQL compliance. The same (and more) can be achieved with the DISTINCT keyword in the ORDERBY clause with the difference that ORDERBY DISTINCT will change the order.
For full SQL compliance it is also possible to give the keyword ALL which is the opposite of DISTINCT, thus all values are returned. This is the default. Because there is an ambiguity between the keyword ALL and function ALL, the first element of the column list cannot be an expression starting with a parenthesis if the keyword ALL is used.

3.2.1 Masked array in column_list

If an expression in the column_list is a masked array, it is possible to create two columns from it: one for the data, one for the mask. This can be done by combining them in parentheses like (DATA,MASK). A possible data type given after the column names only applies to the data column, since the mask column always has data type Bool. For example:

The select results in a masked array containing the means along axis 0. Both column MD and MM are filled with the contents of the masked array. MD (with data type C4) contains the means over the first axis of the unmasked elements; MM contains the resulting mask.

[ AS options]]

3.3 INTO [table] [AS options]

This indicates that the ultimate result of the SELECT command should be written to a table (with the given name). This table can be a reference table, a plain table, or a memory table.

The table argument gives the name of the resulting table. It can be omitted if a memory table is created.

The options argument is optional and can be a single value or a list, enclosed in square brackets, consisting of values and key=value. They can be used to specify the table and storage type. All keys and values are case-insensitive.

TYPE=’value’

specifies the table type.
PLAIN = make a persistent table, thus a true copy of all selected rows/columns.
SCRATCH = as plain, but only as a temporary table.
MEMORY = as plain, but keep everything in memory.
If TYPE is not given, a reference table is made if no expressions are given in the SELECT clause, otherwise a plain table is made.

ENDIAN=’value’

specifies the endianness
BIG = big endian
LITTLE = little endian
LOCAL = native endianness of the machine being used
AIPSRC = as defined in the .casarc file (which usually defaults to LOCAL)
If ENDIAN is not given, it defaults to AIPSRC.

STORAGE=’value’

specifies the storage type
SEPFILE = store as separate files (the old Casacore table format)
MULTIFILE = combine all storage manager files into a single file.
MULTIHDF5 = as MULTIFILE, but use an HDF5 file instead of a regular file.
DEFAULT = use SEPFILE (but might change in a future Casacore version),
AIPSRC = as defined in the .casarc file (which usually defaults to DEFAULT)
If STORAGE is not given, it defaults to AIPSRC.

BLOCKSIZE=n

specifies the blocksize to use for MULTIFILE or MULTIHDF5.

OVERWRITE=F

tells that an existing table with the given name should not be overwritten. By default TaQL will overwrite existing tables.

For backward compatibility, it is possible to specify an option directly without having to use ’key=value’.

MEMORY

to store the result in a memory table.

SCRATCH

to store the result in a scratch table, possibly on disk.

PLAIN

to store the result in a plain table.

PLAIN_BIG

to store the result in a plain table in big-endian format.

PLAIN_LITTLE

to store the result in a plain table in little-endian format.

PLAIN_LOCAL

to store the result in a plain table in native endian format.

The standard TaQL way to define the output table is the GIVING clause. INTO is available for SQL compliance.

If the INTO (or GIVING) clause is not given, the query result will be written into a memory table. In this way queries done in a readonly directory will not fail if a result table cannot created. However, if the result is expected to not fit in memory (which will seldomly be the case), type SCRATCH should be used to make it fit.

If the result is stored in a plain table, it is possible to give detailed data manager info for the result table using the DMINFO clause. See section 8.2 how the data manager info can be specified.

3.4 FROM table_list

The FROM part defines the tables used in the query. It is a comma-separated list of tables, each followed by an optional shorthand (alias).

The full syntax is:

Similar to SQL and OQL the shorthand can also be given using AS or IN. E.g.

Note that if using IN, the shorthand has to precede the table name. It can be seen as an iterator variable.

The shorthand can be used in the query to qualify the table to be used for a column, for example t0.DATA. The first table in the list is the primary table which will be used if a column is not qualified by a shorthand. Often a query uses a single table in which case a shorthand is not needed. Multiple tables require a shorthand and are useful when:

• A keyword in another table is needed.
• Columns from multiple tables are used (an implicit join). In such a case the tables must have the same number of rows. For example, a regression test could be done like:
    SELECT FROM test.MS t1, result.MS t2  
      WHERE not all(near(t1.DATA, t2.DATA))

If the table is normal table with a fully alphanumeric name, the shorthand defaults to that name. In practice a shorthand is always needed if multiple tables are used.

The FROM clause can be omitted, in which case the input is a virtual table with no columns. The number of rows in it is defined by the LIMIT and OFFSET value; it defaults to 1 row. It makes it possible to select column-independent expressions in the SELECT command. Note that these expressions do not need to be constant. For example

creates a temporary table with column Col_1 and 31 rows containing the values 0..30.

A table can be given in a variety of ways.

1. A persistent table can be used by giving its name which can contain path specification and environment variables or the UNIX ~ notation. If the tablename contains a special character, the character can be escaped with a backslash or the table name can be enclosed in single or double quotes.
2. A table name can be taken from a keyword in a previously specified table. This can be useful in a subquery. The syntax for this is the same as that for specifying keywords in an expression. E.g.
     SELECT FROM mytable tab  
       WHERE col1 IN [SELECT subcol FROM tab.col2::key]

   In this example key is a table keyword of column col2 in table mytable (note that tab is the shorthand for mytable and could be left out).
   It can also be used for another table in the main query. E.g.

     SELECT FROM mytable, ::key subtab  
       WHERE col1 > subtab.key1

   In this example the keyword key1 is taken from the subtable given by the table keyword key in the main table.
   If a keyword is used as the table name, the keyword is searched in one of the tables previously given. The search starts at the current query level and proceeds outwards (i.e., up to the main query level). If a shorthand is given, only tables with that shorthand are taken into account. If no shorthand is given, only primary tables are taken into account.

3. Opening a subtable using a path name like my.ms/ANTENNA will fail if my.ms is a reference table instead of the original table. Therefore the path of a subtable should be given using colons instead of slashes like my.ms::ANTENNA which is a slight extension of specifying table names in the previous bullet.
   In this way a subtable can always be found.
4. Similar to OQL it is possible to use a nested query command in the FROM clause. This is a normal query command enclosed in square brackets or parentheses. Besides the SELECT command the COUNT and CREATE TABLE command can also be used. The table created can thereafter be used in the rest of the query command by using the shorthand given to that table. It can also be used in the remainder of the table_list, thus using it as a backreference. Such backreferencing can be useful to avoid multiple equal subqueries. E.g.
     select from MS,  
      [select from MS where sumsqr(UVW[1:2]) < 625]  
       as TIMESEL  
      where TIME in [select distinct TIME from TIMESEL]  
       &&  any([ANTENNA1,ANTENNA2] in  
         [select from TIMESEL giving  
           [iif(UVW[3] < 0, ANTENNA1, ANTENNA2)]])

   is a command to find shadowed antennas for the VLA. Without the query in the FROM command the subqueries in the remainder of the command would have been more complex. Furthermore, it would have been necessary to execute that select twice.
   The command above is quite complex and cannot be fully understood before reading the rest of this note. Note, however, that the command uses the shorthand TIMESEL to be able to use the temporary table in the subqueries.

5. Normally only persistent tables (i.e., tables on disk) can be used. However, it is also possible to use transient tables in a TaQL command given in Python, Glish, or C++. This is done by passing one or more table objects to the function executing the TaQL command. In the TaQL command a $-sign followed by a sequence number has to be given to indicate the correct object containing the transient table. E.g., if two table objects are passed $1 indicates the first table, while $2 indicates the second one.
6. It is possible to use a concatenation of tables with the same description by giving a list of tables enclosed in square brackets. In this way it is, for example, possible to do a query on the combined parts of a MeasurementSet partitioned in time. Each table in the list can be specified in one of the ways mentioned in this section, including another table concatenation.
   For example:
     SELECT FROM [ms.part1, ms.part2, ms.part3] WHERE ...

   does a query on the three parts of an MS which are seen as a single table.
   It is possible to use glob filename patterns in such a list. For example

     SELECT FROM [ms.part*] WHERE ...

   is the same as the example above if no other files with such a name exist. An error is given if no table is found matching the pattern.

   Subtables of the concatenated tables can be concatenated as well. Alternatively, they can be assumed to be the same for all tables meaning that the subtable of the concatenation is the subtable of the first table. For example, when partitioning a MeasurementSet in time, the ANTENNA subtable is the same for all parts, while the POINTING and SYSCAL subtables depend on time, thus have to be concatenated as well. Concatenation of subtables can be achieved by giving them as a comma-separated list of names after the SUBTABLES keyword. For example:

     SELECT FROM [ms.part1, ms.part2 SUBTABLES SYSCAL,POINTING]

   Usually the result of a TaQL query references the table given in the FROM. In this example the FROM table is the concatenation, which is only known during the query. In such a case the concatenation must be made persistent, which can be done by using a GIVING (or INTO) inside the concatenation specification. Only the table name can be given, because the persistent concatenation only keeps the original table names; it does not make a copy of all data.
   For example:

     SELECT FROM [ms.part1, ms.part2 GIVING ms.conc]  
        WHERE ANTENNA1 != ANTENAA2 GIVING ms.cross

   selects the cross-correlation baselines from the concatenation. Note the two GIVING commands. The first one makes the concatenation persistent, the second one is the query result of the query ms.cross. It references the matching rows in the persistent concatenation ms.conc which in its turn references the original parts.

3.5 WHERE expression

It defines the selection expression which must have a boolean scalar result. A row in the primary table is selected if the expression is true for the values in that row. The syntax of the expression is explained in a section 4.

3.6 GROUPBY group_list

It defines how rows have to be grouped. Usually a result per group will be calculated using aggregate functions. A group consists of all rows for which the columns (or expressions) given in the group_list have the same value. The (aggregate) expressions in the SELECT clause are calculated for the entire group. In this way one can get, for example, the mean XX amplitude and the number of time slots per baseline like:

It results in a table containing nbaseline rows with in each row the antenna ids, mean amplitude, and number of rows.
If no aggregate function is used for a column, the value of the last row in the group is used. Note that in this example ANTENNA1 and ANTENNA2 are the same for the entire group. However, if TIME was also selected, only the last time would be part of the result.
Note that each expression in the group_list has to result in a scalar value of type bool, integer, double, date, or string.
Aggregate functions are discussed in more detail in section 5.

3.7 HAVING expression

This clause can be used to select specific groups. Only the groups (defined by GROUPBY) are selected for which the HAVING expression is true.
Note that HAVING can be given without GROUPBY, although that will hardly ever be useful. If no GROUPBY is given, but the SELECT statement contains an aggregate function, the result is a single group. HAVING cannot be used if neither GROUPBY nor SELECT aggregate functions are used.
It is discussed in more detail in section 5.

3.8 ORDERBY sort_list

It defines the order in which the result of the selection has to be sorted. The sort_list is a comma separated list of expressions. It operates on the output of the SELECT, thus after a possible GROUPBY and HAVING clause are executed.
The sort_list can be preceded by the word ASC or DESC indicating if the given expressions are by default sorted in ascending or descending order (default is ASC). Each expression in the sort_list can optionally be followed by ASC or DESC to override the default order for that particular sort key.
To be compliant with SQL whitespace can be used between the words ORDER and BY.

The word ORDERBY can optionally be followed by DISTINCT which means that only the first row of multiple rows with equal sort keys is kept in the result. To be compliant with SQL dialects the word UNIQUE or NODUPLICATES can be used instead of DISTINCT.

An expression can be a scalar column or a single element from an array column. In these cases some optimization is performed by reading the entire column directly.
It can also be an arbitrarily complex expression with exactly the same syntax rules as the expressions in the WHERE clause. The resulting data type of the expression must be a standard scalar one, thus it cannot be a Regex or DateTime (see below for a discussion of the available data types). E.g.

3.9 LIMIT/OFFSET expression

It indicates which of the matching and sorted rows should be selected. If not given, all of them are selected. The word TOP can also be used instead of LIMIT.
LIMIT and OFFSET are applied after ORDERBY and SELECT DISTINCT, so they are particularly useful in combination with those clauses to select, for example, the highest 10 values.

It can be given in two ways:

• In the semi-standard SQL way using LIMIT N to select N rows and/or OFFSET M to skip the first M rows. Similar to Python, N and M can be negative meaning they are counted from the end. E.g., LIMIT -1 means all rows but the last.
• As a Python-style range using LIMIT start:end:incr, where the end is exclusive. Start defaults to 0, end to the number of rows, and incr to 1. As above, start and end can be negative to count from the end. The increment must be positive.

For example:

sorts uniquely by time, skips the first 10 rows, and selects the next two rows.

selects every 100-th row.

[ AS options] — set]

3.10 GIVING [table] [AS options] — set

It indicates that the ultimate result of the SELECT command should be written to a table (with the given name).
Another (more SQL compliant) way to define the output table is the INTO clause. See INTO for a more detailed description including the possible types.

It is also possible to specify a set in the GIVING clause instead of a table name. This is very useful if the result of a subquery is used in the main query. Such a set can contain multiple elements Each element can be a single value, range and/or interval as long as all elements have the same data type. The parts of each element have to be expressions resulting in a scalar.

In the main query and in a query in the FROM command the GIVING clause can only result in a table and not in a set.
To be compliant with SQL dialects, the word SAVETO can be used instead of GIVING. Whitespace can be given between SAVE and TO.

4 Expressions

An expression is the basic building block of TaQL. They are similar to expressions in other languages. An expression is formed by applying an operator or a function to operands which can be a table column or keyword, a constant, or a subexpression. An operand can be a scalar value or an array or set. The next subsections discuss them in detail.

An expression can be used in several places:

• In the WHERE and HAVING clause where the result must be a boolean scalar value. It tells if a table row or group will be selected.
• As a key in the GROUPBY clause where the result must be a scalar value (numeric, bool, or string).
• As a sort key in the ORDERBY clause where the result must be a scalar value (numeric, bool, or string)
• As an element in the set in the GIVING clause. It must be a scalar value of any type except regex.
• As a scalar or array value in the INSERT and UPDATE command.
• As a column expression in the column-list part of the SELECT command. The result can be a scalar or array value.
• As a scalar or array value in the CALC command.
• As a scalar or array value in various ALTER TABLE subcommands

The expression in the clause can be as complex as one likes using arithmetic, comparison, and logical operators. Parentheses can be used to group subexpressions.
The operands in an expression can be table columns, table keywords, constants, units, functions, sets and intervals, and subqueries.
The index operator can be used to take a single element or a subsection from an array expression.
For example,

The last example shows a set with a continuous interval.

4.1 Data Types

Internally TaQL uses the following data types:

Bool

logical values (true/false (case-insensitive) or T/F)

Integer

integer numbers up to 64 bits

Double

64 bit floating point numbers including times/positions

Complex

128 bit complex numbers

String

string values on which operator + can be used (concatenation).

Regex

regular expressions can be used for string matching (see section 4.2). Maximum string distances can also be used in a way similar to regular expressions.

DateTime

representing a date/time. There are several functions acting on a date/time. Operator + and - can be used on them.

Scalars and arbitrarily shaped arrays of these data types can be used. However, arrays of Regex are not possible.
If an operand or function argument with a non-matching data type is used, TaQL will do the following automatic conversions:
- from Integer to Double or Complex.
- from Double to Complex.
- from String or Double to DateTime.

In this document some special data types are used when describing the functions.
- Real means Integer or Double.
- Numeric means Integer, Double, or Complex.
- DNumeric means Double or Complex.

TaQL supports any possible data type of a table column or keyword. In some commands (column list and CREATE TABLE) columns are created where it is possible to specify the data type of a column. The following case-insensitive values can be used to specify a type:

The TIME type is a special data type. It means that the column gets data type DOUBLE and that a MEASINFO record will be defined in the column keywords to designate the column as an epoch.

4.2 Regular Expressions and String Distances

TaQL supports the use of extended regular expressions and string distances. They can be specified in various ways as discussed in section 4.3.6. There are three basic types of regular expressions.

• An SQL-style pattern is quite simple. It has 2 special characters. The underscore (_) means a single arbitrary character and the percent (%) means zero or more arbitrary characters. Special characters can be escaped with a backslash to retain their normal meaning. For example:
    3c\_%

  matches 3c_ and 3c_xx, but not 3caxx.

• A UNIX-style pattern, as often used for wildcarded file names, is more powerful than the SQL-style pattern. It has a few special characters that can be escaped with a backslash.
  • The question mark (?) means a single arbitrary character.
  • The asterisk (*) means zero or more arbitrary characters. For example: 3c_* does the same as the SQL-style pattern above.
  • Square brackets indicate a bracket expression (character choice). For example: [ab] matches a and b, but not c. A few special characters can be used in a bracket expression:
    • A leading ^ or ! means negation. Thus [!ab] matches every character except a and b.
    • A minus sign indicates a range. For example [0-9] matches a digit or [a-z] matches a lowercase letter. If a minus sign cannot be interpreted as a range, it is a literal minus sign like in [-ab] or the second minus sign in [a-z-A].
    • Posix character classes [:xx:] where xx can be:
      - alpha matching any letter
      - lower matching any lowercase letter
      - upper matching any uppercase letter
      - alnum matching any digit or letter
      - digit matching any digit
      - xdigit matching any hexadecimal digit (0-9a-fA-F)
      - space matching any whitespace character
      - print matching any printable character (alnum, punct, space)
      - punct matching any non-alnum visible character (.,!? etc.)
      - graph matching any visible printable character (alnum, punct)
      - cntrl matching any control character.
      For example [_[:isalpha:]][_[:isalnum:]]* to match variable names.
    • A bracket expression cannot be empty, thus if ] is the first character in the bracket expression, it is interpreted literally. Note that is also true if it is the first character after the negation character.
    • A backslash in a character class is always interpreted literally, thus special characters cannot be escaped. However, as shown above they can always be placed such that they are interpreted literally.
  • Braces can be used for a choice between (possible empty) multi-character strings separated by commas. Escape a comma or brace with a backslash to treat it literally. For example:
    *.{h,hpp,c,cc,cpp}
    It is fully nestable, thus choice strings can be patterns. For example:
    *.{[hc]{,pp},c}
    does the same as the example above. Note that the inner choice is between an empty string and pp.
• An awk/egrep-like extended regular expression is most powerful. A full explanation can be found on Wikipedia. Here only a summary of its special characters is given. They can be escaped using backslashes.
  • . matches any character.
  • ^ matches beginning of string.
  • $ matches end of string.
  • Square brackets for a bracket expression. It is the same as described above with the exception that ! cannot be used as negation character.
  • * matches zero or more occurrences of previous character or subexpression.
  • + matches one or more occurrences.
  • ? matches zero or one occurrence.
  • { and } for an interval giving minimum and maximum number of occurrences. For example:
    [a-z]{3,5} matches lowercase string with a minimum of 3 and maximum of 5 characters.
    [a-z]{3} matches exactly 3 characters.
    [a-z]{3,} matches at least 3 characters.
    [a-z]{,5} matches at most 5 characters.
  • | matches left or right substring
  • ( and ) to form subexpressions for operators like *.
  • \1 till \9 mean backreference to a subexpression (first one is \1). A string part matches if it is equal to the string part matching that subexpression. e.g., (a*)x\1 matches x, axa, aaxaa, etc., but not axaa nor aaxa.

  For example:

    .*\.(h|hpp|c|cc|cpp)  
    .*\.[hc](pp)?|cc

  do the same as the pattern examples above.

Furthermore it is possible to specify maximum string distances (known as Levensthein or Edit distance). It is explained in section 4.3.6.

4.3 Constants

Scalar constants of the various data types can be formed in a way similar to Python and Glish. Array constants can be formed from scalar constants.

4.3.1 Bool

A Bool constant is the value T or F (both in uppercase) or the value true or false (any case).

4.3.2 Integer

An integer constant is a numeric value without decimal point or exponent. It can also be given as a hexadecimal value like 0xffff.

4.3.3 Double (and time/position)

A floating-point constant is given with a decimal point and/or exponent. ’E’ or ’e’ can be used to specify the exponent. An integer number followed by a unit is also regarded as a double constant.
Another way to define a Double constant is by means of a Time or Position. Such a constant is always converted to radians. It can be given in several ways:

• An integer or floating-point number immediately followed by a simple unit (thus without whitespace). e.g., 12.43deg
  Some valid units are deg, arcmin, arcsec (or as), rad. The units can be scaled by preceding them with a letter (e.g., mrad is millirad).
• A time/position in HMS format. Seconds can be left out. e.g., 12h34m34.5 or 8h32m
• A position in DMS format. Seconds can be left out. e.g., 12d34m34.5 or 8d0m
• A position as DMS in dot format. Note that all parts must be present. e.g., 12.34.34.5 or 8.0.34.5

4.3.4 Complex

The imaginary part of a Complex constant is formed by an Integer or Double constant immediately followed by a lowercase i or j. A full Complex constant is formed by adding another Integer or Double constant as the real part. E.g.

Note that a full Complex constant has to be enclosed in parentheses if, say, a multiplication is performed on it. E.g.

4.3.5 String

A String constant has to be enclosed in ” or ’ and can be concatenated (as in C++). E.g.

4.3.6 Regular expression and String distance

A regular expression constant can be given directly or using a function.

• An SQL-style pattern can be given directly as a string constant preceded by operator LIKE or NOT LIKE.
• A pattern or regular expression can be given like x/expr/q preceded by operator ~ or !~. Instead of a slash, the characters % and # can also be used as delimiter, as long as the same delimiter is used on both sides. The delimiter can not be part of the expression (not even escaped with a backslash).
  The x denotes the type:
  - p means a pattern matching the full string.
  - f means a regular expression matching the full string.
  - m means a regular expression matching part of the string (a la Perl).
  The q denotes optional qualifiers. Currently only i is supported meaning a case-insensitive match. For example:
    name~p/3[cC]*/  
    name ~ p%3c*%i  
    lower(name) ~ p%3c*%  
    name ~ m/^3c/i  
    name ~ f/3c.*/i  
    filename !~ p#/usr/*.{h,cc}#

  All examples but the last one do the same: matching a name starting with 3c or 3C.
  The last example shows a glob-style pattern to find files on /usr not ending in .h or .cc.

• Apart from these Perl-like specifications, a regular expression can also be formed by applying a function to a string constant. The operator = or != has to be applied to it.
  • Function sqlpattern treats its argument as an SQL-style pattern. For example:
      lower(name) LIKE ’3c%’  
      lower(name) = sqlpattern(’3c%’)

    do the same.

  • Function pattern treats its argument as a UNIX-style pattern.
  • Function regex treats its argument as a full regular expression.

  Case-insensitive matching can only be done as shown in the example above by downcasing the string to be matched.
  Please note that these functions are not limited to constants. They can also be used to form regular expressions from variables.

A maximum string distance constant can be specified in a similar way. Such a distance is known as the Levensthein or Edit distance. It is a measure of the similarity of strings by counting the minimum number of edits (deletions, insertions, substitutions, and swaps of adjacent characters) that need to be done to make the strings equal.

This tests if the strings in the given column are within the maximum distance of the string given in the constant. The following qualifiers can be given (in any order):

• i means a case insensitive test.
• b means that blanks in the strings are ignored.
• nn is an integer value giving the maximum distance. If not given it defaults to 1 + len(string) / 3.

4.3.7 Date/time

DateTime constant can be formed in 2 ways:

1. From a String constant using the datetime function. In this way all possible formats as explained in class MVTime are supported. E.g.
     datetime ("11-Dec-1972")

2. A more convenient way is to specify it directly. Since this makes use of the delimiters space, - or /, it conflicts with the expression grammar as such. However, possible conflicts can be solved by using whitespace in an expression and it is believed that in practice the convenience surpasses the possible conflicts.
   A large subset of the MVTime formats is supported. A DateTime has to be specified as date/time or date-time, where the time part (including the space, -, or / delimiter) is optional. The possible date formats are:
   - YYYY/MM/DD, YYYY-MM-DD, or DD-MM-YYYY where DD and MM must be 2 digits and YYYY 4 digits.
   - DD-MMMMMMMM-YY where the - is optional and MMMMMMM is the case-insensitive name of the month (at least 3 letters). DD can be 1 or 2 digits and YY 1 to 4 digits. 2000 is added if YY$<$50 and 1900 is added if 50$<=$YY$<$100.
   If MM$>$12, YYYY will be incremented accordingly.
   The general time format in a DateTime constant is:
   - hh:mm:ss.s
   where the delimiter h or H can be used for the first colon and m or M for the second. Trailing parts can be omitted. E.g.
     10-02-1997  
     10-February-97  
     10feb97  
     1997/02/10         are all identical  
    
     1May96/3:          : (or h) is mandatory  
     1May96-3:0  
     1May96 3:0:0  
     1May96-3h          h (or :) is mandatory  
     1May96 3H0  
     1May96/3h0M  
     1May96/3hm0.0

   A DateTime constant with the current date/time can be made by using the function datetime without arguments.

4.3.8 Arrays

N-dimensional arrays of all data types can be created with the exception of regular expressions.
It is possible to form a 1-dimensional array from a constant bounded discrete set. When needed such a set is automatically transformed to an array. E.g.

The first example results in an integer array of 10 elements with values 0..9. The others result in a string array of 3 elements. The second version already shows that strings can be concatenated (as explained further on).

A multi-dimensional array can be formed by giving a set of arrays. A nested list resembles the numpy way. For example:

results in a 2-dim array. However, it is also possible to use arrays created in other ways such as arrays in a column or arrays created with the array function described below. For example:

results in a 3-dim array.

Furthermore it is possible to use the array function to create an array of any shape. The values are given in the first argument as a scalar, set, or another array. The shape is given in the latter arguments as scalars or as a set. The array is initialized to the values given which are wrapped if the array has more elements.

The first examples create an array with shape [10,4] containing the values 1..10 in each line. The latter results in a boolean array having the same shape as the DATA array and filled with False.

4.3.9 Masked Arrays

An array can have an optional mask. Similar to numpy’s masked array, a mask value True means that the value is masked off, thus not taken into account in reduce functions like calculating the mean.
Note that this definition is the same as the FLAG column in a MeasurementSet, but is different from a mask in a Casacore Image where True means good and False means bad.

All operations on arrays will take the possible mask into account. Reduce functions like median only use the unmasked array elements. Furthermore, partial reduce functions like medians will set an output mask element to True if the corresponding input array part has no unmasked elements.
Operators like + and functions like cos operate on all array elements. The mask in the resulting array is the logical OR of the input masks. Of course, the result has no mask if no input array has a mask.

A masked array is created by applying a boolean array to an array using the square brackets operator. Both arrays must have the same shape. For example:

The first example applies the FLAG column in a MeasurementSet to the DATA column. The second example masks off high DATA values.

The functions arraydata, arraymask, and flatten can be used to get the array data or mask. The last one flattens the array to a vector while removing all masked elements.

The TaQL commands putting values into a table accept two columns (in parentheses) for a masked array. This is described in more detail in the appropriate sections. For example:

to write the data averaged over the first axis (frequency channel)into column MD. Only the unflagged data points are taken into account. The output contains the resulting flags in column MM; a flag is set to True if all channels were flagged.

4.3.10 Null Arrays

A cell in a table column containing variable shaped arrays, can be empty. Such a cell does not contain an array and is represented in TaQL as a null array. Note it is different from a cell containing an empty array, which is an array without values.

Null arrays can be used with any operator and in any function. If one of the operands or function arguments is a null array, the result will be a null array; only array functions reducing to a scalar (such as sum and mean) give a valid value (usually 0).

The UPDATE and INSERT commands will ignore a null array result; no value is written in that row.

4.4 Table Columns

A table column can be used in a query by giving its name in the expression, possibly qualified with a table shorthand name. A column can contain a scalar or an array value of any data type supported by the table system. It will be mapped to the available TaQL data types. If the column keywords define a unit for the column, the unit will be used by TaQL.

The name of a column can contain alphanumeric characters and underscores. It should start with an alphabetic character or underscore. A column name is case-sensitive.
It is possible to use other characters in the name by escaping them with a backslash. e.g., DATE\-OBS.
In the same way a numeric character can be used as the first character of the column name. e.g., \1stDay.
A reserved word cannot be used directly as a column name. It can, however, be used by escaping it with a backslash. e.g., \IN.
Note that in programming languages like C++ and Python a backslash itself has to be escaped by another backslash. e.g., in Python: tab.query(’DATE\\-OBS$>$10MAR1996’).

If a column contains a record, one has to specify a field in it using the dot operator; e.g., col.fld means use field fld in the column. It is fully recursive, so col.fld.subfld can be used if field fld is a record in itself.
Alas records in columns are not really supported yet. One can specify fields, but thereafter an error message will be given.

4.4.1 Referring to SELECT columns

Usually a column used in an expression will be a column in one of the tables specified in the FROM clause. However, it is possible to use a column created in the SELECT clause, in expressions given in the HAVING or ORDERBY clause. In fact, a column name not preceded by a table shorthand, is first looked up in the SELECT columns and thereafter in the first FROM table.

It can be advantageous to use a SELECT column if that column is an expression; it saves both typing and execution time. because that expression is executed only once.

4.5 Table Keywords

It is possible to use table or column keywords, which can have a scalar or an array value or a record, possibly nested. A table keyword has to be specified as ::key. In an expression the :: part can be omitted if there is no column with the same name. A column keyword has to be specified as column::key.
Note that the :: syntax is chosen, because it is similar to the scope operator in C++.
As explained in the FROM clause, keywords in the primary table and in other tables can be used. If used from another table, it has to be qualified with the (shorthand) name of the table. E.g.,
sh.key or sh.::key
takes table keyword key from the table with the shorthand name sh.

If a keyword value is a record, it is possible to use a field in it using the dot operator. e.g., ::key.fld to use field fld. It is fully recursive, so if the field is a record in itself, a subfield can be used like col::key.fld.subfld

A keyword can be used in any expression. It is evaluated immediately and transformed to a constant value.

4.6 Operators

TaQL has a fair amount of operators which have the same meaning as their C and Python counterparts. The operator precedence order is:

Operator names are case-insensitive. For SQL compliancy some operators have a synonym.

All operators can be used for scalars and arrays and a mix of them. Note that arrays of regular expressions cannot be used.

The following table shows all available operators and the data types that can be used with them.

4.7 Sets and intervals

As in SQL the operator IN can be used to do a selection based on a set. E.g.

The result of operator IN is true if the column value matches one of the values in the set. A set can contain any data type except a regex.

This example shows that (in its simplest form) a set consists of one or more values (which can be arbitrary expressions) separated by commas and enclosed in square brackets. The elements in a set have to be scalars and their data types have to be the same or convertible to a common data type. The square brackets can be left out if the set consists of only one element. For SQL compliance parentheses can be used instead of square brackets.

An array is also a set, so IN can also be used on an array like:

where expr1 is the array result of some expression. It is also possible to use a scalar as the righthand of operator IN. So if expr1 is a scalar, operator IN gives the same result as operator ==.

The lefthand operand of the IN operator can also be an array or set. In that case the result is a boolean array telling for each element in the lefthand operand if it is found in the righthand operand.

An element in a set can be more complicated than a single value. It can define multiple values and intervals. The possible forms of a set element are:

1. A single value as shown in the example above.
2. start:end:incr. This is similar to the way an array index is specified. Incr defaults to 1. End defaults to an open end (i.e., no upper bound) and results in an unbounded set. Start and end can be a real or a datetime. Incr has to be a real. Some examples:
     1:10         means 1,2,...,9  (also 10 when using glish style)  
     1:10:2       means 1,3,5,7,9  
     1::2         means all odd numbers  
     1:           means all positive integer numbers  
     18Aug97::2   means every other day from 18Aug97 on

   These examples show constants only, but start, end, and incr can be any expression.
   Note that :: used here can conflict with the :: in the keywords. e.g., a::b is scanned as a keyword specification. If the intention is start::incr, whitespace should be used as in a: :b. In practice this conflict will hardly ever occur.

3. Continuous intervals can be specified for data type real, string, and datetime. The specification of an interval resembles the mathematical notation 1<x<5, where x is replaced by :. An open interval side is indicated by <, while a closed interval side is indicated by =.
   Another way to specify intervals is using curly and/or angle brackets. A curly bracket is a closed side, the angle bracket is an open side. The following examples show how bounded and half-bounded, (half-)open and closed intervals can be specified.
     1=:=5   {1,5}     means 1<=x<=5   bounded closed  
     1<:<5   <1,5>     means 1<x<5     bounded open  
     1=:<5   {1,5>     means 1<=x<5    bounded right-open  
     1<:=5   <1,5}     means 1<x<=5    bounded left-open  
     1=:  {1,}  {1,>   means 1<=x      left-bounded closed  
     1<:  <1,}  <1,>   means 1<x       left-bounded open  
     :=5  {,5}  <,5}   means x<=5      right-bounded closed  
     :<5  {,5>  <,5>   means x<5       right-bounded open

It is very important to note that the 2nd form of set specification results in discrete values, while the 3rd form results in a continuous interval.

Each element in a set can have its own form, i.e., one element can be a single value while another can be an interval. If a set consists of single or bounded discrete start:end:incr values only, the set will be expanded to an array. This makes it possible for array operators and functions (like mean) to be applied to such sets. E.g.

Another form of constructing a set is using a subquery as described in section 4.11.

4.8 Array Index Operator

It is possible to take a subsection or a single element from an array column, keyword or expression using the index operator [index1,index2,...]. This syntax is similar to that used in Python or Glish. Similar to Python a negative value can be given meaning counting from the end. However, a negative stride cannot be given. Taking a single element can be done as:

Taking a subsection can be done as:

If a start value is left out it defaults to the beginning of that axis. An end value defaults to the end of the axis and an increment defaults to one. If an entire axis is left out, it defaults to the entire axis.
E.g., an array with shape [10,15,20] can be subsectioned as:

The examples show that an index can be a simple constant (as it will usually be). It can also be an expression which can be as complex as one likes. The expression has to result in a real value which will be truncated to an integer.
For fixed shaped arrays checking if array bounds are exceeded is done at parse time. For variable shaped arrays it can only be done per row. If array bounds are exceeded, an exception is thrown. In the future a special undefined value will be assigned if bounds of variable shaped arrays are exceeded to prevent the selection process from aborting due to the exception.

Note that the index operator will be applied directly to a column. This results in reading only the required part of the array from the table column on disk. It is, however, also possible to apply it to a subexpression (enclosed in parentheses) resulting in an array. E.g.

can both be used and have the same result. However, the first form is faster, because only a single element is read (resulting in a scalar) and 1 is added to it. The second form results in reading the entire array. 1 is added to all elements and only then the requested element is taken.
From this example it should be clear that indexing an array expression has to be done with care.

4.9 Units

Units can be given at many places in an expression; in fact, each subexpression can be ended with a unit meaning that the subexpression result gets that unit or will be converted to that unit. A simple unit (only letters) can always be given literally. A non-simple unit can be given literally if only containing digits, underscores and/or dots (e.g., m2, fl_oz. or m.m). Otherwise the unit has to be enclosed in single or double quotes (e.g., ’m/s’) or the backslash has to be used as escape character (e.g., \in or m\/s).
Arguments to functions like sin will be converted to the appropriate unit (radians) as needed. In a similar way, the units of operands to operators like addition, will be converted as needed. An exception is thrown if a unit conversion is not possible.

All units supported by module Quanta can be used. Note that the units are case sensitive. Most common units use lowercase characters. A unit can be preceded by a scaling prefix (like k for kilo). Compound units are created when multiplying or dividing values with units. The available units can be shown using the show units command of the program taql.

Units can be given as follows:

• If a column has a unit defined in column keyword QuantumUnits or UNIT, it automatically gets that unit.
• A constant can immediately be followed by a simple unit. e.g., 2deg.
• The result of several expressions have an implicit unit.
  Constants given as positions are in radians (rad).
  Difference of 2 dates is in days (d).
  Inverse trigonometric functions like asin give radians.
• When combining values with different units in e.g., an interval, a set, an addition, or a function like min, the values are converted to the unit of the first operand or argument with a unit. Values without a unit have by default the unit of the first operand or argument with a unit.
    3mm-7cm          result is -67 mm  
    3+3mm            result is 6 mm  
    3mm<:<3cm        result is interval <3mm,30mm>  
    [3,4cm,5]        result is [3cm, 4cm, 5cm]  
    [5, 7cm, 8mm]    result is [5cm, 7cm, 0.8cm]  
    [5, 7mm, 8cm]    result is [5mm, 7mm, 80mm]  
    max(3mm,2cm)     result is 20 mm  
    5 ’km/h’ + 1 ’m/s’      is 8.6 km/h  
    iif(F,3min,30sec)       is 0.5 min

• Similarly, operands of comparison operators and arguments of comparison functions (like near) are converted to the unit of the first operand or argument with a unit.
• The result of a multiplication and division is a compound unit if both operands have a unit. Otherwise it is the unit of the argument with a unit.
  Before TaQL supported units, it was needed to divide the TIME column in a MeasurementSet by 86400 to convert it to days, so it could be compared with a given date/time. So, for backward compatibility, a division of a value with unit s by a constant 86400 results in unit d.
• Division of units of an equal kind (e.g., km by m) results in a unitless value.
• Division of a unitless value by a value with unit, results in the reciproke unit.
    1km/10m          result is 100  
    1/2s             result is 0.5 ’(s)-1’  
    1Hz + 1/2s       result is 1.5Hz

• The result of functions like SUMSQR and SQRT is a compound unit if the argument has a unit. Note that sqrt(2m) will fail, because the square root of a meter does not exist.
• A (sub)expression can be followed by a simple or compound unit. If the subexpression has no unit, it gets the given unit. Otherwise the resulting value is converted to the unit.
  Note that some units can be the same as a reserved word (e.g., as or in). In that case it has to be escaped or enclosed in quotes.
    COL \in          set/convert column COL to inch  
    3 cm             result is 3 cm  
    3 ’km/s’         result is 3 km/second  
    3mm cm           result is 0.3 cm  
    (3mm cm)m        result is 0.003 m  
    (3+3) cm         result is 6 cm  
    (3+3mm) cm       result is 0.6 cm  
    [3,4,5]mm        result is [3mm, 4mm, 5mm]  
    [3,4cm,5]mm      result is [30mm, 40mm, 50mm]  
        Note: all values in the set first get the same unit cm  
    asin(1)          result is pi/2 radians  
    asin(1) deg      result is 90 degrees  
    (3mm+7cm) m      result is 0.073 m

• If a function argument is expected in a certain unit, values are converted as needed. For example, arguments to functions sin and anycone are automatically converted to radians.
• When adding or subtracting a value from a date, that value is converted to unit d (days).

Units will probably mostly be used in an expression in the WHERE clause or in a CALC command. However, it is also possible to use a unit in the selection of a column in the SELECT clause. For example:

In such a case the selection is an expression and the unit is stored in the column keywords. Thus in this example, TIME is stored in a column TIMED with keyword QuantumUnits=d and the values are converted to days.

4.10 Functions

More than 100 functions exist to operate on scalar and/or array values. Some functions have two names. One name is the CASA/Glish name, while the other is the name as used in SQL. In the following tables the function names are shown in uppercase, while the result and argument types are shown in lowercase. Note, however, that function names are case-insensitive.
Furthermore it is possible to have user defined functions that are dynamically loaded from a shared library. In section Writing user defined functions it is explained how to write user defined functions.
A set of standard UDFs exists dealing with Measure conversions, for example to convert J2000 to apparent. Another set of UDFs deals with values and relations in MeasurementSets and Calibration Tables.

Sets, and in particular subqueries, can result in a 1-dim array. This means that the functions accepting an array argument can also be used on a set or the result of a subquery.

4.10.1 String functions

These functions can be used on a scalar or an array argument.

integer STRLENGTH(string), integer LEN(string)
Returns the number of characters in a string (trailing whitespace is significant).

string UPCASE(string), string UPPER(string)
Convert to uppercase.

string DOWNCASE(string), string LOWER(string)
Convert to lowercase.

string CAPITALIZE(string)
Capitalize a string (make first letter uppercase).

string LTRIM(string)
Removes leading whitespace.

string RTRIM(string)
Removes trailing whitespace.

string TRIM(string)
Removes leading and trailing whitespace.

string SUBSTR(string, integer ST, integer N)
Returns a substring starting at the 0-based position ST with a length of at most N characters. If the string argument is an array of strings, an array with the substring of each string is returned. The arguments ST and N have to be scalar values. If ST is negative, it counts from the end (a la Python). N and the resulting ST will be set to 0 if negative.

string REPLACE(string SRC, PATTERN, string REPL)
Replaces all occurrences of PATTERN in SRC by REPL and returns the result. REPL can be omitted and defaults to the empty string. If the first argument is an array of strings, each element in the array is replaced. The arguments PATTERN and REPL have to be scalar values. PATTERN can be a string or a regular expression (see below). For example:
REPLACE("abcdab", "ab") results in cd
REPLACE("abcdab", REGEX("^ab"), "xyz") results in xyzcdab

4.10.2 Regex functions

Apart from using regex/pattern constants, it is possible to use functions to form a regex or pattern. These functions can only be used on a scalar argument.

regex REGEX(string)
Handle the given string as a regular expression.

regex PATTERN(string)
Handle the given string as a UNIX filename-like pattern and convert it to a regular expression.

regex SQLPATTERN(string)
Handle the given string as an SQL-style pattern and convert it to a regular expression.

A regex formed this way can only be used in a comparison == or !=. E.g.
object == pattern(’3C*’)
to find all 3C objects in a catalogue.

A few remarks:

1. The regex/pattern functions and operator LIKE work on any string, thus they can be used with any string expression.
2. A Regex is case sensitive. One should use function upcase or downcase on the string to test to make it case insensitive or use the i qualifier on a regex constant.
3. Usually a regex/pattern must match the full string, thus not part of it. However, one can use the m// regex constant to do partial matching. Thus something like m/xx/ matches all strings containing xx. Of course, regex(’.*xx.*’) can also be used. In this way the m// regex works the same as in languages like Perl, Python, and Glish.

4.10.3 Date/time functions

These functions make it possible to handle dates/times and can be used on a scalar or an array argument. The syntax of a date/time string or constant is explained in section 4.3.7.

DateTime DATETIME(string)
Parse the string and convert it to a DateTime value.

DateTime MJDTODATE(real)
The real value, which has to be a MJD (ModifiedJulianDate), is converted to a DateTime.

double MJD(DateTime)
Get the DateTime as MJD (ModifiedJulianDate) in days.

DateTime DATE(DateTime)
Get the date (i.e., remove the time part). This function is needed in something like:
DATE(column) == 12Feb1997
if the column contains date/times with times$>$0.

double TIME(DateTime)
Get the time part of the day. It is converted to radians to be compatible with the internal representation of times/positions. In that way the function can easily be used as in:
TIME(date) $>$ 12h

integer YEAR(DateTime)
Get the year (which includes the century).

integer MONTH(DateTime)
Get the month number (1-12).

integer DAY(DateTime)
Get the day number (1-31).

integer WEEK(DateTime)
Get the week number in the year (0 ... 53).
Note that week 1 is the week containing Jan 4th.

integer WEEKDAY(DateTime), integer DOW(DateTime)
Get the weekday number (1=Monday, ..., 7=Sunday).

string CDATETIME(DateTime), string CTOD(DateTime)
Get the DateTime as a string like YYYY/MM/DD/HH:MM:SS.SSS.

string CDATE(DateTime)
Get the date part of a DateTime as a string like DD-MMM-YYYY.

string CTIME(DateTime)
Get the time part of a DateTime as a string like HH:MM:SS.SSS.

string CMONTH(DateTime)
Get the abbreviated name of the month (Jan ... Dec).

string CWEEKDAY(DateTime), string CDOW(DateTime)
Get the abbreviated name of the weekday (Mon ... Sun).

All functions can be used without an argument in which case the current date/time is used. e.g., DATE() results in the current date.
It is possible to give a string argument instead of a date. In this case the string is parsed and converted to a date (i.e., the function DATETIME is used implicitly).
Note that the function STR discussed in the next section can also be used for pretty-printing a date/time. It gives more control over the number of decimals and date format.

4.10.4 Pretty printing functions

Angles (scalar or array) can be returned as strings in HMS and/or DMS format. Currently, they are always formatted with 3 decimals in the seconds.

string HMS(real)
Return angle(s) like 12h34m56.789

string DMS(real)
Return angle(s) like 12d34m56.789

string HDMS(realarray)
Return angles like 12h34m56.789 (even elements) and 12d34m56.789 (odd elements). It is useful for arrays containing RA,DEC values.

The functions mentioned above and the date/time functions in the previous subsection can format a value in a predefined way only.
The STRING (shorthand STR) function makes it possible to convert values to strings using an optional format string or width.precision value. It also makes it possible to format dates, times, and angles in a variety of ways.

string STR(value, [format]), string STRING(value, [format])

The value can be of any type (except Regex) and can be a scalar or array. The optional format must be a scalar string or numeric value. If no format is given, an appropriate default format will be used.

• A numeric format value defines the width and/or precision. For example:
    8        defines width 8 and default precision  
    20.12    defines width 20 and precision 12  
    .8       defines precision 8 and default width

  In this way precision represents all digits, not only the ones behind the decimal point. A default width or precision is used if not given.

• A string format value can contain a printf-style format string, which must include the %-sign. Note that the real and imaginary part of a complex value are formatted separately, so such a format string needs to contain a format specifier for both parts. See printf reference for possible format specifiers. For example:
    %10d     decimal with width 10  
    %010d    decimal with width 10 and filled with zeroes  
    %f+%fi   to format a complex value as a+bi

  Apart from a printf-style format string, it is also possible to define a string to format date/time and angle values (which are automatically converted to radians if containing units).
  Such a format string contains one or more format values as defined in class MVTime. A vertical bar (with optional whitespace) must be used as separator. A string part can be a numeric value defining the precision of the time/angle. Default precision is 6 (thus hh:mm:ss). The optional time/angle formats and modifiers are:

  Format	Description
  YMD	yyyy/mm/dd/hh:mm:ss.sss
  YMD_ONLY	YMD without the time (same as YMD$|$NO_TIME)
  DMY	dd-Mon-yyyy/hh:mm:ss.sss
  FITS	yyyy-mm-ddThh:mm:ss.sss
  BOOST	the same as DMY$|$USE_SPACE
  NO_H, NO_D	suppress the output of hours (or degrees): useful for offsets
  NO_HM, NO_DM	suppress the degrees and minutes
  CLEAN	suppress leading or trailing periods or colons if not all time/angle parts
  	are printed (e.g., when giving NO_H or 4 decimals)
  DAY	precede the output with Day- (e.g., Wed-)
  NO_TIME	suppress printing of time
  ANGLE	+ddd.mm.ss.ttt
  TIME	hh:mm:ss.ttt
  USE_SPACE	use a space between date and time (and day and date)
  DIG2	get angle/time in range -90:+90 or -12:+12
  LOCAL	local time; in FITS mode append time zone as +hh:mm

  For example:

    YMD             format as YYYY/MM/DD/HH:MM:SS  
    DMY|NO_TIME     format as DD-MMM-YYYY  
    DMY | DAY | 8   format as Thu-DD-MMM-YYY/HH:MM:SS.SS  
    TIME            format a datetime or angle as HH:MM:SS  
    ANGLE|9         format an angle as DD.MM.SS.SSS

  If such a format string contains an invalid part, it is assumed that the entire string is a printf-style format string.

4.10.5 Comparison functions

The exact comparison of floating point values is quite tricky. Two functions make it possible to compare 2 double or complex values with a tolerance. They can be used on scalar and array arguments (and a mix of them). The tolerance must be a scalar though.
Note that operator  = is the same as NEAR with a tolerance of 1e-5.

bool NEAR(numeric val1, numeric val2, double tol)
Tests in a relative way if a value is near another. Relative means that the magnitude of the numbers is taken into account.
It returns abs(val2 - val1)/max(abs(val1),abs(val2)) < tol.
If tol<=0, it returns val1==val2. If either val is 0.0, it takes care of area around the minimum number that can be represented. The default tolerance is 1.0e-13.

bool NEARABS(numeric val1, numeric val2, double tol)
Tests in an absolute way if a value is near another. Absolute means that the magnitude of the numbers is not taken into account.
It returns abs(val2 - val1) < tol. The default tolerance is 1.0e-13.

bool ISNAN(numeric val)
Tests if a numeric value is a NaN (not-a-number).

bool ISINF(numeric val)
Tests if a numeric value is infinite (positive or negative).

bool ISFINITE(numeric val)
Tests if a numeric value is a finite number (not NaN or infinite).

4.10.6 Mathematical functions

Standard mathematical can be used on scalar and array arguments (and a mix of them).

double PI()
Return the value of pi.

double E()
Return the value of e (is equal to EXP(1)).

double C()
Return the value of the speed of light (with unit m/s).

dnumeric SIN(numeric)

dnumeric SINH(numeric)

double ASIN(real)

dnumeric COS(numeric)

dnumeric COSH(numeric)

double ACOS(real)

double TAN(real)

double TANH(real)

double ATAN(real)

double ATAN2(real y, real x)
Return ATAN(y/x) in correct quadrant.

dnumeric EXP(numeric)

dnumeric LOG(numeric)
Natural logarithm.

dnumeric LOG10(numeric)

dnumeric POW(numeric, numeric)
The same as operator **.

numeric SQUARE(numeric), numeric SQR(numeric)
The same as **2, but much faster.

dnumeric SQRT(numeric)

complex COMPLEX(real, real)

dnumeric CONJ(numeric)

double REAL(numeric)
Real part of a complex number. Returns argument if real.

double IMAG(numeric)
Imaginary part of a complex number. Returns 0 if argument is real.

real NORM(numeric)

real ABS(numeric), real AMPLITUDE(numeric)

double ARG(numeric), double PHASE(numeric)

numeric MIN(numeric, numeric)

numeric MAX(numeric, numeric)

real SIGN(real)
Return -1 for a negative value, 0 for zero, 1 for a positive value.

real ROUND(real)
Return the rounded value of the number. Negative numbers are rounded in an absolute way. e.g., ROUND(-1.6) = -2.

real FLOOR(real)
Works towards negative infinity. e.g., FLOOR(-1.2) = -2.

real CEIL(real)
Works towards positive infinity.

real FMOD(real, real)
The same as operator %.

Note that the arguments or results of the trigonometric functions are in radians. They are converted automatically if units are given.

4.10.7 Array to scalar reduce functions

The following functions reduce an array to a scalar. They are meant for an array, but can also be used for a scalar.

bool ANY(bool)
Is any element true?

bool ALL(bool)
Are all elements true?

integer NTRUE(bool)
Return number of true elements.

integer NFALSE(bool)
Return number of false elements.

numeric SUM(numeric)
Return sum of all elements.

numeric SUMSQUARE(numeric), numeric SUMSQR(numeric)
Return sum of all squared elements.

numeric PRODUCT(numeric)
Return product of all elements.

real MIN(real)
Return minimum of all elements.

real MAX(real)
Return maximum of all elements.

dnumeric MEAN(numeric), dnumeric AVG(numeric)
Return mean of all elements.

double VARIANCE(real)
Return variance (the sum of
(a(i) - mean(a))**2/(nelements(a) - 1).

double STDDEV(real)
Return standard deviation (the square root of the variance).

double AVDEV(real)
Return average deviation. (the sum of
abs(a[i] - mean(a))/nelements(a).

double RMS(real)
Return root-mean-squares. (the square root of the sum of
(a(i)**2)/nelements(a).

double MEDIAN(real)
Return median (the middle element). If the array has an even number of elements, the mean of the two middle elements is returned.

double FRACTILE(real, doublescalar fraction)
Return the value of the element at the given fraction. Fraction 0.5 is the same as the median, but no mean of the two middle elements is taken.

4.10.8 Array to array reduce functions

These functions reduce an array to a smaller array by collapsing the given axes using the given function. The axes are the last argument(s). They can be given in two ways:
- As a single set argument; for example, maxs(ARRAY,[1,2])
- As individual scalar arguments; for example, maxs(ARRAY,1,2)
For example, using MINS(array,0,1) for a 3-dim array results in a 1-dim array where each value is the minimum of each plane in the cube.
It is important to note that the interpretation of the axes numbers depends on the style being used. e.g., when using glish style, axes numbers are 1-based and in Fortran order, thus axis 1 is the most rapidly varying axis. When using python style, axis 0 is the most slowly varying axis.
Axes numbers exceeding the dimensionality of the array are ignored. For example, maxs(ARRAY,[1:10]) works for arrays of virtually any dimensionality and results in a 1-dim array.
The function names are the ’plural’ forms of the functions in the previous section. They can only be used for arrays, thus not for scalars.

bool ANYS(bool)
Is any element true?

bool ALLS(bool)
Are all elements true?

integer NTRUES(bool)
Return number of true elements.

integer NFALSES(bool)
Return number of false elements.

numeric SUMS(numeric)
Return sum of elements.

numeric SUMSQUARES(numeric), numeric SUMSQRS(numeric)
Return sum of squared elements.

numeric PRODUCTS(numeric)
Return product of elements.

real MINS(real)
Return minimum of elements.

real MAXS(real)
Return maximum of elements.

dnumeric MEANS(numeric), dnumeric AVGS(numeric)
Return mean of elements.

double VARIANCES(real)
Return variance (the sum of
(a(i) - mean(a))**2/(nelements(a) - 1).

double STDDEVS(real)
Return standard deviation (the square root of the variance).

double AVDEVS(real)
Return average deviation. (the sum of
abs(a(i) - mean(a))/nelements(a).

double RMSS(real)
Return root-mean-squares. (the square root of the sum of
(a(i)**2)/nelements(a).

double MEDIANS(real)
Return median (the middle element). If the array has an even number of elements, the mean of the two middle elements is returned.

double FRACTILES(real, doublescalar fraction)
Return the value of the element at the given fraction. Fraction 0.5 is the same as the median.

4.10.9 Array downsampling functions

These functions are a generalization of the functions in the previous section. They downsample an array by taking, say, the mean of every n*m elements. The functions in the previous section downsample by taking the mean of a full line or plane, etc. The most useful one is probably calculating the boxed mean, but the other ones can be used similarly. The width of each window axis has to be given. Missing axes default to 1. Similarly to the partial reduce functions described above, the axes can be given as scalars or as a set.
For example, BOXEDMEAN(array,3,3) calculates the mean in each 3x3 box. At the end of an axis the box used will be smaller if it does not fit integrally.
The functions can only be used for arrays, thus not for scalars.

bool BOXEDANY(bool)
Is any element true?

bool BOXEDALL(bool)
Are all elements true?

double BOXEDMIN(real)
Return minimum of elements.

double BOXEDMAX(real)
Return maximum of elements.

dnumeric BOXEDMEAN(numeric), dnumeric BOXEDAVG(numeric)
Return mean of elements.

double BOXEDVARIANCE(real)
Return variance (the sum of
(a(i) - mean(a))**2/(nelements(a) - 1).

double BOXEDSTDDEV(real)
Return standard deviation (the square root of the variance).

double BOXEDAVDEV(real)
Return average deviation. (the sum of
abs(a(i) - mean(a))/nelements(a).

double BOXEDRMS(real)
Return root-mean-squares. (the square root of the sum of
(a(i)**2)/nelements(a).

double BOXEDMEDIAN(real)
Return median (the middle element).

4.10.10 Array functions operating in running windows

These functions transform an array into an array with the same shape by operating on a rectangular window around each array element. The most useful one is probably calculating the running median, but the other ones can be used similarly. The half-width of each window axis has to be given; the full width is 2*halfwidth + 1. Missing axes default to a half-width of 0. Similarly to the partial reduce functions described above, the axes can be given as scalars or as a set.
For example, RUNNINGMEDIAN(array,1,1) calculates the median in a 3x3 box around each array element. See the examples how it is applied to an image.
In the result the edge elements (i.e., the elements where no full window can be applied) are set to 0 (or False).
The functions can only be used for arrays, thus not for scalars.

bool RUNNINGANY(bool)
Is any element true?

bool RUNNINGALL(bool)
Are all elements true?

double RUNNINGMIN(real)
Return minimum of elements.

double RUNNINGMAX(real)
Return maximum of elements.

dnumeric RUNNINGMEAN(numeric), dnumeric RUNNINGAVG(numeric)
Return mean of elements.

double RUNNINGVARIANCE(real)
Return variance (the sum of
(a(i) - mean(a))**2/(nelements(a) - 1).

double RUNNINGSTDDEV(real)
Return standard deviation (the square root of the variance).

double RUNNINGAVDEV(real)
Return average deviation. (the sum of
abs(a(i) - mean(a))/nelements(a).

double RUNNINGRMS(real)
Return root-mean-squares. (the square root of the sum of
(a(i)**2)/nelements(a).

double RUNNINGMEDIAN(real)
Return median (the middle element).

4.10.11 Type conversion functions

Explicit type conversions can be done using one of the functions below. They can operate on scalars and arrays.

integer INT(numeric or bool or string)
Convert the argument to an integer. A real number is truncated (-10.9 results in -10). For a complex number the truncated real part is taken. A bool is converted to 0 (False) or 1 (True). It does not check if a string represents a valid integer. It is interpreted until the first non-valid character, so a string containing a floating point value is truncated.

double REAL(numeric or bool or string)
Convert the argument to a real number. For a complex number the real part is taken. A bool is converted to 0 (False) or 1 (True). It does not check if a string represents a valid floating point value. A string is interpreted until the first non-valid character.

complex COMPLEX(real,real)
Form a complex number from the given real and imaginary part.

complex COMPLEX(string)
Convert the string to a complex number. The number can be given like (1,2) or 1+2i. In fact, any separator (except whitespace) between real and imaginary part is possible. It does not check if a string represents a valid complex value. The string is interpreted until the first non-valid character, so the last character can be any character (e.g., also j).

bool BOOL(anytype)
Convert the value to a bool. A numeric type (or date) results in False if the value is 0, otherwise True. A string is case-insensitive. False, F, No, N, -, or 0 results in False, otherwise True.

4.10.12 Array creation functions

anytypearray ARRAY(anytype,shape)
This function creates an unmasked array of the given type and shape. The shape is given in the last argument(s). It can be given in two ways:
- As a single set argument; for example, array(0,[3,4])
- As individual scalar arguments; for example, array(0,3,4)
The first argument gives the values the array is filled with. It can be a scalar or an array of any shape. To initialize the created array, the value array is flattened to a 1D array. Its successive values are stored in the created array. If the new array has more values than the value array, the value array is reset to its beginning and the process continues.
Note that a masked array can be created from an (unmasked) array and a mask using the brackets operator like ARRAY[MASK].

anytypearray MARRAY(anytypearray,boolarray)
This function offers another way to create a masked array. The mask must be given in the second argument; its shape must be the same as the shape of the data array.

anytypearray NULLARRAY(anytype)
This function creates a null array. Its data type is determined by the data type of the argument. The argument value itself is not used. It is mainly meant for test purposes.

anytypearray RESIZE(anytypearray,shape[,mode])
This function resizes an array to the given shape and copies the values. The optional mode argument determines how the values are copied. If the argument is not given, the new shape is arbitrary and the dimensionality can change. The values are copied to the same index in the new array. If an axis gets larger, the new values are set to 0 (or an empty string).
If mode is given, the new shape is much more restricted. The dimensionality cannot change and each axis has to be a multiple of the old one. If mode=0, copying the values is done in an upsampling way. E.g., if a new axis is twice the length of the old one, values 1,2,3 are copied as 1,1,2,2,3,3. A good use case is applying the flags of averaged data to the original data. If mode=1, the values in the example above are copied repeatedly as 1,2,3,1,2,3. By giving the mode as a set, it is possible to specify the mode per axis, but that is quite esoteric.

anytypearray TRANSPOSE(anytypearray[,axes])
This function transposes an N-dim array. If no axes are given, the array is fully transposed (thus all axes are reversed). Axes can be specified meaning that those axes will become the first axes in the output array. Non-given axes follow thereafter in their natural order.
A possible mask is transposed as well.

anytypearray DIAGONAL(anytypearray[,firstaxis[,diag]])
This function takes the diagonal of 2-dim subarrays in an N-dim array resulting in an array with 1 dimension less. For a 2-dim array, it is simply the diagonal of the matrix. For a higher dimensional array, it takes the diagonal of each matrix defined by firstaxis and firstaxis+1. e.g., in a 3-dim array the diagonals of each XY-plane can be taken. The default for firstaxis is 0.
The diag argument tells which diagonal has to be taken. The default 0 means the main diagonal. A negative value means below the main diagonal, while positive means above the main diagonal.

anytypearray ARRAYDATA(anytype)
This function returns the array without a mask, thus removes the mask. If the operand is a scalar, it returns a 1-dim array with one element.

anytypearray NEGATEMASK(anytype)
This function returns the array with the negated mask. If the array has no mask, it returns the array with a mask of all Trues. If the operand is a scalar, it returns a 1-dim array with one element.

anytypearray REPLACEMASKED(anytype, anytype)
This function replaces the masked elements in the first argument by the corresponding value in the second argument (which can be a scalar value). If the first argument has no mask, the function is a no-op.

anytypearray REPLACEUNMASKED(anytype, anytype)
This function replaces the unmasked elements in the first argument by the corresponding value in the second argument (which can be a scalar value). If the first argument has no mask, the function is a no-op.

boolarray ARRAYMASK(anytype), boolarray MASK(anytype)
This function returns the mask of an array. If the array has no mask, it returns a boolean array of the same shape with all values set to False. If the operand is a scalar, it returns an 1-dim array with one False element.

anytypearray FLATTEN(anytype)
This function flattens an N-dim array to a 1-dim array keeping the unmasked elements only. If the operand is a scalar, it returns a 1-dim array with one element.

4.10.13 Aggregate functions

The GXXX aggregate functions calculate an aggregated value for all rows in a group, usually defined with a GROUPBY clause. For example, when grouping in TIME, an aggregate function like GNTRUE(FLAG) counts per time slot the number of flagged data points. Aggregate functions can only be used in the SELECT and the HAVING clause.
Most functions listed below reduce the values in a group to a scalar value, also if the value in a row is an array (as in the GNTRUE example above). The arrays in a group can have different shapes.

However, there are several aggregate functions returning an array as done by the last three functions (GHIST, GAGGR, and GROWID) shown below. Furthermore, most scalar functions have a plural form (e.g., GNTRUES) returning an array. They are described at the end of this section.

Note that the aggregate function names differ from their SQL counterparts; they all have the prefix G, because TaQL functions like MAX already exist for array operations. This naming scheme also makes it more clear which TaQL functions are aggregate functions.

A technical detail is how aggregate functions are implemented. TaQL walks sequentially through a table. Non-lazy functions operate directly on the value in a row making the table access purely sequential. It requires that the results of all groups are held in memory. For some functions, in particular GAGGR, this could lead to a very high memory usage. Therefore, some functions are implemented in a lazy way. They keep the row numbers of a group and access the data when the aggregated result of a group is needed. In this way only the data of a single group needs to be held in memory, but the access to the table might be non-sequential making it somewhat slower. Currently, only GAGGR and the User Defined aggregate functions are implemented in a lazy way.

integer GCOUNT(), integer GCOUNT(*)
Return the number of rows per group.

integer GCOUNT(columnname)
Return the number of rows per group for which the column has a value. Note that only a column containing variable sized arrays can contain empty cells.

anytype GFIRST(anytype)
Return the first value of an expression in the group. The values of a column not mentioned in the GROUPBY clause, might differ. This function can be used to return the value of the first row in the group.

anytype GLAST(anytype)
Return the last value of the group (is similar to GFIRST).
Note this function is implicitly used if an expression without aggregate function is used in a group.

bool GANY(bool)
Is any element true?

bool GALL(bool)
Are all elements true?

integer GNTRUE(bool)
Return number of true elements.

integer GNFALSE(bool)
Return number of false elements.

numeric GSUM(numeric)
Return sum of all elements.

numeric GSUMSQUARE(numeric), numeric GSUMSQR(numeric)
Return sum of all squared elements.

numeric GPRODUCT(numeric)
Return product of all elements.

real GMIN(real)
Return minimum of all elements.

real GMAX(real)
Return maximum of all elements.

dnumeric GMEAN(numeric), dnumeric GAVG(numeric)
Return mean of all elements.

double GVARIANCE(real)
Return variance (the sum of
(a(i) - mean(a))**2/(nelements(a) - 1).

double GSTDDEV(real)
Return standard deviation (the square root of the variance).

double GAVDEV(real)
Return average deviation. (the sum of
abs(a[i] - mean(a))/nelements(a).

double GRMS(real)
Return root-mean-squares. (the square root of the sum of
(a(i)**2)/nelements(a).

double GMEDIAN(real)
Return median (the middle element). If the array has an even number of elements, the mean of the two middle elements is returned.

double GFRACTILE(real, doublescalar fraction)
Return the value of the element at the given fraction. Fraction 0.5 is the same as the median.

double GHIST(real, intscalar nbin, realscalar start, realscalar end)
Return the histogram of the data using the given number of bins. The histogram contains an extra bin at the beginning and the end for the outliers. If the rows in the group contain arrays, they can have variable shapes.

anytypearray GAGGR(anytype), anytypearray GSTACK(anytype)
Stack the row values in a group to form an array where the row is the slowest varying axis (similar to numpy’s dstack). Thus if the column contains scalar values, the result is a vector. Otherwise it is an array whose dimensionality is one higher. It requires that all arrays in a group have the same shape.
Note that this function can be very useful for arrays, because it makes it possible to use partial reduce functions like medians to calculate the medians along arbitrary axes.

integerarray GROWID()
Return the row numbers of the rows in the group.

Most functions above have a plural counterpart. They calculate the aggregated value per array index, thus the result has the same shape as the arrays in the group. Similar to function GAGGR, they require that all arrays in a group have the same shape.
For instance, for a MeasurementSet the expression GMEANS(DATA) calculates the mean in a group per channel/polarization. Not only it is a shorthand for MEANS(GAGGR(DATA), 0), but it usually works faster because, unlike GAGGR, it is non-lazy.
The functions available are:

4.10.14 Miscellaneous functions

bool ISNULL(anytype)
Return True if the argument value is a null array.

bool ISDEFINED(anytype)
Return False if the array value in the current row is undefined (is null). It makes it possible to test if a cell in a column with variable shaped arrays contains an array. Furthermore, it can be used to test if a field in a record is defined.
Note that function ISNULL can also be used to test for an undefined array in a row.

bool sh.ISCOLUMN(string)
Return False if no column with the given name exists in the table with the shorthand given before the function name. If no table is given, the first one will be used.

bool sh.ISKEYWORD(string)
Return False if no keyword with the given name exists in the table with the shorthand given before the function name. If no table is given, the first one will be used. The keyword name can be given as described in section 4.5, thus the name of a table keyword or column keyword or a nested field can be specified.

integer NELEMENTS(anytype), integer COUNT(anytype)
Return number of elements in an array (1 for a scalar).

integer NDIM(anytype)
Return dimensionality of an array (0 for a scalar).

integerarray SHAPE(anytype)
Return shape of an array (returns an empty array for a scalar).

integer ROWNUMBER(), integer ROWNR()
Return the row number being tested (first row is row number 0 or 1 depending on the style used).
In combination with function RAND it can, for instance, be used to select arbitrary rows from a table.

integer ROWID()
Return the row number in the original table. This is especially useful for returning the result of a selection of a subtable of a Casacore measurement set (see also subqueries in 4.11 and examples in section 12.1).

double RAND()
Return (per table row) a uniformly distributed random number between 0 and 1 using a Multiplicative Linear Congruential Generator. The seeds for the generator are deduced from the current date and time, so the results are different from run to run.
The function can, for instance, be used to select a random subset from a table.

double ANGDIST(arg1,arg2), double ANGULARDISTANCE(arg1,arg2)
Return the angular distance (in radians) between the positions in arg1 and arg2. Both arguments have to be numeric arrays containing an even number of values. Two subsequent values give the RA and DEC (or longitude and latitude) of positions on a sphere. The result is a 1-dim array containing the angular distance between corresponding positions in arg1 and arg2. If either array contains only one position, the result is the distance between that position and each position in the other array. If both arguments contain only 2 values, the result is a scalar. For example:
angdist(PHASE_DIR[0,], [12h13m45,4d21m39.4, 12h13m49,10d8m4])
returns an array with shape [2] containing the angular distance between the phase center of the field and the two positions given.

double ANGDISTX(arg1,arg2), double ANGULARDISTANCEX(arg1,arg2)
Same as above, but the result is a 2-dim array giving the distance between each position in the first argument and each position in the second argument. Only if both arguments contain a single position, the result is a scalar.

anytype IIF(cond,arg1,arg2)
This is a special funtion which operates like the ternary ?: operator in C++. If all arguments are scalars, the result is a scalar, otherwise an array. In the latter case possible scalar arguments are virtually expanded to arrays. IIF evaluates the condition for each element. If True, it takes the corresponding element of arg1, otherwise of arg2.
If one of the input arrays has a mask, the output array will also have a mask. Each output mask element value is the logical OR of the condition mask element value and the mask value of the element taken from arg1 or arg2.

4.10.15 Cone search functions

Cone search functions make it possible to test if a source is within a given distance of a given sky position. The expression

could be used to test if sources with their sky position defined in columns RA and DEC are within 1 arcmin of the given sky position.
The cone search functions implement this expression making life much easier for the user. Because they can also operate on arrays of positions, searching in multiple cones can be done simultaneously. That makes it possible to find matching source positions in two catalogues as shown in an example at the end of this section.

The arguments of all functions are described below. All of them have to be given in radians. However, usually one does not need to bother because TaQL makes it possible to specify positions in many formats automatically converted to radians.

SOURCES
is a set or array giving the positions of one or more sources (e.g., in equatorial coordinates) to be tested. Normally these are columns in a table. Where argument name SOURCE is mentioned below, only a single source can be used, otherwise multiple sources.
For example:
[RA,DEC] for scalar columns RA and DEC.
SKYPOS for a column SKYPOS containing 2-element vectors with RA and DEC.

CONES
is a set or array giving the center positions and radii of one or more cones (e.g., as RA,DEC,radius). Usually the user will specify it as constants.
For example:
[12h13m54, -5.3.34, 0d1m] for a single cone.
[12h13m54, -5.3.34, 0d1m, 1h2m3, 4.5.6, 0d1m] for two cones.

CONEPOS
is a set or array giving the positions of one or more cone centers (e.g., as RA,DEC).

RADII
is a scalar, set or array giving one or more radii. Each radius is applied to all positions in CONEPOS. Specifying a cone as CONEPOS,RADIUS is easier than specifying it as CONES if the same radius has to be used for multiple cones.
For example:
[12h13m54, -5.3.34, 1h2m3, 4.5.6], 0d1m is the same as the second CONES example above.

The following cone search functions are available.

bool ANYCONE(SOURCE,CONES)
Return T if the source is contained in at least one of the cones. Operator INCONE is a synonym. So ANYCONE(SOURCE,CONES) is the same as SOURCE INCONE CONES.

bool ANYCONE(SOURCE,CONEPOS,RADII)
It does the same as above.

integer FINDCONE(SOURCES,CONES)
Return the index of the first cone containing the source. If a single source is given, the result is a scalar. If multiple sources are given, the result is an array with the same shape as the source array.

integer FINDCONE(SOURCES,CONEPOS,RADII)
It does the same as above. Note that in this case each radius is applied to each cone, so the resulting index array is a combination of the two input arrays (with the radius as the most rapidly varying axis).

bool CONES(SOURCES,CONES)
Return a 2-dim bool array. The length of the most rapidly varying axis is the number of cones. The length of the other axis is the number of sources. When using python style, element (i,j) in the resulting array is T if source i is contained in cone j.

bool CONES(SOURCES,CONEPOS,RADII)
It does the same as above. However, the result is a 3-dim array with the radii as the most rapidly varying axis, cones as the next axis, and sources as the slowest axis.

Please note that ANYCONE(SOURCE,CONES) does the same as any(CONES(SOURCE,CONES)), but is faster because it stops as soon as a cone is found.
Function CONES makes it possible to do catalogue matching. For example, to find sources matching other sources in the same catalogue (within a radius of 10 arcseconds):

Note that in this example the SELECT clause returns an array with positions which are used as the cone centers. So each source in the catalogue is tested against every source. It makes it an N-square operation, thus potentially very expensive. The result is a 4-dim boolean array with shape (in glish style) [1,nrow,1,nrow] which can be processed in Glish. Please note that the CONES function results for each row in a array with shape [1,nrow,1].
The query can be done with multiple radii, for example also with 1 arcsecond and 1 arcminute.

resulting in an array with glish shape [3,nrow,1,nrow]. In this way one can get a better indication how close sources are to the cone centers.

4.10.16 User defined functions

TaQL can be extended with so-called User Defined Functions (UDF). These are dynamically loaded functions, either written in C++ or in Python. In TaQL the name of a UDF written in C++ consists of the name of the library (without lib prefix and extension) followed by a dot and the function name. For example:

denotes function hadec in shared library libmeas.so or libcasa_meas.so. For OS-X the extension .dylib will be used.
The physical shared library name must be fully lowercase, but the UDF name used in TaQL is case-insensitive. The name of a UDF written in Python is like py.module.func where the module part is optional. In the USING STYLE clause it is possible to define synonyms for the UDF library names. By default, mscal is defined as a synonym for derivedmscal and py as a synonym for pytaql.

Usually a UDF will operate on the arguments given to the function and will not itself operate on a table given in a query command. However, some UDFs (most notably the mscal ones) do not have arguments, but operate directly in a specific way on a table. Normally they use the first table given in the FROM clause, but the UDF name can be preceded by a table shorthand to specify another table. For example:

to get the hourangle from two different tables. Of course, both tables need to have the same number of rows.
Note that UDFs not directly operating on a table, will ignore a shorthand.

In section Writing user defined functions it is explained how to write user defined functions.

4.10.17 Special MeasurementSet functions

The Casacore package comes with several predefined UDFs in library libcasa_derivedmscal. It contains four groups of UDFs, all operating on a MeasurementSet and several on a CASA calibration table (both old and new format).
Although the library is called derivedmscal, for ease of use it is possible to use the synonym mscal.

Get derived values
The first group calculates derived values like hourangle and azimuth for the first or second antenna of a baseline. For calibration tables, where a row contains a single antenna, functions like PA1 are the same as PA2. All angles are returned in radians.

double MSCAL.HA()
gives the hourangle of the array center (observatory position).

double MSCAL.HA1()
gives the hourangle of ANTENNA1.

double MSCAL.HA2()
gives the hourangle of ANTENNA2.

double MSCAL.HADEC()
gives the topocentric hourangle/declination of the array center (observatory position).

double MSCAL.HADEC1()
gives the topocentric hourangle/declination of ANTENNA1.

double MSCAL.HADEC2()
gives the topocentric hourangle/declination of ANTENNA2.

doublearray MSCAL.AZEL()
gives the topocentric azimuth/elevation of the array center (observatory position).

doublearray MSCAL.AZEL1()
gives the topocentric azimuth/elevation of ANTENNA1.

doublearray MSCAL.AZEL2()
gives the topocentric azimuth/elevation of ANTENNA2.

double MSCAL.LAST()
gives the local sidereal time of the array center.

double MSCAL.LAST1()
gives the local sidereal time of ANTENNA1.

double MSCAL.LAST2()
gives the local sidereal time of ANTENNA2.

double MSCAL.PA1()
gives the parallactic angle of ANTENNA1.

double MSCAL.PA2()
gives the parallactic angle of ANTENNA2.

doublearray MSCAL.NEWUVW()
gives the 3-vector of UVW coordinates in J2000 in meters. It recalculates them, thus does not return the UVW coordinates stored in the MeasurementSet.

doublearray MSCAL.NEWUVWWVL()
gives the 3-vector of calculated UVW coordinates in J2000 in wavelengths for the reference frequency of the appropriate spectral window.

doublearray MSCAL.NEWUVWWVLS()
gives the nfreq*3-matrix of calculated UVW coordinates in J2000 in wavelengths for all channel frequencies of the appropriate spectral window.

doublearray MSCAL.UVWWVL()
gives the 3-vector of stored UVW coordinates in wavelengths for the reference frequency of the appropriate spectral window.

doublearray MSCAL.UVWWVLS()
gives the nfreq*3-matrix of stored UVW coordinates in wavelengths for all channel frequencies of the appropriate spectral window.

By default all these functions will use the direction given in column PHASE_DIR of the FIELD subtable. It is possible to use another column by giving its name as a string argument (e.g., HA(’DELAY_DIR’)).
Except for the last 2 functions, it is possible to use an explicit direction which must be given as [RA,DEC] in J2000 or as a case-insensitive name of a planetary object (as defined by the Casacore Measures). For example:

The examples above give the azimuth and elevation of the given directions for each selected row in the MeasurementSet, using the position of ANTENNA1 and the times in these rows.

If a string value is given, it is first tried as a planetary object. Theoretically it is possible that a column has the same name as a planetary object. In such a case the name can be escaped by a backslash to indicate that a column name is meant. For example:

means that column SUN in the FIELD table has to be used.

Stokes conversion
The STOKES function makes it possible to convert the Stokes parameters of a DATA column in a MeasurementSet, for instance from linear or circular to iquv. It is also possible to convert the weights or flags, i.e., to combine them in the same way as the data would be combined.

complexarray MSCAL.STOKES(complexarray, string)
converts the data.

doublearray MSCAL.STOKES(doublearray, string)
combines the weights.

boolarray MSCAL.STOKES(boolarray, string)
combines the flags.

In all cases the case-insensitive string argument defines the output Stokes axes. It must be a comma separated list of Stokes names. All values defined in the Casacore class Stokes are possible. Most important are:

• XX, XY, YX, and/or YY.
  LINEAR or LIN means XX,XY,YX,YY.
• RR, RL, LR, and/or LL.
  CIRCULAR or CIRC means RR,RL,LR,LL.
• I, Q, U, and/or V.
  IQUV or STOKES means I,Q,U,V.
• PTOTAL is the polarized intensity (sqrt(Q**2+U**2+V**2))
• PLINEAR is the linearly polarized intensity (sqrt(Q**2+U**2))
• PFTOTAL is the polarization fraction (Ptotal/I)
• PFLINEAR is the linear polarization fraction (Plinear/I)
• PANGLE is the linear polarization angle (0.5*arctan(U/Q)) (in radians)

If not given, the string argument defaults to ’IQUV’. For example:

creates a table with column CIRCDATA containing the circular polarization data.

CASA style selection
The BASELINE function makes it possible to do selection on baselines in a MeasurementSet or CalTable using the special CASA selection syntax described in note 263. Similar functions CORR, TIME, FIELD, FEED, SCAN, SPW, UVDIST, STATE, OBS, and ARRAY can be used to do selection based on other meta data. The functions accept a string containing a selection string and return a Bool value telling if a row matches the selection string. For example,

selects the cross-correlation baselines containing an antenna whose name matches the pattern in the function argument.
Note there is a difference how CASA and TaQL handle unknown antennas given in the baseline selection string. CASA tasks give an error, while TaQL will not complain and not even report it, because doing a selection this way should not behave differently from doing it like NAME=’RTX’.
Also note that in CASA tasks only one selection string per type can be given and the final selection is the AND of them. TaQL has the AND and OR operators making it possible to combine the selections in all kind of ways, possibly using multiple selection strings of the same type.

Get values from a subtable
Several functions exist to get information like the name of an antenna from the subtable for each row in the main table. Basically they do a join of the main table and a subtable. For example:

gets the names of the antennae used in each baseline.

The following functions can be used:

string MSCAL.ANT1NAME()
gives the name of ANTENNA1.

string MSCAL.ANT2NAME()
gives the name of ANTENNA2.

anytype MSCAL.ANT1COL(ColumnName)
gives for ANTENNA1 the value in the given column (in quotes) in the ANTENNA subtable.

anytype MSCAL.ANT2COL(ColumnName)
gives for ANTENNA2 the value in the given column (in quotes) in the ANTENNA subtable.

anytype MSCAL.STATECOL(ColumnName)
gives for STATE_ID the value in the given column (in quotes) in the STATE subtable.

anytype MSCAL.OBSCOL(ColumnName)
gives for OBSERVATION_ID the value in the given column (in quotes) in the OBSERVATION subtable.

anytype MSCAL.SPWCOL(ColumnName)
gives for DATA_DESC_ID the value in the given column (in quotes) in the SPECTRAL_WINDOW subtable.

anytype MSCAL.POLCOL(ColumnName)
gives for DATA_DESC_ID the value in the given column (in quotes) in the POLARIZATION subtable.

anytype MSCAL.FIELDCOL(ColumnName)
gives for FIELD_ID the value in the given column (in quotes) in the FIELD subtable.

anytype MSCAL.PROCCOL(ColumnName)
gives for PROCESSOR_ID the value in the given column (in quotes) in the PROCESSOR subtable.

anytype MSCAL.SUBCOL(SubtableName, ColumnName, idcolumn)
gives for the (integer) id-column the value in the given column in the given subtable. This is the most common form and can be used to join any table with a subtable.

Note that the following are equivalent. The first versions are shorthands for the latter ones.

In the last example the id-column must be given as such, thus must not be a string.

4.10.18 Special Measures functions

These functions make it possible to convert measures like directions, epochs, and positions from one reference frame to another. All conversions supported by Casacore’s Measures are possible. For example:

The first example converts a J2000 position to galactic coordinates. The second example gives the moon’s azimuth/elevation at the WSRT at the current date/time.

The following basic functions are available. Most functions return double angle values with unit rad. Only the RISESET function returns date/time values.
Note that all names used below are case-insensitive.

doublearray MEAS.DIR(toref, direction, epoch, position)
converts a direction to the reference type given by the ’toref’ string. The epoch and position arguments only need to be given if the conversion needs frame information (e.g., when converting J2000 to apparent).
Function name MEAS.DIRECTION can be used as well.

doublearray MEAS.DIRCOS(toref, direction, epoch, position)
Same as function DIR, but returning 3 direction cosines instead of 2 angles.
Function name MEAS.DIRECTIONCOSINE can be used as well.

DateTimearray MEAS.RISESET(direction, epoch, position)
returns the rise and set date/times (UTC) of the sources given in the direction argument for the given dates and positions. Function name MEAS.RISET can be used as well. Note that the source can be invisible all day (results in set$<$rise). If visible all day, rise time is 0h0m and set time is 24 hours later.
The TIME or CTIME function can be used on the result to get the time part only (as double or string).
If the sun is used as a source name (case-insensitive), it can be followed by a hyphen and one of the following case-insensitive suffices indicating what to use. The default is CR which is used in most almanacs.

• CR: use the center of the sun with refraction correction.
• UR: use the upper brim of the sun with refraction correction, thus show when part of the sun is visible.
• LR: use the lower brim of the sun with refraction correction, thus show when the full sun is visible.
• C: use the center of the sun without refraction correction.
• U: use the upper brim of the sun without refraction correction.
• L: use the lower brim of the sun without refraction correction.
• CT: use the civil twilight (6 deg). without artifical light.
• NT: use the nautical twilight (12 deg). visible for navigating.
• AT: use the amateur astronomical twilight (15 deg).
• ST: use the scientific astronomical twilight (18 deg).

The first six suffices can also be used with the Moon.
See stjarnhimlen.se for additional information.

doublearray MEAS.EPOCH(toref, epoch, position)
converts an epoch to the reference type given by the ’toref’ string. The position argument only needs to be given if the conversion needs frame information. By default conversions to sidereal time (e.g, LAST) return the fraction giving the true sidereal time. Only if the toref string starts with ’F-’, ’F_’, ’f-’, or ’f_’ the full sidereal time is returned which includes the number of sidereal days since the start of MJD.

doublearray MEAS.POS(toref, position)
converts a position to the reference type given by the ’toref’ string.
Function name MEAS.POSITION can be used as well.

The available reference types can be shown using the show meastypes command in the program taql. For ease of use several specialized MEAS.DIR and MEAS.POS functions are defined with an implicit ’toref’ argument.

doublearray MEAS.J2000(direction, epoch, position)
converts a direction to J2000.

doublearray MEAS.B1950(direction, epoch, position)
converts a direction to B1950.

doublearray MEAS.APP(direction, epoch, position)
converts a direction to apparent coordinates.
Function name MEAS.APPARENT can be used as well.

doublearray MEAS.HADEC(direction, epoch, position)
converts a direction to hourangle/declination.

doublearray MEAS.AZEL(direction, epoch, position)
converts a direction to azimuth/elevation.

doublearray MEAS.ECL(direction, epoch, position)
converts a direction to ecliptic coordinates.
Function name MEAS.ECLIPTIC can be used as well.

doublearray MEAS.GAL(direction, epoch, position)
converts a direction to galactic coordinates.
Function name MEAS.GALACTIC can be used as well.

doublearray MEAS.SGAL(direction, epoch, position)
converts a direction to supergalactic coordinates.
Function name MEAS.SUPERGAL or MEAS.SUPERGALACTIC can be used as well.

doublearray MEAS.ITRFD(direction, epoch, position)
converts a direction to ITRF coordinates.
Function name MEAS.ITRFDIR or MEAS.ITRFDIRECTION can be used as well.

doublearray MEAS.LAST(epoch, position)
converts an epoch to local sidereal time.
Function name MEAS.LST can be used as well.

doublearray MEAS.ITRFxxx(position)
converts a position to ITRF coordinates.

doublearray MEAS.WGSxxx(position)
converts a position to WGS84 coordinates.

The names of the last two functions have a suffix xxx indicating how positions are returned.
- XYZ means as x,y,z
- LL or LONLAT means as lon,lat
- H or HEIGHT means as height
It defaults to XYZ.

The function arguments can be given in a variety of ways.

• ’toref’ is a constant scalar string giving the reference type to convert to. See the Measure classes MDirection, MEpoch, and MPosition, for an overview of the types.
• ’direction’ gives one or more directions to convert. They can be given in several ways.
  • As a constant scalar or array of strings giving one or more planetary objects like MOON or VENUS and/or giving the name of standard sources (CasA, CygA, TauA, VirA, HerA, HydA, or PerA). In the future support for comets might be added.
    The names are case-insensitive.
  • As 2 constant double scalar arguments giving ra and dec (or longitude and latitude).
  • As a double array with an even number of elements giving ra/dec or longitude/latitude of one or more directions. It can be a constant array (expression), but it can also be a column or an expression using a column.

  If a column or a column slice is given, the reference type stored in the column keywords will be recognized. In other cases the input reference type should be given in the next string argument. If not given, it defaults to J2000.
  For example:

    [’MOON’,’sun’, ’venus’]          # 3 planetary objects  
    12h23m17.5, 23d56m43.8, ’B1950’  # ra/dec as scalar constants (as B1950)  
    [12h23m17.5, 23d56m43.8]         # ra/dec as array (default J2000)  
    PHASE_DIR[0,]                    # direction ra/dec in given column

• ’epoch’ gives one or more epochs to use. Similar to directions the reference type is taken from the column keywords or can be given in the next argument. It defaults to UTC.
  Epochs can be given in three ways:
  • As a scalar or array containing double values. It can be a constant expression or a column (expression).
  • As a scalar or array containing DateTime values.
  • As a scalar or array containing String values representing date/time. They will automatically be converted to DateTime values using function datetime.

  For example:

    datetime()                         # current date/time  
    ’today’                            # current date/time  
    [select unique TIME from my.ms]    # all times from some MS  
    9Sep2011/12:00:00, ’UTC’           # given UTC time

  Note that in the last example ’UTC’ is not necessary, because it is the default.

• ’position’ gives one or more directions to use. They can be given as x,y,z or as lon,lat with an optional height. Usually the unit of the first value defines if x,y,z or lon,lat is used. It is, however, also possible to distinguish between LL and XYZ by using suffices like XYZ or LL in the reference type given in the next argument.
  • As a scalar or array of observatory names using their positions in the Measures Observatory table.
  • As 2 or 3 constant double scalar values giving xyz, lonlat, or lonlat/height.
  • As a double array giving one or more positions in xyz or lonlat. Similar to directions it can be a column (expression) where the reference type is taken from the column keywords.
  • As two constant double arrays giving lonlat and height of one or more positions. The array sizes have to match (thus the size of the lonlat array must be twice the size of the height array).

  If needed, the reference type (with optional suffix) can be given in the next argument. The reference type defaults to ITRF if xyz coordinates are used, otherwise to WGS.
  For example:

    ’WSRT’                             # WSRT position  
    5deg, 52deg                        # 2 scalar constants (WGS84 lonlat)  
    (5deg, 52deg]                      # same, but as array  
    5deg, 52deg, 5m                    # WGS84 lonlat with height  
    [5deg, 52deg], [5m]                # same, but as array  
    3.8288e+06m, 442449, 5.0649e+06    # xyz as scalars (ITRF)  
    [41.84m, 4.835, 55.722], ’WGS’     # xyz as array (WGS84)  
    POSITION                           # POSITION column

A few more elaborate examples are given below.

calculates the local apparent sidereal time for the given date and position.

calculates Jupiter’s azimuth/elevation for WSRT and VLA for all times returned by the subquery (see next section for subqueries).

converts the PHASE_DIR directions in the FIELD table to B1950. Note that no frame information is needed for such a conversion.

calculates the azimuth/elevation (in degrees) of the given source direction for the LOFAR site for 24 hours. The result is an array with shape [24,2]. The direction in the second example is given in B1950, the first as the default J2000.

4.11 Subqueries

As in SQL it is possible to create a set from a subquery. A subquery has the same syntax as a main query, but has to be enclosed in square brackets or parentheses. Basically it looks like:

The subquery on othertable results in a constant set containing the times for which the windspeed matches. Subsequently the main query is executed and selects all rows from the main table with times in that set. Note that like other bounded sets this set is transformed to a constant array, so it is possible to apply functions to it (e.g., min, mean).

This example shows how a subquery is used to join the main table of a MeasurementSet and its ANTENNA subtable. The subquery returns a list with the names of all antennae, which subsequently is indexed with the antenna number to get the antenna name for each row in the main table.

is a newer and easier way to obtain the name of ANTENNA1. It makes use of the new user defined functions in derivedmscal which can do an implicit join of a MeasurementSet and its subtables.

This example contains another subquery to get all windspeeds and to take the mean of them. So the first subquery selects all times where the windspeed is less than the average windspeed.
A subquery result should contain only one column, otherwise an exception is thrown.

It may happen that a subquery has to be executed twice because 2 columns from the other table are needed. E.g.

In this case the other table contains the time range for each windspeed. For big tables it is expensive to execute the subquery twice. A better solution is to store the result of the subquery in a temporary table and reuse it.

However, this has the disadvantage that the table tmptab still exists after the query and has to be deleted explictly by the user. Below a better solution for this problem is shown.

TaQL has a few extensions to support tables better, in particular the Casacore MeasurementSets.

1. The temporary problem above can be circumvented by using the ability to use a SELECT expression in the FROM clause. E.g.
     SELECT FROM maintable,  
         [SELECT FROM othertable WHERE windspeed < 5] tmptab  
         WHERE any(time >= [SELECT starttime FROM tmptab]  
                && time <= [SELECT endtime FROM tmptab])

   However, below an even nicer solution is given.

2. The time range problem above can be solved elegantly by using a set as the result of the subquery. Instead of a table name, it is possible to give an expression in the GIVING clause (as mentioned in section 3.10). E.g.
     select from MY.MS where TIME in  
         [select FROM OTHERTABLE where WINDSPEED < 5  
              giving [TIME-INTERVAL/2 =:= TIME+INTERVAL/2]]

   The set expression in the GIVING clause is filled with the results from the subquery and used in the main query. So if the subquery results in 5 rows, the resulting set contains 5 intervals. Thereafter the resulting intervals are sorted and combined where possible. In this way the minimum number of intervals have to be examined by the main query.

3. In Casacore the other table will often be the name of a subtable, which is stored in a table or column keyword of the main table. The standard keyword syntax can be used to indicate that the other table is the table in the given keyword. Note that for a table keyword the :: part has to be given, otherwise the name is treated as an ordinary table name. E.g.
     select from MY.MS where TIME in  
         [select TIME from ::WEATHER where WINDSPEED < 5]

   In this example the other table is a subtable of table my.ms. Its name is given by keyword WEATHER of my.ms.

4. Often the result of a query on a subtable of a measurement set is used to select columns from the main table. However, several subtables do not have an explicit key, but use the row number as an implicit key. The function ROWID() can be used to return the row number as the subtable query result. E.g.
     select from MY.MS where DATA_DESC_ID in  
         [select from ::DATA_DESCRIPTION where  
            SPECTRAL_WINDOW_ID in [0,2,4] giving [ROWID()]]

   Note that the function ROWNUMBER cannot be used here, because it will give the row number in the selection and not (as ROWID does) the row number in the original table. Furthermore, ROWID gives a 0-relative row number which is needed to be able to use it as a selection criterium on the 0-relative values in the measurement set.

5. Select if any channel has a UV distance $<$ 100 wavelengths.
     select from MY.MS where any(sqrt(sumsqr(UVW[:2])) / c() *  
         [select CHAN_FREQ from ::SPECTRAL_WINDOW][DATA_DESC_ID,]  
            < 100)

   In a MeasurementSet the UVW coordinates are stored in meters, so they have to be multiplied with the frequency and divided by speed of light to get them in wavelengths.
   Because TaQL has no proper join operation, it is not possible to select directly on the DATA_DESC_ID. However, using a nested query and indexing the result with the DATA_DESC_ID has the same effect. It only requires that CHAN_FREQ has the same length in all rows in the subtable.
   Using the new derivedmscal functions, below a much nicer solution is given.

6.   select from MY.MS where any(mscal.uvwwvls() < 100)

   It shows how the UVWWVLS function in derivedmscal can be used to obtain the UVW coordinates in wavelengths.

7. Calculate the angular distance between the Mars and Jupiter as seen from the WSRT for the coming 30 days.
     calc angdist(meas.app(’mars’,    date()+[0:31], ’WSRT’),  
                  meas.app(’jupiter’, date()+[0:31], ’WSRT’))

5 Aggregation, GROUPBY, HAVING

Similar to SQL it is possible to do aggregation and grouping in TaQL and to do selection on the groups using the HAVING clause.

5.1 Aggregation and GROUPBY

One or more aggregated values can be calculated for a group defined by the GROUPBY clause. The aggregate functions described in section 4.10.13 can be used. For example:

A group is formed for the unique values of the columns given in the GROUPBY clause. In the example above a group per baseline is formed. Usually an aggregate function is ued to calculate a value for the group. In the example above the aggregate function gcount() counts the number of rows per baseline.
Often only the GROUPBY columns and aggregated values are part of the SELECT clause, but the example shows that other values (here the baseline length) can also be selected. Non-aggregated values get the values in the last row of a group.

Usually aggregated values and GROUPBY are used jointly, but it is possible to leave out one of them. If GROUPBY is not given, the entire table is a single group. For example:

does not have groups, thus shows the total number of rows in the MS.

does not use aggregate functions, but shows the unique baselines in the MS. Apart from the order, it has the same result as

but is somewhat faster.

In the examples above a sole aggregate function is used, but it is also possible to use it in an expression. Similarly, an expression can be used in the GROUPBY. For example:

groups the MS in chunks of 5 time slots. Note that the nested query gets the TIME of the first time slot. The result is a set, hence the 0th element has to be taken.

Note that an aggregate function can only be used in the SELECT and HAVING clause, so TaQL will give an error message if used elsewhere.

5.2 HAVING

The HAVING clause can be used to select specific groups. For example:

groups by time, but only selects the groups for which the maximum amplitude of the DATA is more than 100. Both examples give the same result, but the first one is more efficient. Not only it is less typing, but it is faster because it reuses the result column MAXA of the SELECT part.
Similar to WHERE, any expression can be used in HAVING, but the result has to be a bool scalar value.

As shown in the example, HAVING will normally use aggregate functions, but it is not strictly needed. However, selections without an aggregate function could as well be done in the WHERE clause.
Usually HAVING will be used in combination with GROUPBY, but it can be used without. It can also be used without an aggregate function in the SELECT. However, it is an error if both are omitted.

6 Some further remarks

6.1 Joining tables

As discussed in some previous sections it is possible to join tables on row number. Two examples show how to do it.

6.1.1 Join on row number

This command can be used to check if the data in mytable is about equal to the data in othertable. Both tables have to have the same number of rows.
The join is done on row number, thus the data in corresponding rows are compared.

6.1.2 Join using an indexed subquery

This example shows how a subquery is used to join the main table of a MeasurementSet with its ANTENNA subtable. The subquery returns a list with the names of all antennae, which subsequently is indexed with the antenna number to get the antenna name for each row in the main table.
The join is done using the ANTENNA1 column which gives the row number in the subtable, thus the index in the subquery result.

6.1.3 Join using a subquery set

This example shows another way to use a subquery for a join of the main table of a MeasurementSet with its ANTENNA subtable. It selects all baselines for which the first station is a core station. The subquery returns a set containing the ids of the core stations, which is used to select the correct stations in the main table.

6.1.4 Join using derivedmscal

Several UDFs in the derivedmscal library make it possible to easily join a MeasuementSet or CASA Calibration Table with a subtable like ANTENNA or SPECTRAL_WINDOW. These functions know which columns to use making the join straightforward like in

The library also contain the more general SUBCOL function making it possible to join any table with a subtable. For example:

to get the parameter name for a LOFAR ParmDB table. A ParmDB table has a subtable NAMES containing the NAME and other info of a parameter. The column NAME_ID is used to reference that subtable.

6.2 Optimization

A lot of development work could be done to improve the query optimization. At this stage only a few simple optimizations are done.

• Constant subexpressions are calculated only once. E.g.
  in COL*sin(180/pi()) the part sin(180/pi()) is evaluated once.
• If a subquery generates intervals of reals or dates, overlapping intervals are combined and eliminated. E.g.
    select from GER.MS where TIME in [select from ::POINTING where  
     sumsqr(DIRECTION[1])>0 giving [TIME-INTERVAL/2=:=TIME+INTERVAL/2]]

  can generate many identical or overlapping intervals. They are sorted and combined where possible to make the set as small as possible.

• If the righthand side of the IN operator is a single value, IN is turned into ==.
• If the righthand side of the IN operator is a set of integer values with a min-max range of $<$=1024*1024, that set is turned into a boolean vector to get linear lookup time.

TaQL does not recognize common subexpressions nor does it attempt to optimize the query. It means that the user can optimize a query by specifying the expression carefully. When using operator $\mid \mid$ or &&, attention should be paid to the contents of the left and right branches. Both operators evaluate the right branch only if needed, so if possible the left branch should be the shortest one, i.e., the fastest to evaluate.

The user should also use functions, operators, and subqueries in a careful way.

• SQUARE(COL) is (much) faster than COL**2 or POW(COL,2), because SQUARE is faster. It is also faster than COL*COL, because it accesses column COL only once.
  Similarly SQRT(COL) is faster than COL**0.5 or POW(COL,0.5)
• SQUARE(U) + SQUARE(V) $<$ 1000**2 is considerably faster than
  SQRT(SQUARE(U) + SQUARE(V)) $<$ 1000, because the SQRT function does not need to be evaluated for each row.
• TIME IN [$0<:<4$] is faster than TIME$>$0 && TIME$<$4, because in the first way the column is accessed only once.
• Returning a column from a subquery can be done directly or as a set. E.g.
    SELECT FROM maintable WHERE time IN  
        [SELECT time FROM othertable WHERE windspeed < 5]

  could also be expressed as

    SELECT FROM maintable WHERE time IN  
        [SELECT FROM othertable WHERE windspeed < 5 GIVING [time]]

  The latter (as a set) is slower. So, if possible, the column should be returned directly. This is also easier to write.
  An even more important optimization for this query is writing it as:

    SELECT FROM maintable WHERE time IN  
        [SELECT DISTINCT time FROM othertable WHERE windspeed < 5]

  Using the DISTINCT qualifier has the effect that duplicates are removed which often results in a much smaller set.

• Testing if a subquery contains at least N elements can be done in two ways:
    count([select column from table where expression]) >= N  
  and  
    exists (select from table where expression limit N)

  The second form is by far the best, because in that case the subquery will stop the matching process as soon as N matching rows are found. The first form will do the subquery for the entire table.
  Furthermore in the first form a column has to be selected, which is not needed in the second form.

• Sometimes operator IN and function ANY can be used to test if an element in an array matches a value. E.g.
    WHERE any(arraycolumn == value)  
  and  
    WHERE value IN arraycolumn

  give the same result. Operator IN is faster because it stops when finding a match. If using ANY all elements are compared first and thereafter ANY tests the resulting bool array.

• It was already shown in the section 4.8 that indexing arrays should be done with care.

7 Modifying a table

Usually TaQL will be used to get a subset from a table. However, as described in the first sections, it can also be used to change the contents of a table using the UPDATE, INSERT, or DELETE command. Note that a table has to be writable, otherwise those commands exit with an error message.

7.1 UPDATE

updates all or some rows in the first table. More input tables can be given in the FROM clause and used in clauses like SET and WHERE. Unlike SQL it is possible to specify more tables in the UPDATE part which is the same as specifying them in the FROM clause. However, using the FROM clause makes it more clear that only the first table is updated.
update_list is a comma-separated list of column=expression parts. Each part tells to update the given column using the expression. Both scalar and array columns are supported. E.g.

to make the antenna numbers zero-based if accidently they were written one-based.

are equivalent. They copy the DATA and FLAG column of that.ms to this.ms for rows where all data in this.ms are flagged. Note the use of the shorthand (alias) t2.

If an array gets an array value, the shape of the array can be changed (provided it is allowed for that table column). Arrays can also be updated with a scalar value causing all elements in the array to be set to that scalar value.

It sets all elements of the arrays in column FLAG to False.

Type promotion and demotion will be done where possible. For example, an integer column can get the value of a double expression (the result will be truncated).
Unit conversion will be done as needed. Thus if a column and its expression have different units, the expression result is automatically converted to the column’s unit. Of course, the units must be of the same type to be able to convert the data.

Note that if multiple column=expression parts are given, the columns are changed in the order as specified in the update-list. It means that if an updated column is used in an expression for a later column, the new value is used when evaluating the expression. e.g., in

the SUMD update uses the new DATA values.

Thus to swap the values of the ANTENNA1 and ANTENNA2 column, one can not do:

To solve this problem a temporary table (in this case in memory) can be used to save the value of e.g., ANTENNA1:

7.1.1 Partial Array Update

It is possible to update part of an array using array indexing and slicing. E.g.,

The first example sets only a single array element, while the second one sets an entire row in the array. Similar to numpy it is also possible to use a mask like

which sets the flag for the DATA values being a NaN. The data and mask must have the same shape. Note this is easier to write than the similar command

Masking and slicing can be combined making it possible to use masking on a part of an array. If the mask is given first, the slice is taken from both the data and mask. If the slice is given first, it is only applied to the data; the mask should have the same shape as the slice. For example:

Both commands set the flag for NaN data in the XX polarization. The first one is somewhat easier to write, but processes the entire DATA and FLAG before taking the slice. The second one only reads and processes the required parts of DATA and FLAG, thus is more efficient.

7.1.2 Update columns from a masked array

If a column is updated with the value of a masked array, only the array part of the masked array is used. However, it is also possible to jointly update the data column and mask column from a masked array by combining them in parentheses like:

It writes the data part into DATA and the mask into FLAG. As above it is possible to use a slice or mask operator on the combination like:

The slice or mask is applied to both columns.

7.2 INSERT

The INSERT command adds rows to the table. It can take three forms:

The first and second form are basically equivalent, but differ in syntax. The first form has the same syntax as the UPDATE command, while the second form is the SQL syntax making it possible to leave out the column names (see below). In both forms it is possible to jointly specify data column and mask column if the value is a masked array. This is done by combining them in parentheses like (DATA,FLAG) as described in the previous subsection for the UPDATE command.

The first form adds one row to the table and puts the values given in the expressions into the columns.
For example:

adds one row, puts 0 in ANTENNA1 and 1 in ANTENNA2.

The second form can add multiple rows to the table. It puts the values given in the expression lists into the columns given in the column list. If the column list is not given, it defaults to all stored columns in the table in the order as they appear in the table description. Multiple expression lists can be given; each list results in the addition of a row. Each expression in the expression list can be as complex as needed; for example, a subquery can also be given. Note that a subquery is evaluated before the new row is added, so the new row is not taken into account if the subquery is done on the table being modified.
It should be clear that the number of columns has to match the number of expressions.
Note that row cells not mentioned in the column list, are not written, thus may contain rubbish in the new rows.
The data types and units of expressions and columns have to conform in the same way as for the UPDATE command; values have to be convertible to the column data type and unit.
For example:

adds two rows, putting 0 and 2 in ANTENNA1 and 1 and 3 in ANTENNA2.

The LIMIT clause can be used to add multiple rows while giving fewer expressions. LIMIT can be given at the beginning or the end of the command. For example:

The first example will add 100 rows where the value in each row is the row number. The second example shows that multiple expression lists can be given. It will iterate through them while adding rows. Thus COL1 and COL2 will have the values 0, 1, 0, 1, and 0 in the new rows.

The third form evaluates the SELECT command and copies the rows found in the selection to the table being modified (which is given in the INTO part). The columns used in the modified table are defined in the column list. As above, they default to all stored columns. The columns used in the selection have to be defined in the column-list part of the SELECT command. They also default to all stored columns.
For example:

appends all rows and columns of my.ms to itself. Please note that only the original number of rows is copied.

copies rows from other.ms where ANTENNA1$>$0. It swaps the values of ANTENNA1 and ANTENNA2. All other columns are not written, thus may contain rubbish.

7.3 DELETE

deletes some or all rows from a table.

deletes the rows matching the WHERE expression.
If no selection is done, all rows will be deleted.
It is possible to specify more than one table in the FROM clause to be able to use, for example, keywords from other tables. Rows will be deleted from the first table mentioned in the FROM part.

8 Creating a new table

TaQL can be used to create a new table. The data managers to be used can be given in full detail. The syntax is:

The command consists of 4 parts, all of them optional.

• The table name and options can be given in the same way as in the GIVING clause.
• The columns are defined in the colspecs part. If not given, a table without any column is created. Below column specification is described in more detail.
• An expression giving the number of rows can be specified in the LIMIT part. If not given, it defaults to 0.
• For expert users data managers can be defined in the optional DMINFO part described further down.

The CREATE TABLE command can be used in a nested query making it possible to fill it immediately. For example:

creates a table with one column and ten rows. The column is filled with the row number. Note that the following command would do the same.

8.1 Column specification

The colspecs part defines the column names, their data types, and optional shapes and units. It can optionally be enclosed in square brackets or parentheses (for SQL compatibility). It is a comma separated list of column specifications. Each specification looks like:

The possible data type strings are given in section 4.1. The part enclosed in square brackets is optional. Zero or more of these keywords can be used. It makes it possible to define array columns and/or default data manager to be used. The square brackets are optional if only one such keyword is used.

• NDIM=n defines if the column contains scalars or arrays.
  A negative value means a scalar, which is the default (unless shape is also given). A value 0 means an array of any dimensionality. A positive value means an array with the given dimensionality.
• SHAPE=[d1,d2,...] makes it possible to define the exact array shape.
  If given and if NDIM is positive, they should be consistent.
• UNIT=’s’ defines the unit to be used for the column.
  It can be any valid unit (simple or compound). It is a string, thus must always be enclosed in quotes.
• COMMENT defines comments for the column.
  It has a string value, thus quotes have to be used.
• DMTYPE, DMGROUP are rather specific and are for the expert user.
  They have a string value, thus quotes have to be used.

8.2 Data manager specification

The datamanagers part makes it possible for the expert user to define the data managers to be used by columns. It is a comma separated list of data manager specifications looking like the output of the table.getdminfo command in Python. Each specification has to be enclosed in square brackets. For example:

The case of the keyword names used (e.g., NAME) is important. They have to be given in uppercase. The following keywords can be given:
NAME defines the unique name of the data manager.
TYPE defines the type of data manager.
SPEC is a list of keywords giving the characteristics of the data manager. This is highly data manager type specific. If shapes have to be given here, they always have to be in Casacore format, thus in Fortran order. TaQL has no knowledge about these internals.
COLUMNS is a list of column names defining all columns that have to be bound to the data manager.

9 Modifying the table structure

TaQL can be used to modify the table structure, i.e., to add, rename, and remove columns and keywords. It is also possible to add rows. The syntax is:

It changes the table with the given name. The tables given in the optional FROM clause can be used in expressions defining keyword values. Any number of subcommands can be given, separated by whitespace and/or comma. The following subcommands can be given. They are explained in the next subsections.

The nouns COLUMN and KEYWORD can also be given in the plural form. The whitespace between verb and noun is optional. For SQL-compatibility DROP can be used instead of DELETE.
For example:

renames column Col1 to Col1A and adds a new column Col1 with data type I4.

Note that TaQL has no way of showing keywords that have a record value. The program showtable can be used for that purpose.

9.1 ADD COLUMN

adds one or more columns to the table. The specification of the columns and the optional data managers is the same as used in the CREATE TABLE command. Thus for each column a data type, dimensionality or shape, and unit can be given. The data manager(s) for the new columns can be specified in the DMINFO part. If not given, StandardStMan will be used. For example:

adds two columns, a 4-byte floating point scalar column and an 8-byte floating point 3-dim array column. They will be stored with StandardStMan.

9.2 RENAME COLUMN

renames one or more columns in a table. For example:

9.3 DELETE COLUMN

removes one or more columns. Note that if multiple columns are combined in a TiledStMan, they have to be removed at the same time. Thus in that case

are not the same, because the second example might fail.

9.4 SET KEYWORD

adds a keyword with the given value or replaces the value if the keyword already exists. The value of a keyword can be a scalar, array, or arbitrarily deeply nested record. See section 4.5 how to specify a keyword name in a column or nested record. The AS dtype part can be used to explicitly set the data type of a new keyword. For an existing keyword, the data type of the new value has to match the data type of the current value.

The value can be an expression, possibly using values from another table given in the FROM clause. It has to be a constant expression, thus cannot depend on column values. Of course, column values can be used when aggregated to a single value. If no data type is given, the data type of the expression result is used. If given, upward and downward coercion is possible (e.g., integer to float and also float to integer). For example:

The 1st example sets table keyword key1 to 4. Its data type is not given, thus is the expression’s data type, in this case I8.
The 2nd example sets key1 to 9, but as an unsigned 4 byte integer. Note that the :: part is redundant.
The 3rd example copies the value of keyword otherkey while converting its data type to I4. Note that if no data type is given, the data type of otherkey is NOT preserved, because it is seen as a TaQL expression which has data type I8 (or R8).
The 4th example sets the ckey.subrec.fld1 in column col to the given vector. It is a nested structure, thus field fld1 in field subrec of column keyword ckey will be set. Its data type will be R8.
Note that the command in the 4th example does not create the higher level records. If not existing yet, the 5th example can be used to create them, where [=] denotes an empty record (it is the old Glish syntax for an empty struct).
The last example shows how to create a key with an empty integer vector as value. In such a case the data type must be given, because it cannot be derived from the value.

Setting a keyword to the value of another keyword is easily possible. For instance:

However, it has two problems.
1) As explained above the data type might not be preserved.
2) Keywords having a record value cannot be copied this way, because TaQL expressions do not support record values.

9.5 COPY KEYWORD

copies the value of keyword otherkey to key. It can be used for any keyword value, thus also for records. The optional AS dtype part can be used to change the data type.

9.6 RENAME KEYWORD

renames one or more table or column keywords. If the old keyword is a field in a column or a nested record, the new name should only contain the new field name, not the full keyword path. For example:

The first example renames the table keyword NAME and the keyword CNAME of column Col1.
The second example renames a field in the nested records of table keyword KEYS.

9.7 DELETE KEYWORD

removes one or more table or column keywords.

9.8 ADD ROW

adds the given number of rows to the table.

where nrows can be any expression. For example,

makes mytab the same size as othertab (assuming it was empty).

10 Counting in a table

Before TaQL had the GROUPBY command, the COUNT command could be used instead of the gcount aggregate function to count the number of occurrences in a table.
For backward compatibility this command can still be used, but its usage is discouraged, also because usually GROUPBY is faster.

The exact syntax is:

It counts the number of rows for each unique tuple in the column list of the table (after the possible WHERE selection is done). For example:

counts the number of rows per timestamp.

counts the number of rows per baseline.

As in the other TaQL commands a column in the column list can be any expression, but that will be slower than straight columns.

11 Calculations on a table

TaQL can be used to get derived values from a table by means of an expression. The expression can result in any data type and value type. For example, if the expression uses an array column, the result might be a vector of arrays (an array for each row). If the expression uses a scalar column, the result might be a vector of scalars or even a single scalar if a reduce function like SUM is used.

The CALC command was developed before the GROUPBY was available and before SELECT could be used without the FROM part. Currently, SELECT is more powerful than the CALC command. For example, multiple expressions can be given in a SELECT command. However, especially in Python sessions CALC has the advantage that it returns the results as a numpy-array or a list instead of a Casacore table.

The exact syntax is:

The part in square brackets can be omitted if no column is (directly) used in the expression. The examples will make clear what that means.
The following syntax is still available for backward compatibility:

is a simple expression not using a table. It shows how the CALC command can be used as a desk calculator to convert 1 inch to cm.

gives a vector of scalars containing the mean per row.

gives a single scalar giving the sum of the means in each row. Note that in this command the CALC command does not need the FROM clause, because it does not use a column itself. Columns are only used in the nested query which has a FROM clause itself.

12 Examples

12.1 Selection examples

Some examples are given starting with simple ones.

12.1.1 Reference table results

The result of the following queries is a reference table, because no expressions have been given in the column-list. This will be the most common case when using TaQL.

SELECT FROM some.ms WHERE ANTENNA1 != ANTENNA2
selects the cross-correlations in a MeasurementSet.

SELECT NAME FROM some.ms::ANTENNA
selects the NAME of all antennae in a MeasurementSet.

SELECT unique ANTENNA1,ANTENNA2 FROM some.ms
gives the baselines used in a MeasurementSet.

SELECT ANTENNA1,ANTENNA2 FROM some.ms GROUPBY ANTENNA1,ANTENNA2
does the same using the GROUPBY clause.

SELECT FROM mytable ORDERBY column0 DESC limit 10
selects the 10 highest values of column0.

SELECT FROM some.MS WHERE near(MJD(1999/03/30/17:27:15), TIME)
selects the rows with the given time from a MeasurementSet.
Note that the TIME is stored in seconds, but will automatically be converted to days.

SELECT FROM some.MS where TIME in
     [{MJD(1999/03/30/17:27:15),MJD(1999/03/30/17:29:15)}]
selects the rows in the given closed time interval.

SELECT FROM some.MS where TIME in
     [MJD(1999/03/30/17:27:15),MJD(1999/03/30/17:29:15)]
selects the rows having one of the given times.
Note the difference with the previous example where an interval was given. Here a set of two individual time values is given.

SELECT NAME FROM some.ms::ANTENNA WHERE NAME ! p/[CR]S*/
selects the names of the international LOFAR stations (not core or remote).

SELECT FROM some.ms WHERE ntrue(FLAG) >= 3
selects rows where at least 3 visibilities are flagged.

SELECT FROM book.table WHERE nelements(author) > 1
selects books with more than 1 author.

SELECT FROM some.ms WHERE any(ANTENNA1==[0,0,1] && ANTENNA2==[1,3,2])
selects the antenna pairs (baselines) 0-1, 0-3, and 1-2.
It requires some explanation. The two comparisons result in boolean vectors (with 3 elements). It matching elements are both true, the baseline in the table row matches. Thus the vectors are and-ed to see if any two matching elements are true.

SELECT FROM some.ms WHERE ANTENNA1 in [0,0,1] && ANTENNA2 in [1,3,2])
looks the same as above, but will select all baselines between the two sets, thus also 1-1, 1-3, and 0-2.

SELECT FROM some.ms t1, that.ms t2 WHERE !all(near(t1.DATA, t2.DATA, 1e-5))
selects all rows where the DATA columns in both tables are not equal (with some tolerance). Note the use of shorthands t1 and t2.

SELECT FROM mytable WHERE cos(0d1m) <=
     sin(52deg) * sin(DEC) + cos(52deg) * cos(DEC) * cos(3h30m - RA)
selects observations with an direction (in say J2000) inside a cone with a radius of 1 arcmin around (3h30m, 52deg). To find them the condition DISTANCE$<=$RADIUS must be fulfilled, which is equivalent to COS(RADIUS)$<=$COS(DISTANCE).

SELECT FROM mytable WHERE [RA,DEC] INCONE [3h30m, 52deg, 0d1m]
does the same as above in an easier (and faster) way.

SELECT FROM mytable WHERE angdist([RA,DEC], [3h30m, 52deg]) <= 0d1m]
is another way to do the above.

SELECT FROM mytable WHERE object == pattern("3C*") &&
     [RA,DEC] INCONE [3h30m, 52deg, 0d1m]
finds all 3C objects inside that cone.

SELECT ANTENNA1,ANTENNA2,sqrt(sumsqr(UVW[:2]))
     FROM some.ms GROUPBY ANTENNA1,ANTENNA2
finds the 10 longest baselines. It groups by ANTENNA1 and ANTENNA2 to get the unique baselines. UVW[:2] denotes the U and V coordinate giving the baseline length.

select from MY.MS where DATA_DESC_ID in [select from ::DATA_DESCRIPTION where
     SPECTRAL_WINDOW_ID in [0,2,4] giving [ROWID()]]
finds all rows in a measurement set matching the given spectral windows. It uses a nested query to find the DATA_DESC_ID for each spectral window.

select from MY.MS where TIME in [select from ::SOURCE where REST_FREQUENCY < 180MHz
     giving [TIME-INTERVAL/2 =:= TIME+INTERVAL/2]]
finds all rows in a measurement set observing sources with a rest frequency less than 180 Mhz.

select from VLA.MS,
     [select from VLA.MS where sumsqr(UVW[:2]) < 625] as TIMESEL
     where TIME in [select distinct TIME from TIMESEL]
     && any([ANTENNA1,ANTENNA2] in [select from TIMESEL giving
     [iif(UVW[2] < 0, ANTENNA1, ANTENNA2)]])
selects rows where an antenna (VLA has 25 m diameter) is shadowed.
The query in the FROM command finds all rows where an antenna is shadowed (i.e., its UV-distance less than 25 meters) and creates a temporary table. This selection (named TIMESEL) is done first otherwise two 2 equal selections are needed in the main WHERE command. The last line determines which antenna is shadowed (based on the W coordinate). The two lines above selects the times and baselines where an antenna is shadowed.

select from MS
     where DATA_DESC_ID in [select from ::DATA_DESCRIPTION
     where SPECTRAL_WINDOW_ID in [select from ::SPECTRAL_WINDOW
     where NET_SIDEBAND==1 giving [ROWID()]] giving [ROWID()]]
finds all rows in the MeasurementSet with the given NET_SIDEBAND.
The MeasurementSet uses a table to map spectral-window-id to data-desc-id. Hence two nested subqueries are needed.

select findcone(REFERENCE_DIR[0,],
     [16h34m33.805,62d45m36.83, 12h29m06.7,2d3m9], 1arcsec)
     from MS/FIELD
compares the direction given in the first argument with the directions given in the second function argument using the search radius given in the third argument. It returns the index of the first matching cone (thus 0 or 1). If no cone matches, it returns -1.
It can be used in the following example to find the name of the source matching a direction.

select [’unknown’,’3C343’,’3C273’][1+findcone(...)] from MS/FIELD where ... is the findcone argument list given in the previous example. 1 is added to cope with the case that no cone matches.

12.1.2 Plain table results

The following examples result in a plain table, thus in a deep copy of the query results, because the column-list contains an expression or a data type.

SELECT column0+column1 FROM mytable
creates a table of 1 column with name Col_1. Its data type is on the expression data type.

SELECT column0+column1 Res I4 FROM mytable
creates a table of 1 column with name Res. Its data type is a 4 byte signed integer.

SELECT colx colx R4 FROM mytable
creates a table of 1 column with name colx. The sole purpose of this selection is to convert the data type of the column.

SELECT means(DATA,0) AS DATA_MEAN C4 FROM my.ms
creates a table of 1 column with name DATA_MEAN. Column DATA in a Casacore MeasurementSet is a 2-dimensional array with axes polarization and frequency. This command calculates and stores the mean in each polarization. If no data type was given, the means would have been stored as double precision complex (which is the expression data type).
Note that this command is valid when using python style; in glish style MEANS(DATA,2) should be used.

12.2 Modification examples

update MY.MS set VIDEO_POINT=MEANS(DATA,2) where isdefined(DATA)
sets the VIDEO_POINT of each correlation to the mean of the DATA for that correlation. Note that the 2 indicates averaging over the second axis, thus the frequency axis.

update MY.MS set FLAG_ROW=T where isdefined(FLAG) && all(FLAG)
sets FLAG_ROW in the rows where the entire FLAG array is set.

delete from MY.MS where FLAG_ROW
deletes all flagged rows.

insert into MY.MS select from OTHER.MS where !FLAG_ROW
copies all unflagged rows from OTHER.MS to MY.MS.

insert into MY.MS/DATA_DESCRIPTION
     (SPECTRAL_WINDOW_ID,POLARIZATION_ID,FLAG_ROW)
     values (1,0,F)
adds a row to the DATA_DESCRIPTION subtable and initializes it.

12.2.1 Applying running median to an image

The following command shows how a running median can be applied to an Casacore image.

The running medians are subtracted from the data in the copy. It uses a half window size of 25x25, thus the full window is 51x51.
When doing this, one should take care that in case of a spectral line cube the image is not too large, otherwise it won’t fit in memory. If too large, it should be done in chunks like:

where sc and ec are the start and end frequency channel. In this example it is assumed that the axes of the image are RA, DEC, freq, Stokes.
Note that the image is updated, so it should have been copied before if the original data needs to be kept.

12.3 Table creation examples

create table mytab (col1 I4, col2 I4, col3 R8)
creates table mytab of 3 scalar columns.

create table mytab
creates an empty table.

create table mytab colarr R4 ndim=0
creates a table of 1 array column with arbitrary dimensionality.

create table mytab colarr R4 [shape=[4,128], dmtype=’TiledColumnStMan’]
creates a table of 1 array column with the given shape. The column is stored with the TiledColumnStMan storage manager using its default settings.

create table mytab colarr R4 shape=[4,128]
     dminfo [TYPE=’TiledColumnStMan’, NAME=’TCSM’,
     SPEC=[DEFAULTTILESHAPE=[4,32,64]], COLUMNS=[’colarr’]]
creates a table of 1 array column with the given shape. The column is stored with the TiledColumnStMan storage manager using the given settings.

12.4 Calculation examples

calc 1+2
uses TaQL as a desktop calculator.

calc 7-Apr-2007 - 20-Nov-1979
calculates the number of days between these dates.

calc sum([select from MY.MS giving [ntrue(FLAG)]])
determines the total number of flags set in the measurement set.

calc mean(abs(DATA))
from [select from MY.MS where ANTENNA1==0]

calculates for each row the mean of the data for the selected subset of the measurement set.

calc mean([select from MY.MS where ANTENNA1==0
     giving [mean(abs(DATA))]])

looks like the previous example. It, however, calculates the mean of the mean of the data in each row for the selected subset of the measurement set.

calc max([select from MY.MS where isdefined(DATA)
     giving [max(abs(VIDEO_POINT-MEANS(DATA,0)))]])

shows the maximum absolute difference between VIDEO_POINT of each correlation and the mean of the DATA for that correlation. Note that the 2 indicates averaging over the first axis, thus the frequency axis.

12.5 Aggregation/groupby examples

select gcount(*) from my.ms
counts the number of rows in the table.

select TIME, gcount(*) from my.ms groupby TIME
counts the number of rows (usually number of baselines) per time slot.

select ANTENNA1,ANTENNA2,gfirst(TIME),glast(TIME),gcount()
     from my.ms groupby ANTENNA1,ANTENNA2
counts the number of rows (usually number of time slots) and shows the first and last time per baseline.

select gmean(DATA) from my.ms
     groupby int((TIME - [select TIME from my.ms limit1][0]) / INTERVAL / 10))
calculates the average of DATA for every 10 time slots. Note it also averages in frequency and polarization. The following example shows how to average each frequency channel and polarization.

select boxedmean(gaggr(DATA), 10, 4) from my.ms
     groupby int((TIME - [select TIME from my.ms limit1][0]) / INTERVAL / 10))
calculates the average of DATA per polarization for every 10 time slots and 4 frequency channels. Note it it first combines the data of each 10 time slots in a single array, after which the boxedmean function is used to average every [10,4,1] box.

12.5.1 Obtaining the flux density from visibility data

The Miriad program uvflux estimates the source I flux density and its standard deviation at the phase center without having to make an image. A single, not too complicated TaQL command (courtesy Dijkema, Heald) provides the same functionality on a MeasurementSet. For LOFAR it is best to use baselines with a length between 5 and 10 km. The command shows various aspects of TaQL that are explained below. The numbers at the beginning of the lines point to the text following the example.

A subquery is used to get the average flux (I = 0.5*(XX+YY)) per time slot.

1. For each baseline it gets the mean of the channels. Note that it uses sum/n to ignore flagged visibilities. The iif function tells to use 0 for them. Also note that XX is the 1st and YY the 4th polarisation, hence [0:4:3] (or [::3]) indexes these polarisations. Once masked arrays are supported by TaQL, it could be written as: mean(DATA[,::3][FLAG[,::3]])
2. It only uses the baselines with lengths between 5 and 10 km where not all visibilities are flagged. Note that a mscal user defined function is used for the baseline selection as described in section 4.10.17.
3. Thereafter the average flux per time slot is determined in the subquery using the gmean aggregation and GROUPBY functionality. The result is an intermediate table with one column called SUMMED and a row per time slot.
4. Finally, the outer query uses aggregate functions to calculate the overall mean, standard deviation, and number of time slots from the result of the subquery. The final result is a table with 1 row and 3 columns.

12.5.2 Number of fully flagged baselines per antenna

The example below counts per antenna the number of fully flagged baselines, excluding the autocorrelations. It uses grouping and aggregate functions twice; first per baseline, thereafter per antenna. It also uses the concatenation and backreferencing features. It also shows the results when using the time keyword, which shows that the processing time is dominated by the first query.

1. The first inner select counts the number of fully flagged baselines per baseline and stores the result with the antennas making up a baseline in a temporary table with shorthand t0.. In this way the expensive counting part needs to be executed only once.
2. This table has to be summed for both antennas of the baselines, which is done in the second part creating a concatenated table with shorthand t1. It backreferences the first table t0 twice. First to use it directly to count for ANTENNA1, thereafter to count for ANTENNA2 by doing a select to make the ANTENNA2 the ANTENNA column.
3. The table does not contain the antennas having no fully flagged baselines. Therefore the third part of the concatenation copies the ANTENNA table to insert zero counts for all antennas.
4. Finally the outer select (in the first line) sums the values in the concatenated table per antenna. It also retrieves the name of the antenna by indexing in the selection of all antenna names.

13 Interface to TaQL

User and a programmer interfaces to TaQL are available. The program taql and some Python and Glish functions form the user interface, while C++ classes and functions form the programmer interface.

13.1 Python interface python-casacore

The main TaQL interface in Python is formed by the query function in module table. The function can be used to compose and execute a TaQL command using the various (optional) arguments given to the query function. E.g.

The first command opens the table mytable. The second command does a simple query resulting in a temporary table. That temporary table is used in the next command resulting in a persistent table. The latter function call is transformed to the TaQL command:

During execution $1 is replaced by table seltab1.
Note that the name argument generates the GIVING part to make the result persistent.

The functions sort and select exist as convenience functions for a query consisting of a sort or column selection only. Both functions have an optional second name parameter to make the result persistent.

The calc function can be used to execute a TaQL calc command on the current table. The result can be kept in a variable. For example, the following returns a vector containing the median of the DATA column in each table row:

It is possible to embed Python variables and expressions in a TaQL command using the syntax $variable and $(expression). A variable can be a standard numeric or string scalar or vector. It can also be a table tool. An expression has to result in a numeric or string scalar or vector. E.g

These three queries give the same result.
The substitution mechanism is described in more detail in pyrap.util.

The most generic function that can be used is taql (or its synonym tablecommand). The full TaQL command has to be given to that command. The result is a table object. E.g.

By default, these commands will use the Python style for a TaQL statement. The style argument can be used to choose another style.

13.2 Interface to Glish

The Glish interface is formed by script table.g. By default, it will use the Glish style for a TaQL statement. For example:

13.3 Program taql

The program taql makes it possible to execute TaQL commands from the shell. Commands can be given in different ways:

• The TaQL command can be given directly as command line arguments to the taql program. The arguments will be combined to a single command (separated by a space). Note that using multiple arguments instead of a single (quoted) argument makes it easier to use tab-completion for the table name. It will execute the command, show the result, and exit.
• Using the -f option, the name of a file containing one or more TaQL commands can be given. The commands can be split over multiple lines, where a # can be used for comments. A semicolon has to be used to separate commands. The special commands listed below for the interactive case, can also be given.
• The program is run interactively if command nor -f is given. It will run until the user stops via the command exit, quit, or q or by giving Ctrl/D. Command editing and recall is possible, unless taql was built without readline support. Of course, a file can be used as input by redirection to stdin. This is more or less the same as using -f, but a command cannot be split over multiple lines and no semicolon is needed to separate commands. Interactive commands are kept in $HOME/.taql_history making command recall possible across multiple taql sessions.
  A few types of commands can be given:
  • help, h or ? shows brief help information.
  • show can be used to show to available units or measure types.
  • A full TaQL command. If no GIVING part is given, the resulting table is not made persistent. It shows the number of matching rows.
  • A full TaQL command preceded by varname=, where varname is the name under which the resulting table is kept in this session. Thus the result is not a persistent table (unless GIVING was given), but it is kept temporarily. The varname can be used in subsequent commands like SELECT FROM $varname.
  • varname= without a further command removes the temporary result.
  • varname shows the number of rows in the temporary result. It can be followed by one or more question marks to show the column names and details about them.
    Note that if an unknown varname is given, it is treated as a TaQL command resulting in a parse error.
  • Comments can be given after the hash (#) character. Empty lines are ignored.

If columns are selected in the TaQL command, their values are printed. By default they are separated by tabs, but the -d option can be used to define another delimiter. Epoch, position, and direction measures are printed in a formatted way. Note that TaQL’s str function offers nice formatting for any value. Also the result of CALC commands is printed. Otherwise only the number of selected (or updated or deleted) rows is printed. Options like -nops can be given to suppress printing the results.

By default the program taql uses the Python style. The -s option can be used to specify a style.

13.4 C++ interface

The C++ programmer can use TaQL commands and expressions at various levels,

13.4.1 TaQL query string

The function tableCommand in TableParse.h can be used to execute a TaQL command. The result is a Table object. E.g.,

These examples do the same as the Python ones shown above.
Note that in the second function call the table name $1 is replaced by the object seltab1 passed to the function.
There is no style argument, so if an explicit style is needed it should be the first part of the TaQL statement. Note that the Glish style is the default style.

13.4.2 Expression string

The function parse in RecordGram.h can be used to parse a TaQL expression. The result is a TableExprNode object that can be evaluated for each row in the table. E.g.

The example above does the same as the first example in the previous section. There are, however, better ways to use this functionality.

The example above shows a boolean scalar expression, but it can also be a numeric expression or an array expression as shown in the example below. Note that TaQL expression results have data type Bool, Int64, Double, DComplex, String, or MVTime.

Class RecordGram can also be used to apply TaQL to C++ vectors of values or Records. The RecordGram class documentation and its test program describe these features in more detail.

13.4.3 Expression classes

The other expression interface is a true C++ interface having the advantage that C++ variables can be used directly. Class Table contains functions to sort a table or to select columns or rows. When selecting rows class TableExprNode (in ExprNode.h) has to be used to build a WHERE expression which can be executed by the overloaded function operator in class Table. E.g.

does the same as the first example shown above. See classes Table, TableExprNode, and TableExprNodeSet for more information on how to construct a WHERE expression.

14 Writing user defined functions

A C++ user defined function has to be written as a class derived from the abstract base class UDFBase. The documentation of this base class describes how to write a UDF. Furthermore one can look at class UDFMSCal that contains the UDFs described in subsection User defined functions.

It is possible to write a UDF that operates on an individual expression (for each table row) and returns the result. It is, however, also possible to write a UDF acting as an aggregate function. In that case it will return a result based on the values of all rows in a group. See the desription of the GROUPBY clause for more information on the GROUPBY clause and aggregate functions.

Note that a class can contain multiple UDFs as done in UDFMSCal. Also note that a single UDF can operate on multiple data types which is similar to a function like min that can operate on scalars and arrays of different data types.

A UDF class can contain a HELP function, which should return help information. This function is called by a help command like

It returns an overview of the functions in the UDF class and possible other information. The optional subtype argument can be used to return more specific information. Note that the same result is given by

TaQL finds a UDF by looking in a dictionary mapping the UDF name to a function constructing an object of the UDF class. If not found, it tries to load the shared library with the lowercase name of the library part of the UDF (like in derivedmscal.pa1). If the load is successful, it calls an initialization function in the shared library that should add all UDF functions in the library to the dictionary. The description of the UDFBase class shows how this should be done.

14.1 UDFs in Python

NOTE: This section is for a future version of TaQL. It has not been fully implemented yet.

For performance reasons User Defined Functions will usually be implemented in C++. It is, however, possible to implement them in Python, both regular functions and aggregate functions. This can be done by means of the pytaql module of Casacore.

A UDF has to be implemented in Python by subclassing pytaqlbase, that can be imported from Casacore.python. The subclass has to implement a few functions, some are optional. The functions are called in the order given below.

• __init__(self)
  The class constructor must call the __init__ function of the superclass.
• needTable(self)
  This optional function tells if the UDF needs the Table object of the table being queried. If True is returned, the function setTable is called thereafter. The Table object can be used by UDFs needing extra info (e.g., keywords) from the table being queried or from its subtables (comparable to derivedmscal). It requires the import of pyrap.tables at the beginning of the UDF script.
• setTable(self, tab)
  This function makes it possible to keep the Table object. It must be implemented if needTable returns True. The object should be kept like:
      self.tab = pyrap.tables.table (tab, _oper=3)

• setup(self, valuetypes, datatypes, units)
  This function is called once at the beginning. It gets the value types, data types, and units of the function arguments. The length of the sequences tells the number of arguments.
   valuetypes  int seq   value type of each argument  
                         0=scalar 1=array 2=set  
   datatypes   int seq   data type of each argument  
                         0=bool 1=int 2=float 3=complex  
                         4=string 6=date  
   units       strings   unit of each argument (empty=no unit)

  The UDF should check if the argument types are correct and determine the result type. It has to return a dict containing the following fields:

   ndim      int       dimensionality of result  
                       -1=scalar 0=unknown  
   shape     int seq   shape (if fixed, otherwise empty sequence)  
   dtype     int       data type of result  
   unit      string    optional unit of result  
   isaggr    bool      True = UDF is aggregate function

• getValue(self, argValues, rownr)
  This function must return the function value for the given argument values. It is only called if setup does not set isaggr=True. Normally the rownr argument is not needed, but it could be used to obtain special info from the Table object for that row.
• getAggrValue(self, rownrs)
  This function must return the value of the aggregate function for the given rows. The argument values are not passed, because their sizes may exhaust memory. Instead the list of row numbers in the aggregation group are given. For each row the following function must be called to get a list of the argument values for that row.
      getArgValues(self, rownr)

Such a UDF can be called in TaQL like py.module.class where the class defaults to the module name.
An example of UDFs in Python is given below. The first one is a regular UDF, the second one an aggregate UDF.

15 Possible future developments

In the near or far future TaQL can be enhanced by adding new features and by doing optimizations.

• Add ROLLUP/CUBE to the GROUPBY clause.
• Implement the OVER/PARTITION clause.
• Add explicit JOIN clause (probably only equi-joins).
• Add UNION, INTERSECTION, and DIFFERENCE.
• Handle invalid subexpressions (e.g., exceeding array bounds) as null arrays which can be tested with the function ISNULL.