CREATE FUNCTION Statement
Creates a user-defined function (UDF), which you can use to implement custom logic during SELECT or INSERT operations.
The syntax is different depending on whether you create a scalar UDF, which is called once for each row and implemented by a single function, or a user-defined aggregate function (UDA), which is implemented by multiple functions that compute intermediate results across sets of rows.
To create a scalar UDF, issue a CREATE FUNCTION statement:
CREATE FUNCTION [IF NOT EXISTS] [db_name.]function_name([arg_type[, arg_type...]) RETURNS return_type LOCATION 'hdfs_path' SYMBOL='symbol_or_class'
To create a UDA, issue a CREATE AGGREGATE FUNCTION statement:
CREATE [AGGREGATE] FUNCTION [IF NOT EXISTS] [db_name.]function_name([arg_type[, arg_type...]) RETURNS return_type LOCATION 'hdfs_path' [INIT_FN='function] UPDATE_FN='function MERGE_FN='function [PREPARE_FN='function] [CLOSEFN='function] [SERIALIZE_FN='function] [FINALIZE_FN='function] [INTERMEDIATE type_spec]
Statement type: DDL
If the underlying implementation of your function accepts a variable number of arguments:
- The variable arguments must go last in the argument list.
- The variable arguments must all be of the same type.
- You must include at least one instance of the variable arguments in every function call invoked from SQL.
- You designate the variable portion of the argument list in the CREATE FUNCTION statement by including ... immediately
after the type name of the first variable argument. For example, to create a function that accepts an INT argument, followed by a BOOLEAN, followed by one or more STRING arguments, your CREATE FUNCTION statement would look like:
CREATE FUNCTION func_name (INT, BOOLEAN, STRING ...) RETURNS type LOCATION 'path' SYMBOL='entry_point';
See Variable-Length Argument Lists for how to code a C++ UDF to accept variable-length argument lists.
Scalar and aggregate functions:
The simplest kind of user-defined function returns a single scalar value each time it is called, typically once for each row in the result set. This general kind of function is what is usually meant by UDF. User-defined aggregate functions (UDAs) are a specialized kind of UDF that produce a single value based on the contents of multiple rows. You usually use UDAs in combination with a GROUP BY clause to condense a large result set into a smaller one, or even a single row summarizing column values across an entire table.
You create UDAs by using the CREATE AGGREGATE FUNCTION syntax. The clauses INIT_FN, UPDATE_FN, MERGE_FN, SERIALIZE_FN, FINALIZE_FN, and INTERMEDIATE only apply when you create a UDA rather than a scalar UDF.
The *_FN clauses specify functions to call at different phases of function processing.
- Initialize: The function you specify with the INIT_FN clause does any initial setup, such as initializing member variables in internal data structures. This function is often a stub for simple UDAs. You can omit this clause and a default (no-op) function will be used.
- Update: The function you specify with the UPDATE_FN clause is called once for each row in the original result set, that is, before any GROUP BY clause is applied. A separate instance of the function is called for each different value returned by the GROUP BY clause. The final argument passed to this function is a pointer, to which you write an updated value based on its original value and the value of the first argument.
- Merge: The function you specify with the MERGE_FN clause is called an arbitrary number of times, to combine intermediate values produced by different nodes or different threads as Impala reads and processes data files in parallel. The final argument passed to this function is a pointer, to which you write an updated value based on its original value and the value of the first argument.
- Serialize: The function you specify with the SERIALIZE_FN clause frees memory allocated to intermediate results. It is required if any memory was allocated by the Allocate function in the Init, Update, or Merge functions, or if the intermediate type contains any pointers. See the UDA code samples for details.
- Finalize: The function you specify with the FINALIZE_FN clause does any required teardown for resources acquired by your UDF, such as freeing memory, closing file handles if you explicitly opened any files, and so on. This function is often a stub for simple UDAs. You can omit this clause and a default (no-op) function will be used. It is required in UDAs where the final return type is different than the intermediate type. or if any memory was allocated by the Allocate function in the Init, Update, or Merge functions. See the UDA code samples for details.
If you use a consistent naming convention for each of the underlying functions, Impala can automatically determine the names based on the first such clause, so the others are optional.
For end-to-end examples of UDAs, see Impala User-Defined Functions (UDFs).
Complex type considerations:
Currently, Impala UDFs cannot accept arguments or return values of the Impala complex types (STRUCT, ARRAY, or MAP).
- You can write Impala UDFs in either C++ or Java. C++ UDFs are new to Impala, and are the recommended format for high performance utilizing native code. Java-based UDFs are compatible between Impala and Hive, and are most suited to reusing existing Hive UDFs. (Impala can run Java-based Hive UDFs but not Hive UDAs.)
- The body of the UDF is represented by a .so or .jar file, which you store in HDFS and the CREATE FUNCTION statement distributes to each Impala node.
- Impala calls the underlying code during SQL statement evaluation, as many times as needed to process all the rows from the result set. All UDFs are assumed to be deterministic, that is, to always return the same result when passed the same argument values. Impala might or might not skip some invocations of a UDF if the result value is already known from a previous call. Therefore, do not rely on the UDF being called a specific number of times, and do not return different result values based on some external factor such as the current time, a random number function, or an external data source that could be updated while an Impala query is in progress.
- The names of the function arguments in the UDF are not significant, only their number, positions, and data types.
- You can overload the same function name by creating multiple versions of the function, each with a different argument signature. For security reasons, you cannot make a UDF with the same name as any built-in function.
- In the UDF code, you represent the function return result as a struct. This struct contains 2 fields. The first field is a boolean representing whether the value is NULL or not. (When this field is true, the return value is interpreted as NULL.) The second field is the same type as the specified function return type, and holds the return value when the function returns something other than NULL.
- In the UDF code, you represent the function arguments as an initial pointer to a UDF context structure, followed by references to zero or more structs, corresponding to each of the arguments. Each struct has the same 2 fields as with the return value, a boolean field representing whether the argument is NULL, and a field of the appropriate type holding any non-NULL argument value.
- For sample code and build instructions for UDFs, see the sample UDFs in the Impala github repo.
- Because the file representing the body of the UDF is stored in HDFS, it is automatically available to all the Impala nodes. You do not need to manually copy any UDF-related files between servers.
- Because Impala currently does not have any ALTER FUNCTION statement, if you need to rename a function, move it to a different database, or change its signature or other properties, issue a DROP FUNCTION statement for the original function followed by a CREATE FUNCTION with the desired properties.
- Because each UDF is associated with a particular database, either issue a USE statement before doing any CREATE FUNCTION statements, or specify the name of the function as db_name.function_name.
If you connect to different Impala nodes within an impala-shell session for load-balancing purposes, you can enable the SYNC_DDL query option to make each DDL statement wait before returning, until the new or changed metadata has been received by all the Impala nodes. See SYNC_DDL Query Option for details.
Impala can run UDFs that were created through Hive, as long as they refer to Impala-compatible data types (not composite or nested column types). Hive can run Java-based UDFs that were created through Impala, but not Impala UDFs written in C++.
The Hive current_user() function cannot be called from a Java UDF through Impala.
Currently, Impala UDFs and UDAs are not persisted in the metastore database. Information about these functions is held in the memory of the catalogd daemon. You must reload them by running the CREATE FUNCTION statements again each time you restart the catalogd daemon.
Cancellation: Cannot be cancelled.
HDFS permissions: This statement does not touch any HDFS files or directories, therefore no HDFS permissions are required.