Writing UDFs in Python

Dependencies

The UDF Python API consists of one file, kinetica_proc.py. This needs to be included in the UDF source code. There are no external dependencies beyond the Python standard library.

To take advantage of GPU processing within a UDF, the CUDA Toolkit must be downloaded & installed from the NVIDIA Developer Zone.

Also, any Python packages the UDF may need to use should be installed on all cluster nodes. The gpudb-pip.sh script is provided to install any given package into the bundled Python installation on all nodes. To run:

$ /opt/gpudb/udf/api/python/gpudb-pip.sh install <package name>

Defining the Function

A UDF must first get a handle to the ProcData class imported from kinetica_proc.py. This will parse the primary control file and set up all the necessary structures. It will return ProcData object, which is used to access everything else. All configuration information is cached, so repeated calls to get a ProcData will not reload any configuration files.

Once the UDF has been initialized, the following calls are possible:

Input

Call	Description
proc_data.request_info	Returns a map of basic information about the /execute/proc request, map values being accessed using: proc_data.request_info[<map_key>] The full set of map keys is listed below, under Request Info Keys.
proc_data.params	Returns a map of string-valued parameters that were passed into the UDF
proc_data.bin_params	Returns a map of binary-valued parameters that were passed into the UDF
proc_data.input_data	Returns an InputDataSet object for accessing input table data that was passed into the UDF

Output

Call	Description
proc_data.results	Returns a map that can be populated with string-valued results to be returned from the UDF
proc_data.bin_results	Returns a map that can be populated with binary-valued results to be returned from the UDF
proc_data.output_data	Returns an OutputDataSet object for writing output table data that will be written to the database

The UDF must finish with a call to proc_data.complete. This writes out some final control information to indicate that the UDF completed successfully.

NOTE: If this call is not made, the database will assume that the UDF didn’t finish and will return an error to the caller.

The InputDataSet & OutputDataSet objects contain InputTable & OutputTable objects, which, in turn, contain InputColumn & OutputColumn, holding the actual data sets. Tables and columns can be accessed by index or by name. For example, given a customer table at InputDataSet index 5 and a name column at that InputTable's index 1, either of the following calls will retrieve the column values associated with customer.name:

proc_data.input_data["customer"]["name"]
proc_data.input_data[5][1]

Unlike the other Kinetica APIs, the UDF Python API does not process data using records or schemas, operating in terms of columns of data instead. The raw column values returned closely map to the data types used in the tables being accessed:

Column Category	Column Type	UDF Value Type
Numeric	int	int
	int8	int
	int16	int
	long	int
	float	float
	double	float
	decimal	decimal.Decimal
String	string	str
	char[N]	str
	ipv4	int
Date/Time	date	datetime.date
	time	datetime.time
	timestamp	int
Binary	bytes	str

Column data values can be accessed through array indices:

column[i]

For example, to retrieve the value for the 10th record:

column[9]

To output data to a table, the size of the table must be set in order to allocate enough space in all of the columns to hold the correct number of values. To do this, call:

table.size = <total number of output records>

Table column values can then be assigned by index to each OutputColumn of each OutputTable.

Any output from the UDF to stdout is written to the /opt/gpudb/core/logs/gpudb.log file.

Request Info Keys

A variety of details about the executing UDF can be extracted from the request information map made available to each running proc. The full list of keys follows.

General Information

Map Key	Description
proc_name	The name of the proc (UDF) being executed
run_id	The run ID of the proc being executed; this is also displayed in Gadmin on the UDF page in the Status section as a link you can click on to get more detailed information; note that although this is an integer, it should not be relied upon as such, as its format may change in the future
rank_number	The rank number on which the current proc instance is executing; for distributed procs [1..n], for non-distributed procs 0
tom_number	The TOM number within the rank on which the current proc instance is executing; for distributed procs [0..n-1], where n is the number of TOMs per rank, for non-distributed procs it is not provided, since these do not run on a TOM
<option_name>	Any options passed in the options map in the /execute/proc request will also be in the request info map

Data Segment Information

Data is passed into procs in segments. Each segment consists of the entirety of the data on a single TOM and is processed by the proc instance executing on that TOM. Thus, there is a 1-to-1 mapping of data segment and executing proc instance, though this relationship may change in the future.

Running the same proc multiple times should result in the same set of segments, assuming the same environment and system state across runs.

Map Key	Description
data_segment_id	A unique identifier for the segment of the currently executing proc instance; all of the data segment IDs for a given proc execution are displayed in Gadmin when you click on the run ID; note that although it is possible to identify rank and TOM numbers from this ID, it should not be relied upon, as its format may change in the future
data_segment_count	The total cluster-wide count of data segments for distributed procs; for non-distributed procs 1
data_segment_number	The number of the current data segment or executing proc instance [0..data_segment_count-1]

Kinetica API Connection Parameters

These can be used to connect back to Kinetica using the regular API endpoint calls. Use with caution in distributed procs, particularly in large clusters, to avoid overwhelming the head node. Also note, multi-head ingest may not work from a proc in some cases without overriding the worker URLs to use internal IP addresses.

Map Key	Description
head_url	The URL to connect to
username	The username to connect as (corresponds to the user that called /execute/proc)
password	The password for the username

Note: username and password are not the actual login credentials of the user; they are randomly generated temporary credentials, which, for security reasons, should still not be printed or output to logs, etc., as they are live credentials for a period of time.