MaxCompute allows you to write user-defined aggregate functions (UDAFs) in Java or Python to extend the capabilities of MaxCompute functions and accommodate your business requirements. This topic describes the types, limits, usage notes, and development process of UDAFs. This topic also describes how to use UDAFs.
Background information
A many-to-one mapping is established between the input and output data of a UDAF. Multiple input records are aggregated to generate one output value. MaxCompute allows you to write UDAFs in Java or Python. The following table describes the two types of UDAFs.
| UDAF type | Description |
|---|---|
| Java UDAF | This type of UDAF is written in Java to implement the function logic. For more information, see Java UDAFs. |
| Python UDAF | This type of UDAF is written in Python to implement the function logic. Python UDAFs are classified into Python 2 UDAFs and Python 3 UDAFs.
|
Limits
You cannot access the Internet by using UDFs. If you want to access the Internet by using UDFs, fill in the network connection application from based on your business requirements and submit the application. After the application is approved, the MaxCompute technical support team will contact you and help you establish network connections. For more information about how to fill in the network connection application form, see Network connection process.
Usage notes
Before you use UDFs, take note of the following items:
- UDFs cannot compete with built-in functions in performance. We recommend that you preferentially use built-in functions to implement your business logic.
- If you use a UDF in SQL statements, the memory usage of a computing job may exceed the default allocated memory size if a large amount of data is computed and data skew occurs. In this case, you can run the
set odps.sql.udf.joiner.jvm.memory=xxxx;command at the session level to resolve the issue. For more information about the MaxCompute UDF FAQ, see FAQ about MaxCompute UDFs. - If the name of a UDF is the same as that of a built-in function, the UDF is preferentially called. For example, if UDF CONCAT and built-in function CONCAT both exist in MaxCompute, the system automatically calls UDF CONCAT instead of the built-in function CONCAT. If you want to call the built-in function, you must add the symbol
::before the built-in function, for example,select ::concat('ab', 'c');.
Development process
This section describes the development process of a UDAF.
- The following figure demonstrates how to write a MaxCompute UDF in Java.
No. Required Description Platform References 1 No Before you can use Maven to write code, you must add the related SDK dependencies to the POM file. This ensures that the code can be compiled. The following SDK dependency shows an example:
You can search for<dependency> <groupId>com.aliyun.odps</groupId> <artifactId>odps-sdk-udf</artifactId> <version>0.29.10-public</version> </dependency>odps-sdk-udffrom Maven repositories to obtain the version of the SDK dependency.IntelliJ IDEA (Maven) None 2 Yes Write a UDF based on your business requirements. IntelliJ IDEA (Maven) and MaxCompute Studio Develop a UDF in Java 3 Yes Debug the UDF by running it on your on-premises machine or by performing unit testing to check whether the result meets expectations. 4 Yes Debug the UDF code to ensure that the code is packaged into a JAR file after it is successfully run on your on-premises machine. 5 Yes Upload the JAR file as a resource to your MaxCompute project. MaxCompute client (odpscmd), MaxCompute Studio, and DataWorks - MaxCompute client
- MaxCompute Studio
Package a Java program, upload the package, and create a MaxCompute UDF
- DataWorks
6 Yes Create a UDF based on the JAR file that you uploaded. 7 No Call the UDF in the query data code. None - The following figure demonstrates how to write a MaxCompute UDF in Python.
No. Required Description Platform References 1 Yes Write a UDF based on your business requirements. MaxCompute Studio Develop a Python UDF 2 Yes Debug the UDF by running it on your on-premises machine or by performing unit testing to check whether the result meets expectations. 3 Yes Upload Python files or required resources, such as file resources, table resources, and third-party packages, to a MaxCompute project. MaxCompute client (odpscmd), MaxCompute Studio, and DataWorks - MaxCompute client
- MaxCompute Studio
- DataWorks
4 Yes Create a UDF based on the uploaded Python files or required resources. 5 No Call the UDF in the query data code. None
Instructions
Use the following methods to call UDFs:
- Use a UDF in a MaxCompute project: The method is similar to that of using built-in functions.
- Use a UDF across projects: Use a UDF of Project B in Project A. The following statement shows an example:
select B:udf_in_other_project(arg0, arg1) as res from table_t;. For more information about resource sharing across projects, see Cross-project resource access based on packages.