[ https://issues.apache.org/jira/browse/FLINK-22605?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]
Godfrey He updated FLINK-22605: ------------------------------- Fix Version/s: 1.17.0 (was: 1.16.0) > Python UDF should only be created once regardless of how many times it is > invoked > --------------------------------------------------------------------------------- > > Key: FLINK-22605 > URL: https://issues.apache.org/jira/browse/FLINK-22605 > Project: Flink > Issue Type: Improvement > Components: API / Python > Reporter: Yik San Chan > Priority: Minor > Labels: auto-deprioritized-major > Fix For: 1.17.0 > > > Follow up to > http://apache-flink-user-mailing-list-archive.2336050.n4.nabble.com/I-call-Pandas-UDF-N-times-do-I-have-to-initiate-the-UDF-N-times-tt43576.html. > Currently, when we call Python UDF N times, the Python UDF is constructed N > times. This may become a concern to performance, when we want to load large > resources in the open() method of the UDF, which is quite common in machine > learning use cases. > I propose we optimize in PyFlink framework level s.t. no matter how many > times a UDF is called in the execution environment, it is only initiated once. -- This message was sent by Atlassian Jira (v8.20.10#820010)