[
https://issues.apache.org/jira/browse/FLINK-31623?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Fan Hong updated FLINK-31623:
-----------------------------
Description:
Current implementation employs two-level sampling method.
However, when data instances are imbalanced distributed among partitions
(subtasks), the probabilities of instances to be sampled are different in
different partitions (subtasks), i.e., not a uniform sampling.
In addition, one side-effect of current implementation is: one subtask has a
memory footprint of `2 * numSamples * sizeof(element)`, which could cause
unexpected OOM in some situations.
was:
Current implementation employs two-level sampling method.
However, when data instances are imbalanced distributed among partitions
(subtasks), the probabilities of instances to be sampled are different in
different partitions (subtasks), i.e., not a uniform sampling.
In addition, one side-effect of current implementation is: one subtask has a
memory footprint twice as large as the other subtasks, which could cause
unexpected OOM in some situations.
> Change to uniform sampling in DataStreamUtils#sample method
> -----------------------------------------------------------
>
> Key: FLINK-31623
> URL: https://issues.apache.org/jira/browse/FLINK-31623
> Project: Flink
> Issue Type: Bug
> Components: Library / Machine Learning
> Reporter: Fan Hong
> Priority: Major
>
> Current implementation employs two-level sampling method.
> However, when data instances are imbalanced distributed among partitions
> (subtasks), the probabilities of instances to be sampled are different in
> different partitions (subtasks), i.e., not a uniform sampling.
>
> In addition, one side-effect of current implementation is: one subtask has a
> memory footprint of `2 * numSamples * sizeof(element)`, which could cause
> unexpected OOM in some situations.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
