[
https://issues.apache.org/jira/browse/BEAM-13268?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17445154#comment-17445154
]
Sam Whittle commented on BEAM-13268:
------------------------------------
Actually it looks like Pablo improved the BQ services to issue parallel rpcs
for larger batches. So I think we can just ungroup the grouped input and rely
on that parallelization. That also cleans up some code duplication.
> Reduce latency by parallelizing BQ inserts when flushing due to row limit
> -------------------------------------------------------------------------
>
> Key: BEAM-13268
> URL: https://issues.apache.org/jira/browse/BEAM-13268
> Project: Beam
> Issue Type: Improvement
> Components: io-java-gcp
> Reporter: Sam Whittle
> Assignee: Sam Whittle
> Priority: P2
>
> InsertBatchedElements consumes the output of GroupIntoBatches which flushes
> after default 500 elements to respect the max items per streaming insert
> request to BQ.
> However InsertBatchedElements flushes rows synchrounously meaning that
> latencies of writes accumulates. It could instead initiate writes in
> ProcessElement and block on write completion in FinishBundle. There could be
> some limited configurable parallelism if desired to limit memory usage.
--
This message was sent by Atlassian Jira
(v8.20.1#820001)
