mbutrovich opened a new pull request, #3553:
URL: https://github.com/apache/datafusion-comet/pull/3553
## Which issue does this PR close?
<!--
We generally require a GitHub issue to be filed for all bug fixes and
enhancements and this helps us generate change logs for our releases. You can
link an issue to this PR using the GitHub syntax. For example `Closes #123`
indicates that this PR will close issue #123.
-->
N/A.
## Rationale for this change
<!--
Why are you proposing this change? If this is already explained clearly in
the issue then this section is not needed.
Explaining clearly why changes are proposed helps reviewers understand your
changes and offer better suggestions for fixes.
-->
Seeing heavy CPU utilization on I/O, latency-bound workloads (e.g., 10s of
thousands of small Iceberg FileScanTasks on an object store). It looks like a
busy-poll, but we've done some work to try to address that already (#2937,
#2938, #3063).
I think it's because the task executor thread does a `block_on` the stream,
it's Pending, and then we await. However, if the I/O tasks aren't done, we just
wake up again, check the stream, it's still Pending, and we await again. This
essentially degrades to behavior equivalent to a busy-poll, and results in a
ton of scheduling overhead in tokio.
## What changes are included in this PR?
<!--
There is no need to duplicate the description in the issue here but it is
sometimes worth providing a summary of the individual changes in this PR.
-->
For scenarios where we don't have any scans that need to pull batches from
JVM, we set up a channel and pass that task into the tokio worker pool. This
lets the tokio worker pool handle stream execution and allows the executor task
thread from the JVM (that made the call into `executePlan` in jni_api.rs) to
properly wait on a batch arriving.
## How are these changes tested?
<!--
We typically require tests for all PRs in order to:
1. Prevent the code from being accidentally broken by subsequent changes
2. Serve as another way to document the expected behavior of the code
If tests are not included in your PR, please explain why (for example, are
they covered by existing tests)?
-->
Existing tests. I also benchmarked before and after with a workload that 1)
creates an Iceberg table in Minio (S3) with 10,000 small data files 2) runs a
query that reads all of the data. I couldn't simulate the latency added by the
object store, but you can see the difference:
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
