spark-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From James Yu <>
Subject Re: Performance Problems Migrating to S3A Committers
Date Thu, 05 Aug 2021 21:22:06 GMT
See this ticket  It may help your team.
From: Johnny Burns <>
Sent: Tuesday, June 22, 2021 3:41 PM
To: <>
Cc: data-orchestration-team <>
Subject: Performance Problems Migrating to S3A Committers


I’m Johnny, I work at Stripe. We’re heavy Spark users and we’ve been exploring using
s3 committers. Currently we first write the data to HDFS and then upload it to S3. However,
now with S3 offering strong consistency guarantees, we are evaluating if we can write data
directly to S3.

We’re having some troubles with performance, so hoping someone might have some guidance
which can unblock this.

File Format
We are using parquet as the File Format. We do have iceberg tables as well, and they are indeed
able to commit directly to S3 (with minimal local disk usage). We can’t migrate all of our
jobs to iceberg right now. Hence, we are looking for a committer that is performant and can
directly write parquet files to S3 (with minimal local disk usage).
What have we tried?
We’ve tried using both the “magic” and “directory” committers. We're setting the
following configs (in addition to the "magic/directory"<>).





Both committers have shown performance regressions on large jobs. We’re currently focused
on trying to make the directory committer work because we’ve seen fewer slowdowns with that
one, but I’ll describe the problems with each.

We’ve been testing the committers on a large job with 100k tasks (creating 7.3TB of output).
Observations for magic committer

Using the magic committer, we see slowdowns in two places:

  *   S3 Writing (inside the task)

  *   The slowdown seems to occur just after the s3 multipart write. The finishedWrite<>
function tries to do some cleanup and kicks off this deleteUnnecessaryFakeDirectories<>

  *   This causes 503’s due to hitting AWS rate limits on

  *   I'm not sure what directories are actually getting cleaned up here (I assume the _magic
directories are still needed up until the job commit).

  *   Job Commit

  *   Have not dug down into the details here, but assume it is something similar to what
we’re seeing in the directory committer case below.

Observations for directory committer

We’ve observed that the “directory” s3committer performance is on-par with our existing
HDFS commit for task execution and task commit. The slowdowns we’re seeing are in the job
commit phase.

The job commit happens almost instantaneously in the HDFS case, vs taking about an hour for
the s3 directory committer.

We’ve enabled DEBUG logging for the s3 committer. It seems like that hour is mostly spent
doing things which you would expect (completing 100k delayedComplete s3 uploads). I've attached
an example of some of the logs we see repeated over-and-over during the 1 hour job commit
(I redacted some of the directories and SHAs but the logs are otherwise unchanged).

One thing I notice is that we see object_delete_requests += 1 in the logs. I’m not sure
if that means it’s doing an s3 delete, or it is deleting the HDFS manifest files (to clean
up the task).

Alternatives - Should we check out directCommitter?
We’ve also considered using the directCommitter. We understand that the directCommitter
is discouraged because it does not support speculative execution (and for some failure cases).
Given that we do not use speculative execution at Stripe, would the directCommitter be a viable
option for us? What are the failure scenarios to consider?

Alternatives - Can S3FileIO work well with parquet files?

Netflix has a tool called s3FileIO<>. We’re
wondering if it can be used with spark, or only with Iceburg.

View raw message