[SPARK-46796][SS] Ensure the correct remote files (mentioned in metadata.zip) are used on RocksDB version load #44837
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
…data.zip) are used on RocksDB version load
sql/core/src/main/scala/org/apache/spark/sql/execution/streaming/state/RocksDB.scala
Show resolved
Hide resolved
HyukjinKwon
changed the title
|
Thanks! Merging to master/3.5. |
HeartSaVioR
pushed a commit
that referenced
this pull request
Jan 24, 2024
…ata.zip) are used on RocksDB version load This PR ensures that RocksDB loads do not run into SST file Version ID mismatch issue. RocksDB has added validation to ensure exact same SST file is used during database load from snapshot. Current streaming state suffers from certain edge cases where this condition is violated resulting in state load failure. The changes introduced are: 1. Ensure that the local SST file is exactly the same DFS file (as per mapping in metadata.zip). We keep track of the DFS file path for a local SST file, and re download the SST file in case DFS file has a different UUID in metadata zip. 2. Reset lastSnapshotVersion in RocksDB when Rocks DB is loaded. Changelog checkpoint relies on this version for future snapshots. Currently, if a older version is reloaded we were not uploading snapshots as lastSnapshotVersion was pointing to a higher snapshot of a cleanup database. We need to ensure that the correct SST files are used on executor during RocksDB load as per mapping in metadata.zip. With current implementation, its possible that the executor uses a SST file (with a different UUID) from a older version which is not the exact file mapped in the metadata.zip. This can cause version Id mismatch errors while loading RocksDB leading to streaming query failures. See https://issues.apache.org/jira/browse/SPARK-46796 for failure scenarios. No Added exhaustive unit testcases covering the scenarios. No Closes #44837 from sahnib/SPARK-46796. Authored-by: Bhuwan Sahni <bhuwan.sahni@databricks.com> Signed-off-by: Jungtaek Lim <kabhwan.opensource@gmail.com> (cherry picked from commit f25ebe5) Signed-off-by: Jungtaek Lim <kabhwan.opensource@gmail.com>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
What changes were proposed in this pull request?
This PR ensures that RocksDB loads do not run into SST file Version ID mismatch issue. RocksDB has added validation to ensure exact same SST file is used during database load from snapshot. Current streaming state suffers from certain edge cases where this condition is violated resulting in state load failure.
The changes introduced are:
Why are the changes needed?
We need to ensure that the correct SST files are used on executor during RocksDB load as per mapping in metadata.zip. With current implementation, its possible that the executor uses a SST file (with a different UUID) from a older version which is not the exact file mapped in the metadata.zip. This can cause version Id mismatch errors while loading RocksDB leading to streaming query failures.
See https://issues.apache.org/jira/browse/SPARK-46796 for failure scenarios.
Does this PR introduce any user-facing change?
No
How was this patch tested?
Added exhaustive unit testcases covering the scenarios.
Was this patch authored or co-authored using generative AI tooling?
No