Fix performance bug in Java API get() functions when key is NotFound #13049
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
trazfr
added a commit
to trazfr/rocksdbjni-miss-perfs
that referenced
this pull request
Oct 3, 2024
alanpaxton
force-pushed
the
eb/expensive-jni-misses-13023-jmh-fix
branch
from
3519c07
to
e028be5
Compare
|
Hello @alanpaxton, looks good for me. Only one question, can merge operation exit with status not found? The reason why I'm asking is because we changes signature of method
|
@rhubner No, I can't see how NotFound can come back from a merge. It's always a valid operation to merge with no previously written value, if I understand the semantics correctly. |
alanpaxton
force-pushed
the
eb/expensive-jni-misses-13023-jmh-fix
branch
from
e028be5
to
479acd0
Compare
Performance of NotFound case has been bad - see facebook#13023 Some jmh tests to measure this, prior to fixing it.
Handling NotFound should not go through the (horribly slow) C++ exception mechanism, because it is the expected/standard path in many cases. Change KVException::ThrowOnError to return a status when ok() or NotFound(), and the caller has to handle that status. Retain the exception mechanism for exceptions, it does simplifiy the flow and reduce checking.
get / getDirect / getForUpdate Repair performance by avoiding the C++ exception path on NotFound. Previous commit on this change altered behaviour of ROCKSDB_NAMESPACE::KVException::ThrowOnError to not throw on NotFound, this change is also necessary to reinstate correct behaviour.
Add extra tests for getDirect() —> NotFound where a potential gap existed Add a couple of extra tests around NotFound for comfort
alanpaxton
force-pushed
the
eb/expensive-jni-misses-13023-jmh-fix
branch
from
479acd0
to
b2b87da
Compare
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.
When a key did not exist in the system, various
get()calls from the Java API were using an exception-based return path within the C++ stubs used by the Java/JNI API, because we considered that a non-existent key was a problem. But it's a common pattern and throwing an exception was reducing the overall performance of NotFoundget()s by about 3x in the cached case we measured.We have removed NotFound from the error path; exceptions are still used to handle other exceptional cases, but we explicitly handle
ok()andIsNotFound()status from the underlyingget()calls without using exceptions. This alsoapplies to various transaction versions of
get().Finally we noted (provoked in our testing) a couple of cases where some variants of calls to get() a key were not tested for non-existent keys, and we added tests.
Closes #13023