You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Fix memory leak when running topic compaction. (apache#6485)
Fixesapache#6482
### Motivation
Prevent topic compaction from leaking direct memory
### Modifications
Several leaks were discovered using Netty leak detection and code review.
* `CompactedTopicImpl.readOneMessageId` would get an `Enumeration` of `LedgerEntry`, but did not release the underlying buffers. Fix: iterate though the `Enumeration` and release underlying buffer. Instead of logging the case where the `Enumeration` did not contain any elements, complete the future exceptionally with the message (will be logged by Caffeine).
* Two main sources of leak in `TwoPhaseCompactor`. The `RawBacthConverter.rebatchMessage` method failed to close/release a `ByteBuf` (uncompressedPayload). Also, the return ByteBuf of `RawBacthConverter.rebatchMessage` was not closed. The first one was easy to fix (release buffer), to fix the second one and make the code easier to read, I decided to not let `RawBacthConverter.rebatchMessage` close the message read from the topic, instead the message read from the topic can be closed in a try/finally clause surrounding most of the method body handing a message from a topic (in phase two loop). Then if a new message was produced by `RawBacthConverter.rebatchMessage` we check that after we have added the message to the compact ledger and release the message.
### Verifying this change
Modified `RawReaderTest.testBatchingRebatch` to show new contract.
One can run the test described to reproduce the issue, to verify no leak is detected.
0 commit comments