Goodness Ayinmode created IGNITE-25719: ------------------------------------------
Summary: Prevent OOM in GridDhtTxPrepareFuture and GridCacheMvcc
Key: IGNITE-25719
URL: https://issues.apache.org/jira/browse/IGNITE-25719
Project: Ignite
Issue Type: Improvement
Components: cache
Affects Versions: 2.16
Reporter: Goodness Ayinmode
In large transactions involving thousands of keys, the methods
[checkNeedRebalanceKeys()|https://github.com/apache/ignite/blob/2027c2ae1fe5bf516e1c6ec01764843e62b57a83/modules/core/src/main/java/org/apache/ignite/internal/processors/cache/distributed/dht/GridDhtTxPrepareFuture.java#L1142],
[forceRebalanceKeys()|https://github.com/apache/ignite/blob/2027c2ae1fe5bf516e1c6ec01764843e62b57a83/modules/core/src/main/java/org/apache/ignite/internal/processors/cache/distributed/dht/GridDhtTxPrepareFuture.java#L1171],
and
[readyLocks()|https://github.com/apache/ignite/blob/2027c2ae1fe5bf516e1c6ec01764843e62b57a83/modules/core/src/main/java/org/apache/ignite/internal/processors/cache/distributed/dht/GridDhtTxPrepareFuture.java#L662]
within GridDhtTxPrepareFuture can cause excessive memory usage and potential
OOM. These methods are called inside the
[prepare()|https://github.com/apache/ignite/blob/2027c2ae1fe5bf516e1c6ec01764843e62b57a83/modules/core/src/main/java/org/apache/ignite/internal/processors/cache/distributed/dht/GridDhtTxPrepareFuture.java#L1077]
where the number of keys they handle grows proportionally with the transaction
size and remain in memory throughout.
I also think that there is a risk of indefinite memory growth with
[locs|https://github.com/apache/ignite/blob/e457dcc98e53cf131a90db78d1d5814a1e3a565b/modules/core/src/main/java/org/apache/ignite/internal/processors/cache/GridCacheMvcc.java#L273]
inside
[GridCacheMvcc|https://github.com/apache/ignite/blob/2027c2ae1fe5bf516e1c6ec01764843e62b57a83/modules/core/src/main/java/org/apache/ignite/internal/processors/cache/GridCacheMvcc.java#L50],
which holds local lock candidates. Methods like
[add0()|https://github.com/apache/ignite/blob/e457dcc98e53cf131a90db78d1d5814a1e3a565b/modules/core/src/main/java/org/apache/ignite/internal/processors/cache/GridCacheMvcc.java#L267],
[candidates|https://github.com/apache/ignite/blob/e457dcc98e53cf131a90db78d1d5814a1e3a565b/modules/core/src/main/java/org/apache/ignite/internal/processors/cache/GridCacheMvcc.java#L1408],
and [localOwners()
|https://github.com/apache/ignite/blob/e457dcc98e53cf131a90db78d1d5814a1e3a565b/modules/core/src/main/java/org/apache/ignite/internal/processors/cache/GridCacheMvcc.java#L142]operate
over locs. Under failure scenarios like timeouts or transaction aborts locs,
cleanup logic (e.g.
[releaseLocal(|https://github.com/apache/ignite/blob/2027c2ae1fe5bf516e1c6ec01764843e62b57a83/modules/core/src/main/java/org/apache/ignite/internal/processors/cache/GridCacheMvcc.java#L1200]))
may be skipped or delayed. This can result in stale entries remaining in
{{{}locs{}}}, causing it to grow indefinitely and increasing the risk of memory
pressure or OOM.
Would it be possible and beneficial to introduce batching or timeout-based
cleanup to mitigate these potential memory issues?
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
