klion26 commented on a change in pull request #10329:
[FLINK-12785][StateBackend] RocksDB savepoint recovery can use a lot of
unmanaged memory
URL: https://github.com/apache/flink/pull/10329#discussion_r354262325
##########
File path:
flink-state-backends/flink-statebackend-rocksdb/src/main/java/org/apache/flink/contrib/streaming/state/RocksDBWriteBatchWrapper.java
##########
@@ -113,4 +126,16 @@ public void close() throws RocksDBException {
}
IOUtils.closeQuietly(batch);
}
+
+ private void flushIfNeeded() throws RocksDBException {
+ boolean needFlush = batch.count() == capacity || (batchSize > 0
&& batch.getDataSize() >= batchSize);
+ if (needFlush) {
+ flush();
+ }
+ }
+
+ @VisibleForTesting
+ long getDataSize() {
+ return batch.getDataSize();
Review comment:
Using our own implemented `getDataSize()` we can gain 2% performance but
need to track the format envolving in RocksDB, on the other hand, using JNI
call directly we don't need to track the format envolving in RocksDB,
As the performance gained is not much big, so I'm biased on using the JNI
call directly also.
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
With regards,
Apache Git Services
