lanyuanxiaoyao/hudi
1
0
Fork 0
Code Activity

[HUDI-2794] Guarding table service commits within a single lock to commit to both data table and metadata table (#4037)

Browse Source 
* Fixing a single lock to commit table services across metadata table and data table

* Addressing comments

* rebasing with master
This commit is contained in:
Sivabalan Narayanan
2021-11-25 14:19:30 -05:00
committed by GitHub
parent b972aa5bf2
commit 7bb90e8caf
4 changed files with 49 additions and 74 deletions
Download Patch File Download Diff File Expand all files Collapse all files

32
hudi-client/hudi-flink-client/src/main/java/org/apache/hudi/client/HoodieFlinkWriteClient.java
View File

@@ -362,12 +362,19 @@ public class HoodieFlinkWriteClient<T extends HoodieRecordPayload> extends
String compactionCommitTime) {
this.context.setJobStatus(this.getClass().getSimpleName(), "Collect compaction write status and commit compaction");
List<HoodieWriteStat> writeStats = writeStatuses.stream().map(WriteStatus::getStat).collect(Collectors.toList());
finalizeWrite(table, compactionCommitTime, writeStats);
// commit to data table after committing to metadata table.
writeTableMetadata(table, metadata, new HoodieInstant(HoodieInstant.State.INFLIGHT, HoodieTimeline.COMPACTION_ACTION, compactionCommitTime));
LOG.info("Committing Compaction {} finished with result {}.", compactionCommitTime, metadata);
CompactHelpers.getInstance().completeInflightCompaction(table, compactionCommitTime, metadata);
try {
HoodieInstant compactionInstant = new HoodieInstant(HoodieInstant.State.INFLIGHT, HoodieTimeline.COMPACTION_ACTION, compactionCommitTime);
this.txnManager.beginTransaction(Option.of(compactionInstant), Option.empty());
finalizeWrite(table, compactionCommitTime, writeStats);
// commit to data table after committing to metadata table.
// Do not do any conflict resolution here as we do with regular writes. We take the lock here to ensure all writes to metadata table happens within a
// single lock (single writer). Because more than one write to metadata table will result in conflicts since all of them updates the same partition.
table.getMetadataWriter().ifPresent(w -> w.update(metadata, compactionInstant.getTimestamp(), table.isTableServiceAction(compactionInstant.getAction())));
LOG.info("Committing Compaction {} finished with result {}.", compactionCommitTime, metadata);
CompactHelpers.getInstance().completeInflightCompaction(table, compactionCommitTime, metadata);
} finally {
this.txnManager.endTransaction();
}
if (compactionTimer != null) {
long durationInMs = metrics.getDurationInMs(compactionTimer.stop());
try {
@@ -399,19 +406,6 @@ public class HoodieFlinkWriteClient<T extends HoodieRecordPayload> extends
throw new HoodieNotSupportedException("Clustering is not supported yet");
}
private void writeTableMetadata(HoodieTable<T, List<HoodieRecord<T>>, List<HoodieKey>, List<WriteStatus>> table,
HoodieCommitMetadata commitMetadata,
HoodieInstant hoodieInstant) {
try {
this.txnManager.beginTransaction(Option.of(hoodieInstant), Option.empty());
// Do not do any conflict resolution here as we do with regular writes. We take the lock here to ensure all writes to metadata table happens within a
// single lock (single writer). Because more than one write to metadata table will result in conflicts since all of them updates the same partition.
table.getMetadataWriter().ifPresent(w -> w.update(commitMetadata, hoodieInstant.getTimestamp(), table.isTableServiceAction(hoodieInstant.getAction())));
} finally {
this.txnManager.endTransaction();
}
}
@Override
protected HoodieTable<T, List<HoodieRecord<T>>, List<HoodieKey>, List<WriteStatus>> getTableAndInitCtx(WriteOperationType operationType, String instantTime) {
HoodieTableMetaClient metaClient = createMetaClient(true);