showuon commented on code in PR #13561:
URL: https://github.com/apache/kafka/pull/13561#discussion_r1281568519
##########
core/src/main/java/kafka/log/remote/RemoteLogManager.java:
##########
@@ -618,6 +625,230 @@ public void run() {
}
}
+ public void handleLogStartOffsetUpdate(TopicPartition topicPartition,
long remoteLogStartOffset) {
+ if (isLeader()) {
Review Comment:
I agree with @divijvaidya 's suggestion. In this scenario:
1. replica 1 is the leader, and doing remote log segment deletion
2. leadership changed to replica 2
3. replica 1 entering this `handleLogStartOffsetUpdate` method
Under current implementation, we won't update log start offset since it is
not the leader anymore. But we should update it! @satishd , thoughts?
##########
core/src/main/java/kafka/log/remote/RemoteLogManager.java:
##########
@@ -634,6 +642,241 @@ public void run() {
}
}
+ public void handleLogStartOffsetUpdate(TopicPartition topicPartition,
long remoteLogStartOffset) {
+ if (isLeader()) {
+ logger.debug("Updating {} with remoteLogStartOffset: {}",
topicPartition, remoteLogStartOffset);
+ updateRemoteLogStartOffset.accept(topicPartition,
remoteLogStartOffset);
+ }
+ }
+
+ class RemoteLogRetentionHandler {
+
+ private final Optional<RetentionSizeData> retentionSizeData;
+ private final Optional<RetentionTimeData> retentionTimeData;
+
+ private long remainingBreachedSize;
+
+ private OptionalLong logStartOffset = OptionalLong.empty();
+
+ public RemoteLogRetentionHandler(Optional<RetentionSizeData>
retentionSizeData, Optional<RetentionTimeData> retentionTimeData) {
+ this.retentionSizeData = retentionSizeData;
+ this.retentionTimeData = retentionTimeData;
+ remainingBreachedSize = retentionSizeData.map(sizeData ->
sizeData.remainingBreachedSize).orElse(0L);
+ }
+
+ private boolean
deleteRetentionSizeBreachedSegments(RemoteLogSegmentMetadata metadata) throws
RemoteStorageException, ExecutionException, InterruptedException {
+ if (!retentionSizeData.isPresent()) {
+ return false;
+ }
+
+ boolean isSegmentDeleted = deleteRemoteLogSegment(metadata, x
-> {
+ // Assumption that segments contain size >= 0
+ if (remainingBreachedSize > 0) {
+ long remainingBytes = remainingBreachedSize -
x.segmentSizeInBytes();
+ if (remainingBytes >= 0) {
+ remainingBreachedSize = remainingBytes;
+ return true;
+ }
+ }
+
+ return false;
+ });
+ if (isSegmentDeleted) {
+ logStartOffset = OptionalLong.of(metadata.endOffset() + 1);
+ logger.info("Deleted remote log segment {} due to
retention size {} breach. Log size after deletion will be {}.",
+ metadata.remoteLogSegmentId(),
retentionSizeData.get().retentionSize, remainingBreachedSize +
retentionSizeData.get().retentionSize);
+ }
+ return isSegmentDeleted;
+ }
+
+ public boolean
deleteRetentionTimeBreachedSegments(RemoteLogSegmentMetadata metadata)
+ throws RemoteStorageException, ExecutionException,
InterruptedException {
+ if (!retentionTimeData.isPresent()) {
+ return false;
+ }
+
+ boolean isSegmentDeleted = deleteRemoteLogSegment(metadata,
+ x -> x.maxTimestampMs() <=
retentionTimeData.get().cleanupUntilMs);
+ if (isSegmentDeleted) {
+ remainingBreachedSize = Math.max(0, remainingBreachedSize
- metadata.segmentSizeInBytes());
+ // It is fine to have logStartOffset as
`metadata.endOffset() + 1` as the segment offset intervals
+ // are ascending with in an epoch.
+ logStartOffset = OptionalLong.of(metadata.endOffset() + 1);
+ logger.info("Deleted remote log segment {} due to
retention time {}ms breach based on the largest record timestamp in the
segment",
+ metadata.remoteLogSegmentId(),
retentionTimeData.get().retentionMs);
+ }
+ return isSegmentDeleted;
+ }
+
+ private boolean
deleteLogStartOffsetBreachedSegments(RemoteLogSegmentMetadata metadata, long
startOffset)
+ throws RemoteStorageException, ExecutionException,
InterruptedException {
+ boolean isSegmentDeleted = deleteRemoteLogSegment(metadata, x
-> startOffset > x.endOffset());
+ if (isSegmentDeleted && retentionSizeData.isPresent()) {
+ remainingBreachedSize = Math.max(0, remainingBreachedSize
- metadata.segmentSizeInBytes());
+ logger.info("Deleted remote log segment {} due to log
start offset {} breach", metadata.remoteLogSegmentId(), startOffset);
+ }
+
+ // No need to update the logStartOffset.
+ return isSegmentDeleted;
+ }
+
+ // It removes the segments beyond the current leader's earliest
epoch. Those segments are considered as
+ // unreferenced because they are not part of the current leader
epoch lineage.
+ private boolean
deleteLogSegmentsDueToLeaderEpochCacheTruncation(EpochEntry earliestEpochEntry,
RemoteLogSegmentMetadata metadata) throws RemoteStorageException,
ExecutionException, InterruptedException {
+ boolean isSegmentDeleted = deleteRemoteLogSegment(metadata, x
->
+
x.segmentLeaderEpochs().keySet().stream().allMatch(epoch -> epoch <
earliestEpochEntry.epoch));
+ if (isSegmentDeleted) {
+ logger.info("Deleted remote log segment {} due to leader
epoch cache truncation. Current earliest epoch: {}, segmentEndOffset: {} and
segmentEpochs: {}",
+ metadata.remoteLogSegmentId(), earliestEpochEntry,
metadata.endOffset(), metadata.segmentLeaderEpochs().keySet());
+ }
+
+ // No need to update the log-start-offset as these
epochs/offsets are earlier to that value.
+ return isSegmentDeleted;
+ }
+
+ private boolean deleteRemoteLogSegment(RemoteLogSegmentMetadata
segmentMetadata, Predicate<RemoteLogSegmentMetadata> predicate)
+ throws RemoteStorageException, ExecutionException,
InterruptedException {
+ if (predicate.test(segmentMetadata)) {
Review Comment:
+1, or at least a debug level.
##########
core/src/main/java/kafka/log/remote/RemoteLogManager.java:
##########
@@ -581,11 +588,18 @@ public void run() {
if (isLeader()) {
// Copy log segments to remote storage
copyLogSegmentsToRemote();
+ // Cleanup/delete expired remote log segments
+ cleanupExpiredRemoteLogSegments();
Review Comment:
@jeqo , What Satish meant, is in most remote storage case, the deletion API
won't wait until data deleted in remote storage, instead, it'll mark file as
deleted and return immediately. And run background GC in remote storage to
delete the deleted flagged file.
--
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.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
