noob-se7en commented on code in PR #16492:
URL: https://github.com/apache/pinot/pull/16492#discussion_r2259208340
##########
pinot-controller/src/main/java/org/apache/pinot/controller/helix/core/realtime/PinotLLCRealtimeSegmentManager.java:
##########
@@ -796,9 +796,10 @@ private String createNewSegmentMetadata(TableConfig
tableConfig, IdealState idea
// segment. For tables with pauseless mode enabled, this size is
unavailable at this step because the
// segment has not yet been built.
- createNewSegmentZKMetadata(tableConfig, streamConfigs.get(0),
newLLCSegment, newSegmentCreationTimeMs,
+ createNewSegmentZKMetadataWithOffsetAutoReset(tableConfig,
streamConfigs.get(0), newLLCSegment,
+ newSegmentCreationTimeMs,
committingSegmentDescriptor, committingSegmentZKMetadata,
instancePartitions, partitionIds.size(),
- numReplicas);
+ numReplicas, false);
Review Comment:
If there are existing partitions, why this flag needs to be true for newly
added partitions if the new partition start offset is before the SLA?
##########
pinot-controller/src/main/java/org/apache/pinot/controller/helix/core/realtime/PinotLLCRealtimeSegmentManager.java:
##########
@@ -962,6 +962,72 @@ private void createNewSegmentZKMetadata(TableConfig
tableConfig, StreamConfig st
persistSegmentZKMetadata(realtimeTableName, newSegmentZKMetadata, -1);
}
+ /**
+ * Creates and persists segment ZK metadata for the new CONSUMING segment.
+ */
+ private void createNewSegmentZKMetadataWithOffsetAutoReset(TableConfig
tableConfig, StreamConfig streamConfig,
+ LLCSegmentName newLLCSegmentName, long creationTimeMs,
CommittingSegmentDescriptor committingSegmentDescriptor,
+ @Nullable SegmentZKMetadata committingSegmentZKMetadata,
InstancePartitions instancePartitions, int numPartitions,
+ int numReplicas, boolean skipAutoReset) {
+ String oldStartOffset = committingSegmentDescriptor.getNextOffset();
+ String startOffset = skipAutoReset ? oldStartOffset : computeStartOffset(
+ oldStartOffset, streamConfig, newLLCSegmentName.getPartitionGroupId());
+ createNewSegmentZKMetadata(tableConfig, streamConfig, newLLCSegmentName,
creationTimeMs,
+ committingSegmentDescriptor, committingSegmentZKMetadata,
instancePartitions, numPartitions, numReplicas,
+ startOffset);
+ }
+
+ private String computeStartOffset(
+ String nextOffset, StreamConfig streamConfig, int partitionId) {
+ long timeThreshold = streamConfig.getOffsetAutoResetTimeSecThreshold();
+ int offsetThreshold = streamConfig.getOffsetAutoResetOffsetThreshold();
+ if (timeThreshold <= 0 && offsetThreshold <= 0) {
+ return nextOffset;
+ }
+ String clientId =
+ PinotLLCRealtimeSegmentManager.class.getSimpleName() + "-" +
streamConfig.getTableNameWithType() + "-"
+ + streamConfig.getTopicName();
+ StreamConsumerFactory consumerFactory =
StreamConsumerFactoryProvider.create(streamConfig);
+ StreamPartitionMsgOffset offsetAtSLA;
+ StreamPartitionMsgOffset latestOffset;
+ try (StreamMetadataProvider metadataProvider =
consumerFactory.createPartitionMetadataProvider(
+ clientId, partitionId)) {
+ // Fetching timestamp from an offset is an expensive operation which
requires reading the data,
+ // while fetching offset from timestamp is lightweight and only needs to
read metadata.
+ // Hence, instead of checking if latestOffset's time - nextOffset's time
< SLA, we would check
+ // (CurrentTime - SLA)'s offset > nextOffset.
+ // TODO: it is relying on System.currentTimeMillis() which might be
affected by time drift. If we are able to
+ // get nextOffset's time, we should instead check (nextOffset's time +
SLA)'s offset < latestOffset
+ latestOffset =
metadataProvider.fetchStreamPartitionOffset(OffsetCriteria.LARGEST_OFFSET_CRITERIA,
5000);
+ LOGGER.info("Latest offset of topic {} and partition {} is {}",
+ streamConfig.getTopicName(), partitionId, latestOffset);
+ offsetAtSLA = metadataProvider.getOffsetAtTimestamp(
+ partitionId, System.currentTimeMillis() - timeThreshold * 1000);
+ LOGGER.info("Offset at SLA of topic {} and partition {} is {}",
+ streamConfig.getTopicName(), partitionId, offsetAtSLA);
+ } catch (Exception e) {
+ LOGGER.warn("Not able to fetch the offset metadata, skip auto resetting
offsets", e);
+ return nextOffset;
+ }
+ try {
+ if (timeThreshold > 0 && offsetAtSLA != null
+ && Long.valueOf(offsetAtSLA.toString()) > Long.valueOf(nextOffset)) {
Review Comment:
What I meant was to only rely on the internal implementation of `compareTo`
methods of the StreamPartitionMsgOffset (for kafka, kinesis, etc).
I see we are already creating StreamConsumerFactory object in this method,
we can use that to get `StreamPartitionMsgOffsetFactory` and convert string to
offset.
Or there might be other better ways.
--
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]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
