[
https://issues.apache.org/jira/browse/GOBBLIN-2118?focusedWorklogId=928888&page=com.atlassian.jira.plugin.system.issuetabpanels:worklog-tabpanel#worklog-928888
]
ASF GitHub Bot logged work on GOBBLIN-2118:
-------------------------------------------
Author: ASF GitHub Bot
Created on: 06/Aug/24 10:48
Start Date: 06/Aug/24 10:48
Worklog Time Spent: 10m
Work Description: arpit09 commented on code in PR #4009:
URL: https://github.com/apache/gobblin/pull/4009#discussion_r1705323654
##########
gobblin-modules/gobblin-kafka-common/src/main/java/org/apache/gobblin/source/extractor/extract/kafka/KafkaSource.java:
##########
@@ -482,20 +487,54 @@ private static void skipWorkUnit(WorkUnit workUnit) {
workUnit.setProp(ConfigurationKeys.WORK_UNIT_HIGH_WATER_MARK_KEY,
workUnit.getLowWaterMark());
}
- private WorkUnit getWorkUnitForTopicPartition(KafkaPartition partition,
SourceState state,
+ /**
+ * Get the workunits of all the partitions passed, this method fetches all
the offsets for the partitions
+ * at once from kafka, and for each partiton creates a workunit.
+ * @param partitions
+ * @param state
+ * @param topicSpecificState
+ * @return
+ */
+ private Map<KafkaPartition, WorkUnit>
getWorkUnits(Collection<KafkaPartition> partitions, SourceState state,
Optional<State> topicSpecificState) {
- Offsets offsets = new Offsets();
-
- boolean failedToGetKafkaOffsets = false;
-
+ final Map<KafkaPartition, Offsets> partitionOffsetMap = Maps.newHashMap();
+ final Set<KafkaPartition> fetchOffsetsFailedPartitions = Sets.newHashSet();
try (Timer.Context context =
this.metricContext.timer(OFFSET_FETCH_TIMER).time()) {
- offsets.setOffsetFetchEpochTime(System.currentTimeMillis());
-
offsets.setEarliestOffset(this.kafkaConsumerClient.get().getEarliestOffset(partition));
-
offsets.setLatestOffset(this.kafkaConsumerClient.get().getLatestOffset(partition));
- } catch (Throwable t) {
- failedToGetKafkaOffsets = true;
- LOG.error("Caught error in creating work unit for {}", partition, t);
+ // Fetch the offsets for all the partitions at once
+ final Map<KafkaPartition, Long> earliestOffsetMap =
this.kafkaConsumerClient.get().getEarliestOffsets(partitions);
+ final Map<KafkaPartition, Long> latestOffsetMap =
this.kafkaConsumerClient.get().getLatestOffsets(partitions);
+ for (KafkaPartition partition : partitions) {
+ Offsets offsets = new Offsets();
+ offsets.setOffsetFetchEpochTime(System.currentTimeMillis());
+ if (earliestOffsetMap.containsKey(partition) &&
latestOffsetMap.containsKey(partition)) {
Review Comment:
Added, this is to check if both earliest offset and latest offset available
for the partitions. If there is a discrepancy, we failed to fetch the offsets
and we should move this to failedOffsetPartitionslist.
Added the same in the comments
Issue Time Tracking
-------------------
Worklog Id: (was: 928888)
Time Spent: 1h (was: 50m)
> Reduce no of network calls while fetching kafka offsets during startup
> ----------------------------------------------------------------------
>
> Key: GOBBLIN-2118
> URL: https://issues.apache.org/jira/browse/GOBBLIN-2118
> Project: Apache Gobblin
> Issue Type: Improvement
> Reporter: Arpit Varshney
> Priority: Major
> Time Spent: 1h
> Remaining Estimate: 0h
>
> During starting while creating work unit, in Kafkasource there are network
> calls that tries to fetch the kafka offsets (both earliest and latest) to
> find out the watermark (to find the offsets where the gobblin job will start
> consuming from)
> These calls are fetched for each topic and each partition in the topic. For
> each partition, there is a separate call that goes to kafka client, which
> increases the no of network calls. If there are cross colo calls (calls to
> different datacenters in different regions) this increase the time to fetch
> and results in timeout which leads to skipping of the topic partition to
> fetch leading to starvation.
> This ticket targets to reduce the no of network calls, rather than doing a
> call for each partition. Utilize kafka source to fetch the offsets for all
> the paritions at once from kafka.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
