Nikolay Izhikov created IGNITE-19369:
----------------------------------------

             Summary: metadata topic offset must be stored only after commit
                 Key: IGNITE-19369
                 URL: https://issues.apache.org/jira/browse/IGNITE-19369
             Project: Ignite
          Issue Type: Improvement
          Components: extensions
            Reporter: Ilya Shishkov
            Assignee: Ilya Shishkov


Currently, when CDC through Kafka is used there are possible delays in 
replication between clusters when {{KafkaToIgniteCdcStreamerApplier}} tries to 
update binary metadata and marshaller mappings.

Delays caused by calls of {{KafkaConsumer#poll}} in 
{{KafkaToIgniteMetadataUpdater#updateMetadata}} , when meta topic is empty:
# When first {{KafkaToIgniteCdcStreamerApplier}} meets {{META_UPDATE_MARKER}} 
it calls {{KafkaToIgniteMetadataUpdater#updateMetadata}} which in turn calls 
{{KafkaConsumer#poll}}, which returns immediately [1] when data is present in 
metadata topic. If there are few binary types and mappings to update, first 
{{KafkaToIgniteCdcStreamerApplier}} will consume all entries from metadata 
topic. 
# {{KafkaToIgniteCdcStreamerApplier}} consequently call 
{{KafkaToIgniteMetadataUpdater#updateMetadata}} for each partition with meta 
update marker.  All further consequent calls will wait for {{kafkaReqTimeout}}.
# Also there is a bottleneck, when multiple applier threads tries to update 
metadata and call synchronized method 
{{KafkaToIgniteMetadataUpdater#updateMetadata}}, because 
{{KafkaToIgniteMetadataUpdater}} is shared between applier threads.
# Because {{META_UPDATE_MARKER}} is sent twice to each Kafka partition of event 
topic from every node: firstly, in case of type mappings updates, secondly, in 
case of binary types update there are possible delays up to {{clusterSize x 
(topicPartitions x 2 - 1) x kafkaReqTimeout}}. 
# Data updates are blocked for Kafka partitions with unprocessed update markers.
# For example for default timeout and 16 Kafka partitions _last partition will 
be consumed after 1.5 minutes_ in case of two one-node clusters.

Links:
 # 
[https://kafka.apache.org/27/javadoc/org/apache/kafka/clients/consumer/KafkaConsumer.html#poll-java.time.Duration-]



--
This message was sent by Atlassian Jira
(v8.20.10#820010)

Reply via email to