You are viewing an old version of this page. View the current version.

Compare with Current View Page History

« Previous Version 7 Next »

Status

Current state: Under Discussion

Discussion thread: here

JIRA: KAFKA-14112

Please keep the discussion on the mailing list rather than commenting on the wiki (wiki discussions get unwieldy fast).

Motivation

The current Kafka architecture lacks a built-in mechanism to directly track the replication offset lag, i.e the number of to-be-replicated records. This could be essential for monitoring and maintaining the health and performance of data replication processes. The replication offset lag, defined as the difference between the last end offset of the source partition (LEO) and the last replicated source offset (LRO), is beneficial for understanding the progress and potential bottlenecks in data replication scenarios.

Public Interfaces

  • org/apache/kafka/connect/mirror/MirrorMetrics - new metrics templates for max, min, avg and value replication-offset-lag

    MirrorMetrics.java
     replicationOffsetLag = new MetricNameTemplate(

 "replication-offset-lag", SOURCE_CONNECTOR_GROUP,

 "Count of records to be replicated from source to target cluster.", partitionTags);

 replicationOffsetLagMax = new MetricNameTemplate(

 "replication-offset-lag-max", SOURCE_CONNECTOR_GROUP,

 "Max count of records to be replicated from source to target cluster.", partitionTags);

 replicationOffsetLagMin = new MetricNameTemplate(

 "replication-offset-lag-min", SOURCE_CONNECTOR_GROUP,

 "Min count of records to be replicated from source to target cluster.", partitionTags);

 replicationOffsetLagAvg = new MetricNameTemplate(

 "replication-offset-lag-avg", SOURCE_CONNECTOR_GROUP,

 "Average count of records to be replicated from source to target cluster.", partitionTags)

Proposed Changes

This proposal aims to enhance Kafka's monitoring capabilities by introducing a new metric to track the replication offset lag for a given topic-partition. The metric will be calculated by taking the difference between the LEO, which will be fetched during the source task's poll loop, and the LRO, which will be stored in an in-memory "cache" and updated during the task's producer callback.

The proposed changes involve the following steps:

  1. LRO Tracking Mechanism:

    • Introduce a new in-memory "cache" to store the LRO for each topic-partition.
    • Update the producer callback logic to keep this LRO cache up-to-date.

  2. LEO Acquisition during Poll Loop:

    • Modify the source task's poll loop to retrieve the LEO from the consumer.
    • Calculate the replication offset lag as LEO - LRO 

  3. Expose Replication Offset Lag Metric:

    • Introduce a new MirrorMaker metric, named "replication_offset_lag", which represents the calculated replication offset lag.

Compatibility, Deprecation, and Migration Plan

  • This proposal ensures backward compatibility by introducing new metrics and modifying existing poll loop and callback mechanisms we do not impact existing Kafka features, APIs, or configurations.

Test Plan

Metrics will be tested in org.apache.kafka.connect.mirror.MirrorSourceTaskTest in a unit test fashion. Unfortunately there are no system tests that I could find where this change could be tested, but I am happy to do it if there is a suitable place where it is already done and that I have missed, or to implement a new system test.

Rejected Alternatives

It might be argued that the existing "replication-latency-ms" metric already provides satisfactory information about the replication lag, but I think providing an exact amount of record lagging behind is beneficial in addition to providing a time-lag.

No other alternatives were considered.

  • No labels