Shekhar Prasad Rajak created KAFKA-21097:
--------------------------------------------
Summary: Avoid duplicate share-partition next-fetch offset scans
Key: KAFKA-21097
URL: https://issues.apache.org/jira/browse/KAFKA-21097
Project: Kafka
Issue Type: Improvement
Components: core
Reporter: Shekhar Prasad Rajak
Assignee: Shekhar Prasad Rajak
Current flow :
DelayedShareFetch acquires the share-partition fetch lock, calls
canAcquireRecords(), then calls nextFetchOffset() to build the log read.
canAcquireRecords() already calls nextFetchOffset(), so a partition with
released or timed-out records can scan cachedState and per-offset state twice
under the write lock for one fetch.
Proposed flow
Add one atomic nextFetchOffsetIfAcquirable() call. Under one lock, it
computes the next fetch offset once and decides whether the partition can fetch
due
to available redelivery records or remaining in-flight capacity.
DelayedShareFetch uses that returned offset directly.
This removes the duplicate scan and lock gap without changing share-record
state transitions or introducing a new availability index.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
