Alexandre Dupriez created KAFKA-15678:
-----------------------------------------
Summary: [Tiered Storage] Stall remote reads with long-spanning
transactions
Key: KAFKA-15678
URL: https://issues.apache.org/jira/browse/KAFKA-15678
Project: Kafka
Issue Type: Bug
Components: Tiered-Storage
Affects Versions: 3.6.0
Reporter: Alexandre Dupriez
Hi team,
I am facing an issue on the remote data path for uncommitted reads.
As mentioned in [the original
PR|https://github.com/apache/kafka/pull/13535#discussion_r1166887367], if a
transaction spans over a long sequence of segments, the time taken to retrieve
the producer snapshots from the remote storage can, in the worst case, become
redhibitory and block the reads if it consistently exceed the deadline of fetch
requests ({{{}fetch.max.wait.ms{}}}).
Essentially, the method used to compute the uncommitted records to return have
an asymptotic complexity proportional to the number of segments in the log.
That is not a problem with local storage since the constant factor to traverse
the files is small enough, but that is not the case with a remote storage which
exhibits higher read latency. An aggravating factor was the lock contention in
the remote index cache which was then mitigated by KAFKA-15084. But
unfortunately, despite the improvements observed without said contention, the
algorithmic complexity of the current method used to compute uncommitted
records can always defeat any optimisation made on the remote read path.
Maybe we could start thinking (if not already) about a different construct
which would reduce that complexity to O(1) - i.e. to make the computation
independent from the number of segments and irrespective of the transaction
spans.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
