[
https://issues.apache.org/jira/browse/CASSANDRA-21455?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Sam Tunnicliffe updated CASSANDRA-21455:
----------------------------------------
Change Category: Operability
Complexity: Normal
Fix Version/s: 6.x
Reviewers: Sam Tunnicliffe
Status: Open (was: Triage Needed)
I've combined the patches for CASSANDRA-21454, CASSANDRA-21455 &
CASSANDRA-21456 and am running CI against that branch.
> Unable to catch up TCM Log from peer with gaps in log sequence
> --------------------------------------------------------------
>
> Key: CASSANDRA-21455
> URL: https://issues.apache.org/jira/browse/CASSANDRA-21455
> Project: Apache Cassandra
> Issue Type: Improvement
> Components: Transactional Cluster Metadata
> Reporter: Sam Tunnicliffe
> Assignee: Jon Meredith
> Priority: Normal
> Fix For: 6.x
>
>
> The triggering scenario for this is a non-CMS peer receiving a snapshot when
> catching up from a peer/CMS node, leaving a gap near the tail of its local
> metadata log. If that node then subsequently receives catchup requests itself
> the heuristic that prioritises sending recent log entries over recent
> snapshots is counterproductive.
> A node receives log entries with epochs A, B, misses entries C, D, E and
> catches up from a peer, receiving snapshot at epoch F
> Later, when another node asks this peer for its log since epoch B:
> 1. listSnapshotsSince(B) returns exactly 1 snapshot (at epoch F)
> 2. Since snapshotEpochs.size() <= 1, the code tries getEntries(B) - gets
> empty (or entries only up to B with nothing after)
> 3. Empty entries are vacuously isContinuous() → returns LogState(null, [])
> without the snapshot
> 4. The requesting node is stuck - it keeps getting empty responses and never
> advances past the gap
> Proposed fix is in LogReader.getLogState. After confirming entries are
> continuous, check if there's a snapshot beyond the latest entry. If so, fall
> through to serve the snapshot instead.{quote}
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
