[CASSANDRA-11828] Commit log needs to track unflushed intervals rather than positions

• Type: Bug
• Status: Resolved
• Priority: Normal
• Resolution: Fixed
• Fix Version/s: 2.1.16, 2.2.8, 3.0.9, 3.8
• Component/s: Legacy/Local Write-Read Paths
• Severity: Normal
• Assignee: Branimir Lambov Assign to me
• Reporter: Branimir Lambov

Authors

Branimir Lambov

Reviewers

Sylvain Lebresne

• Created: 18/May/16 10:49
• Updated: 16/Apr/19 09:30
• Resolved: 05/Aug/16 13:49

Description

In CASSANDRA-11448 in an effort to give a more thorough handling of flush errors I have introduced a possible correctness bug with disk failure policy ignore if a flush fails with an error:

• we report the error but continue
• we correctly do not update the commit log with the flush position
• but we allow the post-flush executor to resume
• a successful later flush can thus move the log's clear position beyond the data from the failed flush
• the log will then delete segment(s) that contain unflushed data.

After CASSANDRA-9669 it is relatively easy to fix this problem by making the commit log track sets of intervals of unflushed data (as described in CASSANDRA-8496).

Issue Links

• blocks: CASSANDRA-8496 Remove MemtablePostFlusher
• is related to: CASSANDRA-11448 Running OOS should trigger the disk failure policy
• is related to: CASSANDRA-12436 Under some races commit log may incorrectly think it has unflushed data