[KUDU-1131] Crash in compaction due to overlapping flush/undo snapshots

• Type: Bug
• Status: Resolved
• Priority: Critical
• Resolution: Fixed
• Affects Version/s: Private Beta
• Fix Version/s: 0.9.0
• Component/s: tablet
• Labels: crash
• Target Version/s: 0.9.0
• Code Review: http://gerrit.cloudera.org:8080/#/c/3073/
• Assignee: Todd Lipcon
• Reporter: Todd Lipcon
• Created: 11/Sep/15 02:52
• Updated: 17/May/16 04:53
• Resolved: 17/May/16 04:53

Description

Binglin is triggering a crash reasonably regularly under load:

• a tablet is flushed with a snapshot that has at least one txn in flight, but a txn with a later timestamp already committed. eg:
  • txn 1 and 3 committed, 2 in flight. This gives a flush snapshot txn <= 1 or txn == 3.
• as of KUDU-987, we don't wait for all in-flight transactions to commit during flush (necessary since the txn might be in flight for a while)
• because txn 3 was committed, the UNDO delta has a ts range of [1, 3]
• we then select the newly-flushed rowset for compaction, and txn 2 is still not committed
  • at this point, we hit a CHECK failure because we see an UNDO file which can't be fully ignored by a compaction (its time range overlaps with uncommitted ranges in the current snapshot)