Details
-
Bug
-
Status: Reopened
-
Minor
-
Resolution: Fixed
-
None
-
None
-
New
Description
CMS hard-stalls incoming threads for denial-of-service protection when merging cannot keep up with whatever is producing new segments.
When you call IW.rollback, it asks all merges to abort, and a running merge will periodically check to see if it should abort.
However, a stalled merge fails to check, which means rollback can take indefinitely long; I've seen this in Elasticsearch causing shutdown to take > 10 sec.