Details
-
Improvement
-
Status: Resolved
-
Major
-
Resolution: Fixed
-
None
-
None
-
None
Description
There are a number of errors that currently crash Kudu tablet servers (e.g. disk failure). In the push to keep tablet servers alive in spite of these failures, the affected tablets should no longer service any type of request. Writes should not proceed, scans should be bounced to another tablet server, flushes and compactions should exit early, etc. The tablet should act as though it were deleted, with the exception that its data is not yet deleted, for the sake of durability in case it is the last remaining replica.
This mode may be need to be entered from a myriad of tablet operations: from corruptions in reading a cfile, from disk failures when flushing to disk (while locks may be held), etc.