Caleb Rackliffe created CASSANDRA-21515:
-------------------------------------------
Summary: Avoid Rebuilding Per-SSTable SAI Components Unless All
Column Indexes are Being Rebuilt
Key: CASSANDRA-21515
URL: https://issues.apache.org/jira/browse/CASSANDRA-21515
Project: Apache Cassandra
Issue Type: Bug
Reporter: Caleb Rackliffe
Assignee: Caleb Rackliffe
When a table has multiple SAI indexes, they share per-SSTable index components
(the primary key map, token/partition-key blocks, group metadata). Running
nodetool rebuild_index naming only one of those indexes triggers a "full
rebuild" path that unlinks and rewrites those shared components on every
SSTable, but only the rebuilt index is notified: its view is refreshed against
the newly-written state, while the bystander indexes are never told. Those
bystanders keep serving queries through their existing {{SSTableIndex}}
objects, which hold ref-counted shared copies of the old {{SSTableContext}}, so
they silently continue resolving row IDs through the now-unlinked (but still
mmap'd) old primary key map, and only release that stale reference when the
sstable is compacted, the bystander is itself dropped or rebuilt, or the node
restarts. Nothing in the code path fails, marks the bystanders non-queryable,
or checks that the rewritten mapping matches the old one; correctness holds
only because row-ID ordering happens to be deterministic given the same input
sstable, and the cost is a hidden disk/FD leak plus under-reported per-SSTable
metrics until the stale references are finally released.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
