-
Type:
Task
-
Resolution: Unresolved
-
Priority:
Major - P3
-
None
-
Affects Version/s: None
-
Component/s: None
-
None
-
Storage Execution
-
None
-
None
-
None
-
None
-
None
-
None
-
None
Flip both ends of the rid pipeline from the per-collection recordIdsReplicated catalog option to groupId membership in one commit: a primary emits rids iff it holds a groupId, and an applier honors them iff it holds a groupId.
Details
- Emission: the op observer already sets the rid (oplogEntry.setRecordId in op_observer_impl.cpp). The change is gating it on the membership plus the in-scope predicate instead of the per-collection catalog option. Holding a groupId implies flag-on plus latest FCV, since miniting only happens under both.
- Honoring: relax tassert for rid presence iff collection option to a membership-plus-in-scope check.
- In-group: inserts at the replicated rid, updates and deletes located by rid (existing computeEffectiveRid machinery).
- Out-of-group: inserts generate local rids (today's non-rrid behavior), updates and deletes fall back to _id-based lookup.
- Honoring is active in all application modes: steady state, initial sync application phase, and startup oplog recovery.
- Emission and honoring must flip atomically. Emitting membership-based while honoring catalog-based (or vice versa) creates states where an in-group applier expects rids a groupless primary never wrote, or a primary emits rids an out-of-group applier ignores, both surfacing only as data divergence.
- No behavior change for converged sets: an all-in-group replica set behaves exactly like today's flag-on behavior, so the existing featureFlagRecordIdsReplicated-tagged tests stay green.
Acceptance criteria
- Entries carry rids for in-scope collections iff the primary hols a groupId and none when groupless.
- In-group nodes apply inserts at the replicated rid and locate updates and deletes by rid.
- Out-of-group nodes ignore rid entirely.
- Unit tests for in-group and out-of-group application of insert, update, and delete.
- jstest with a mixed-membership replica set validating both paths.
- Existing featureFlagRecordIdsReplicated-tagged tests pass.
- depends on
-
SERVER-138106 Mint and announce the groupId at election (new-primary noop)
-
- Open
-
-
SERVER-138110 Clear groupId on mismatched new-primary noop (C1), with batcher fence
-
- Open
-
-
SERVER-138221 Found the groupId at setFCV upgrade completion
-
- Open
-
-
SERVER-138223 Clear the groupId on FCV downgrade (C2)
-
- Open
-
-
SERVER-138224 Clear the groupId on flag-off startup (C3)
-
- Open
-
-
SERVER-138231 Logical initial sync joins the RecordIds group
-
- Open
-
-
SERVER-139451 Restrict the fixture dbHash rid opt-in to disagg until groupId membership lands
-
- In Code Review
-
-
SERVER-138123 Add the shared in-scope predicate for set-wide Replicated RecordIds
-
- Closed
-
- is depended on by
-
SERVER-117561 Drop the overrideRecordIdsReplicatedFalse failpoint from dbcheck_logs_minkey_maxkey.js
-
- Open
-
-
SERVER-122753 Remove recordIdsReplicated from listCollections output
-
- Open
-
-
SERVER-138127 Re-enable rid-sensitive dbHash checks in test fixtures with groupId-membership-aware opt-in
-
- Open
-
-
SERVER-138242 Retire the per-collection recordIdsReplicated option
-
- Open
-
-
SERVER-139504 Drop the overrideRecordIdsReplicatedFalse failpoint from index_downgrade_fcv.js
-
- Open
-