Hi Zhijie Hou and others, On Sat, 21 Nov 2026 at 22:39, Zhijie Hou <[email protected]> wrote:
> Hi, > > While reviewing the dynamic logical decoding status control code > (introduced in commit 67c20979ce7), I noticed a race condition when > creating a logical replication slot on a standby while logical decoding > is being dynamically enabled. > > During recovery, ReplicationSlotReserveWal() sets the new logical slot's > restart_lsn to GetXLogReplayRecPtr(). However, logical decoding can > become enabled before GetXLogReplayRecPtr() advances past the > corresponding XLOG_LOGICAL_DECODING_STATUS_CHANGE record in two cases: > > 1. During WAL replay, xlog_redo() calls EnableLogicalDecoding() and sets > logical_decoding_enabled to true before ApplyWalRecord() updates > lastReplayedEndRecPtr to the end of the status-change record. > > 2. At the end of recovery, UpdateLogicalDecodingStatusEndOfRecovery() > sets logical_decoding_enabled to true (even before writing the > status-change record) while RecoveryInProgress() is still true and > lastReplayedEndRecPtr remains at EndOfLog, never advancing past that > record. > > In either window, a concurrent logical slot creation on the standby > observes logical decoding as enabled, reserves a restart_lsn preceding > the status-change record, and fails in xlog_decode() with: > > ERROR: unexpected logical decoding status change 1 > > Attached is a patch to fix this by Holding LogicalDecodingControlLock in > UpdateLogicalDecodingStatusEndOfRecovery() until the status-change > record is written and recording its end LSN in last_replayed_enable_lsn > when enabling logical decoding (alongside EnableLogicalDecoding() during > replay), and clamping the reserved restart_lsn in > ReplicationSlotReserveWal() to at least GetLogicalDecodingEnabledLSN() > during recovery. > > Note that the test in the patch covers only case 2 (reusing the existing > end-of-recovery injection point) to avoid adding a new injection point > in xlog_redo() just for case 1. > > And I kept the field name last_replayed_enable_lsn to keep diff minimal, > though we could rename it (e.g., to last_enable_lsn) if others prefer > since it now also records an end-of-recovery LSN. > > -- > Best Regards, > Zhijie Hou > I'm ready to investigate it by reviewing your patch file, and I'll send my review here and if time permits me then I'm sending my patch here too.
