changefeed lag continues to increase after ticdc rolling restart

[Lily2025]

What did you do?

1、run sysbench
2、add index is executed three times
3、ticdc rolling restart

What did you expect to see?

after ticdc rolling restart，changefeed lag can be restored to 5s

What did you see instead?

changefeed lag continues to increase after ticdc rolling restart

Versions of the cluster

./cdc version
Release Version: v8.5.4-nextgen.202510.5-74-g7a206e9
Git Commit Hash: 7a206e9
Git Branch: HEAD
UTC Build Time: 2026-02-14 03:48:33
Go Version: go version go1.25.5 linux/amd64
Failpoint Build: false
Kernel Type: Next Generation
2026-02-22T02:47:46.517+0800

[3AceShowHand]

The coordinator switched once at the time when the checkpoint lag increase.

This is not related to the resolved-ts, but the event service send resolved-ts lag

[lidezhu]

Timeline observed:
(1) A dispatcher REGISTER arrives and eventStore.RegisterDispatcher() succeeds.
(2) Immediately after, schemaStore.RegisterTable() fails with [CDC:ErrNewStore] new store failed: context canceled (often during PD/TSO client rebuild), so eventBroker.addDispatcher() returns early and the dispatcher is not stored in the broker’s dispatcher map.
(3) The downstream EventCollector receives the Ready event from EventService and sends a RESET request (epoch increments).
(4) EventService handles RESET, but eventBroker.resetDispatcher() cannot find the dispatcher and logs “reset a non-exist dispatcher, ignore it”, so the dispatcher stays at epoch=0/seq=0 (no handshake).
(5) The collector has already marked itself ready and will ignore subsequent Ready signals, so it never retries RESET.
(6) EventService keeps printing “dispatcher not reset” for this dispatcher, and because this table’s progress never advances, the changefeed resolved ts can stall