Lab: commit and visibility across two CNs¶
Use two real CN endpoints in one isolated cluster and confirm their service/CN IDs; two clients on one CN are insufficient. Pin v4.2.5 build/configuration/clocks and follow cluster deployment. Single-node CI does not execute topology-dependent blocks.
Source CN¶
DROP DATABASE IF EXISTS kernel_cn_518;
CREATE DATABASE kernel_cn_518;
USE kernel_cn_518;
CREATE TABLE events(id INT PRIMARY KEY, label VARCHAR(20));
BEGIN;
INSERT INTO events VALUES(1,'committed');
COMMIT;
SELECT id,label FROM events ORDER BY id;
Destination CN¶
After acknowledged source commit, open a new session on the other CN:
USE kernel_cn_518;
SELECT id,label FROM events ORDER BY id;
The eventual result must be (1,'committed'). Retain source acknowledgement, destination read time/snapshot, subscription/applied watermarks and logs. Blocking/waiting is an observation; no fixed millisecond SLA is promised. Preserve destination errors instead of only polling.
Reconnection and boundaries¶
Reconnect or restart the destination lab CN, wait for subscription/readiness and reread without missing/duplicate rows. Record source session, another source-CN session and destination session separately. A reused long transaction requires isolation/snapshot analysis first.
For batches/cancellation/network failures check uncommitted rows remain absent and reconcile unknown commits through business state. Record actual catch-up rather than a heartbeat-derived SLA. Delete the lab database through an available CN. See visibility.