ns-ui / truncation-word-count
The migration finished at 02:40 with zero dropped writes, but the interesting part is what did not happen. The dual-write shim stayed in place for eleven days, and in that window we diffed every row pair nightly: 14 mismatches total, all of them timestamps rounded differently by the old ORM. Nothing user-visible. The cutover itself was a single DNS flip behind the connection pooler, so rollback stayed one command away the whole time. What we'd do differently: start the nightly diffs a week earlier, and put the shim's latency cost (about 6ms p50) on a dashboard from day one instead of discovering it during review. The old cluster stays warm until Friday, then we snapshot and tear it down.
Connection pooling moves to the new PgBouncer pair next sprint. The rollout mirrors this one: dual configuration for a week, nightly comparisons of connection counts and wait times, and a one-command rollback path via the load balancer. The only open question is whether the analytics batch jobs keep their dedicated pool or share the general one; benchmarks land Thursday and will settle it.
Read replicas come next quarter. No schema changes planned before then.
the fold states exactly what it hides — a measured word count, not an ellipsis. text that fits gets no control.