Notes for the problems that show up after launch: bad plans, awkward migrations, index debt, vacuum pressure, replica lag, and the small decisions that make PostgreSQL easier to operate.
EXPLAIN tells you which plan MySQL chose; optimizer_trace tells you which plans it rejected and why. Per-session capture, reading considered_execution_plans, cost numbers, truncation traps, and a worked wrong-index case.
Checksum failures in the error log, a server that crashes 40 seconds after every start, and the six levels of innodb_force_recovery. What each level really skips, the dump-and-rebuild runbook, and why replicas beat recovery.
Multi-source replication lets one replica fan in from many primaries with FOR CHANNEL syntax — but there is no conflict detection and no last-write-wins. Setup, per-channel monitoring, GTID bookkeeping, and when a pipeline beats fan-in.
An InnoDB FULLTEXT index is a family of hidden auxiliary tables with its own parser, stopwords, and deferred deletes. Where the edges are, which knobs demand a rebuild, and when to move search out of MySQL.
Updating one key in an 18KB JSON document should not rewrite 18KB into the binlog. MySQL 8.0 can patch documents in place and log just the diff — when the change qualifies, and when it silently does not.
binlog_expire_logs_seconds makes old files eligible for purge — it does not promise disk headroom, and it does not protect an offline replica. What pins purging, how to purge safely, and how to size retention from write rate.
FTWRL waits for every in-flight statement, then makes everything else wait for it. MySQL 8.0's LOCK INSTANCE FOR BACKUP and performance_schema.log_status give backup tools a lighter path — and you a shorter outage.
A failover that should have taken two minutes took forty-seven because the log had 30,000 virtual log files. How autogrowth wrote that layout, what log_reuse_wait_desc is telling you, and how to rebuild the layout exactly once.
One RLS policy with a memberships subquery turned our indexed tenant lookups into per-row nested loops overnight — p95 went from 9 ms to 1.4 seconds. Here's how we found it and the pattern that fixed it.
Columnstore cut a report from four minutes to six seconds until trickle inserts rotted the rowgroups. Rowgroups, segment elimination, the deltastore lifecycle, batch mode on rowstore, and when a plain btree still wins.
Same procedure, two seconds for one tenant, thirty-second timeouts for another. How to prove parameter sniffing with plan variance, and the fix toolkit ranked from better indexes to 2022's Parameter Sensitive Plan optimization.
synchronous_commit is the rare setting that trades durability for latency openly. Used per transaction, it is the fastest safe win on a write-heavy pipeline.