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.
Both Aurora flavors share the same storage layer and carry very different databases on top. Here is how I compare them: feature deltas, the extension ecosystem, performance tests you should run yourself, and migration gravity.
MariaDB's InnoDB is a maintained fork with its own defaults: the change buffer is gone in 11.0, flushing defaults to O_DIRECT, the redo log was rewritten, and half of every MySQL tuning guide quietly stops applying.
The sync job double-counted rows for three weeks because someone said 'change tracking' and someone else heard 'CDC'. What each feature actually captures, what each one costs, and a four-question test for picking the right one.
Every shared MariaDB box eventually meets the query that should never have been allowed to run for six hours. max_statement_time is the built-in kill switch — here is how to scope it, what it actually kills, and why it will not save you from a retry loop.
When the MySQL optimizer guesses wrong by a factor of a thousand, plans collapse. Column histograms fix skewed, non-indexed estimates, if you know when they are ignored.
Two sessions inserting different rows can still deadlock under REPEATABLE READ. Gap locks, next-key locks, insert intention locks, and how to read the deadlock report.
MySQL's default is one thread per connection, and it degrades gracefully until it falls off a cliff. When the thread pool plugin actually helps, when it just rearranges the queue, and how to size max_connections without guessing.
One SET GLOBAL gives MariaDB per-account, per-client, per-table, and per-index counters — enough to answer 'who is hammering this table' in a single query. Here is what userstat collects, what it costs, and where the unused-index caveats hide.
The dashboard said SYNCHRONIZED while the readable secondary served twenty-minute-old data. Send queue and redo queue are different numbers; here is what each one means, what synchronous commit really guarantees, and how to measure RPO.
mysqldump, XtraBackup, and the clone plugin map surprisingly well onto pg_dump, pg_basebackup, and pgBackRest. The concepts transfer; the tools, PITR mechanics, and failure modes do not.
Concurrent FOR SHARE locks and foreign key checks consume MultiXact IDs, and multixact has its own wraparound emergency separate from XID wraparound. How to monitor mxid_age(datminmxid), SLRU pressure, and vacuum's freezing before it stops your database.
Dead tuples climbed for days, autovacuum ran constantly, and pg_stat_activity showed nothing. The culprit was one row in pg_prepared_xacts, eleven days old, left by a Java service that crashed mid-commit.