Thoughts, tutorials, and insights from the trenches.
A `jobs` table with a status column and a worker that polls it every few seconds is a perfectly reasonable way to start. The signs that it's time for a real message queue show up gradually, and they're not really about scale.
Redis usually gets introduced as 'just a cache.' A year later it is holding session data, rate-limit counters, and critical queues. The instance never changed, but its job did. Its configuration absolutely must change with it.
Galera's pitch of synchronous multi-master replication where every node has the exact same data is genuinely true. It just comes with a handful of operational realities that aren't obvious from the pitch, and that I learned the slow, painful way.
A pg_dump cron job feels exactly like a real backup, right up until restore day. Here is how I set up pgBackRest so restores are a routine event, not a guessing game.