Ask HN: Has distributed SQL come of age?
At work we're exploring ways of scaling our database layer, ideally achieving horizontal scalability in both read and write throughput. A few years ago horizontal write scalability could only mean you either explicitly shard or adopt one of the distributed NoSQL key-value stores.
I've sat on the sidelines though as "distributed SQL" became a new category- first with Cloud Spanner, then with CockroachDb, TiDB, MariaDb XPand, etc. They promise near-linear horizontally scalability, atomic transactions across instances (required for us), real online schema changes, and a compliant SQL interface.
It seems almost too good to be true if you're already tied to relational databases, which we are. I never hear about real companies with massive scaling needs using them though. Are they mature enough to be trusted? Does anyone have real-world data about performance or reliability? The literature is sparse. It's definitely an interesting category right now - I've been following Citus and Vitess with interest for a while since they let you use PostgreSQL and MySQL respectfully with magic sharding sauce. I haven't actually tried either yet - like you I'm waiting for more success stories (plus I don't currently have any problems that warrant that approach). Slack adopted Vitess a few years ago. Vitess and Citus seem to be half-way between manual sharding and distributed SQL. They really discourage cross-shard updates, for instance, whereas they're a really natural part of distributed SQL:
> The authors of Vitess recommend that you design your VSchema so that cross-shard updates (and 2PC) are not required.
https://vitess.io/docs/reference/features/two-phase-commit/ Maybe though it's just a fundamental coordination problem one can't work around. I'm very OK with that limitation - when I'm designing a sharded system I'm happy to take shards into account and arrange my entities such that I can apply updates to them grouped on a single shard. The value I want is help directing those write queries to the correct shards, and some assistant with cross-shard read queries. Apache ignite does this