PlanetScale built the thing everyone else copied: a database where you branch schema like code, review the diff, and deploy it without locking a table. Vitess underneath meant horizontal scale was a real option rather than a someday project. It is a genuinely good product, and most people looking for an alternative are not looking because it is bad.
They're looking for one of four specific reasons, and each points somewhere different. I build PandaStack, which runs managed Postgres on microVMs — relevant to two of the four and irrelevant to the others, and I'll say which.
The four reasons people leave
1. Cost, especially for small projects
The removal of the hobby tier was the moment a lot of side projects and early-stage products started shopping. If your database is small and your budget is smaller, you're looking for a generous free or cheap tier, and that's a different search than 'best MySQL platform'. Candidates: Neon, Supabase, Turso, Xata, or a small managed instance from a general-purpose cloud provider.
2. You'd rather be on Postgres
PlanetScale is MySQL-lineage. If your team's expertise, your ORM ergonomics, or a specific extension — PostGIS, pgvector, TimescaleDB — points at Postgres, that's a legitimate reason to move and it's the most expensive one, because it's a real migration and not a swap. Do not undertake it casually: dialect differences, type mappings, and index behaviour all bite. But if you're going to be on Postgres for the next decade, doing it at a hundred thousand rows beats doing it at a hundred million.
3. Constraints and the Vitess model
Vitess's sharding architecture is why foreign-key support was historically restricted, and while that has improved, teams that want ordinary referential integrity enforced by the database without thinking about it sometimes prefer a plain single-node database. Related: if you never needed horizontal sharding, you were paying architectural complexity for capability you weren't using.
4. You need control of the database itself
Custom extensions, a specific version, unusual configuration, superuser for something. Every multi-tenant managed database restricts this, necessarily — you can't let tenants load arbitrary C into a shared instance. If you're blocked here, no managed MySQL or Postgres service will unblock you; you need a database that is a machine you control.
The options, and who each is for
- Neon — serverless Postgres with instant branching, which is the closest thing to PlanetScale's workflow in the Postgres world. Best fit if branching was the reason you were there and Postgres is where you want to be. Watch cold starts on idle databases.
- Supabase — Postgres plus auth, storage, realtime, and edge functions. Right if you want the surrounding platform; more product than you need if you only wanted a database.
- Amazon RDS / Aurora — the boring answer for MySQL specifically. If you like MySQL and just want it hosted without the Vitess layer, this is a direct, well-understood landing spot with excellent operational tooling and more setup.
- Turso — SQLite at the edge with embedded replicas. A completely different architecture: extremely fast reads near the user, and a data model you should evaluate on its own terms rather than as a MySQL replacement.
- Xata and similar — developer-experience-first Postgres with branching and a nice API layer. Worth a look if DX was what you liked about PlanetScale.
- PandaStack (mine) — each database is Postgres 16 on its own Firecracker microVM with a durable volume. Because it's a machine rather than a shared service, the extension surface is the machine's, and clone-to-a-point-in-time produces a genuinely independent database from the archive rather than a shared-storage view. The trade: creating one takes tens of seconds rather than being instant, because you are getting a VM, and it's Postgres — so this is only relevant if you were already moving off MySQL.
- Self-managed MySQL or Postgres on a VM — cheapest and most controlled, and you now own backups, failover, patching, and the 3am page.
Keep the branching habit
The workflow worth preserving is: every pull request gets a database that is a real copy of production's shape and a plausible amount of its data, and destructive tests against it cannot touch anything real. Different platforms implement it as copy-on-write branches, restore-from-snapshot, or a scripted restore of a recent backup into a fresh instance.
Three properties to compare, in order of how often they matter: whether the clone is genuinely independent, how long it takes, and whether it duplicates storage cost.
# A per-PR database from a point in time, independent of production
pandastack db clone prod-db \
--label pr-1042 \
--target-time 2026-08-19T08:00:00Z
# Source is untouched; the clone gets its own id and connection URLMigrating off, without a bad weekend
- Decide first whether this is a swap (MySQL to MySQL) or a port (MySQL to Postgres). They are different projects with different risk profiles, and conflating them is how migrations overrun.
- Inventory the things that are not tables: stored procedures, triggers, generated columns, JSON functions, collation and charset behaviour, and anything your ORM emits that is dialect-specific.
- Rehearse the full dump and restore with real data volume and time it. That number is your maintenance window, and it is usually longer than the estimate.
- For a port, run both databases in parallel with dual writes or a replication tool, and diff a sample of reads before cutting over. Do not trust a single successful restore.
- Keep the old database running and readable for at least a week after cutover. Not a snapshot — actually running.
Short version
Cost on a small project: Neon or Supabase free tiers. Staying on MySQL, just elsewhere: RDS or Aurora. Wanting Postgres and keeping the branching workflow: Neon. Wanting Postgres with full control of the instance and independent clones: a per-database VM platform like PandaStack. Blocked on extensions or configuration: a machine, not a service. And if none of the four reasons is actually yours, staying put is a perfectly good answer.