sailfork / t3code-github / .repos/ alchemy-effect/ website/ src/ content/ docs/ planetscale/ guides/ preview-branches.mdx

PlanetScale's branch model maps directly onto alchemy stages: the database is the expensive, long-lived tier; branches are cheap forks. Split your stages the same way and every PR gets an isolated schema and data set without provisioning a cluster per preview:

  • staging-* stages own the long-lived database.
  • pr-* stages reference that database and own only the ephemeral pieces — a branch, a credential, and compute.

The pattern is identical for both engines; only the resource names and the credential model differ. New here? Set up credentials first.

Read the active stage

The create-vs-reference decision keys off the stage being deployed, which the Stack service exposes:

import * as Alchemy from "alchemy";

const { stage } = yield* Alchemy.Stack;

stage is whatever was passed to alchemy deploy --stage <name> — dev_sam, staging-pr-42, or pr-42.

Create or reference the database

PR stages reference the database with PostgresDatabase.ref; every other stage creates it:

import * as Planetscale from "@alchemy.run/planetscale";

const database = stage.startsWith("pr-")
  ? yield* Planetscale.PostgresDatabase.ref("app-db", {
      stage: `staging-${stage}`,
    })
  : yield* Planetscale.PostgresDatabase("app-db", {
      region: { slug: "us-east" },
      clusterSize: "PS_10",
    });

Resource.ref(id, { stage }) reads the deployed resource's attributes from another stage's state at plan time — the lookup is keyed by { stack, stage, id }, so the logical ID "app-db" must match what the owning stage created. database is typed Planetscale.PostgresDatabase on both paths, and downstream resources don't care which one ran. Note the ref targets staging-${stage}, not a global "staging": pr-42 reads from a parallel staging-pr-42 stage, so each PR's data lives in its own database. Prefer one database shared by every PR? Use stage: "staging".

A branch per stage

Below the database line, nothing is conditional. Every stage — including each PR stage — forks its own branch:

const branch = yield* Planetscale.PostgresBranch("app-branch", {
  database,
  migrations: "./migrations",
});

Branches fork from main by default and carry their own migrations: pending .sql files are applied to the branch on every deploy, so each preview migrates itself independently — see Migrations for ordering, hashing, and the tracking table. To fork data as well as schema, seedData: "last_successful_backup" restores the last backup into the new branch.

Mint a per-stage credential

The credential is scoped to the branch, so it's per-stage too:

const role = yield* Planetscale.PostgresRole("app-role", {
  database,
  branch,
  inheritedRoles: ["postgres"],
});

role.origin (direct) and role.pooledOrigin (PSBouncer) are what you hand to the runtime consumer — see Postgres for the role model and the direct-vs-pooled distinction.

The MySQL variant

The MySQL (Vitess) family follows the same two tiers with its own resource trio — MySQLDatabase.ref, a non-production MySQLBranch, and a MySQLPassword in place of the Postgres role:

const database = stage.startsWith("pr-")
  ? yield* Planetscale.MySQLDatabase.ref("app-db", {
      stage: `staging-${stage}`,
    })
  : yield* Planetscale.MySQLDatabase("app-db", {
      region: { slug: "us-east" },
      clusterSize: "PS_10",
    });

const branch = yield* Planetscale.MySQLBranch("app-branch", {
  database,
  isProduction: false,
  migrations: "./migrations",
});

const password = yield* Planetscale.MySQLPassword("app-password", {
  database,
  branch,
  role: "readwriter",
});

Non-production branches run on the PS_DEV size, and the password's role is a Vitess access level ("reader", "writer", "admin", "readwriter") rather than inherited Postgres roles. Migrations behave slightly differently on Vitess — DDL commits implicitly, so prefer one schema change per file — see the engine differences in Migrations.

Deploy the owner first

A ref can only resolve against state that exists, so the owning stage deploys before the PR stage that references it:

alchemy deploy --stage staging-pr-42   # creates the database
alchemy deploy --stage pr-42           # references it, forks a branch

Deploying pr-42 first fails at plan time with InvalidReferenceError — there's nothing under { stack, stage: "staging-pr-42", id: "app-db" } to read yet.

Tear down

alchemy destroy --stage pr-42

Destroy removes the per-stage resources — branch, credential, compute — but leaves the referenced database untouched: the PR stage doesn't own it, so it can't delete it.

Wire it into your runtime

The snippets above end at a credential. Connecting it to compute is runtime-specific:

Complete working projects: cloudflare-planetscale-postgres-drizzle and cloudflare-planetscale-mysql-drizzle.

Where next