feat(db): auto-apply tracked script data migrations in db:migrate (#5497)

* feat(db): auto-apply tracked script data migrations in db:migrate

* fix(db): reset session lock_timeout before script migrations, guard journal insert

* improvement(db): re-verify advisory-lock session before script migrations
This commit is contained in:
Theodore Li
2026-07-07 20:55:42 -04:00
committed by GitHub
parent 41fdcdb5d3
commit 12fb4a9db1
10 changed files with 226 additions and 150 deletions
+1 -4
View File
@@ -11,10 +11,7 @@
* repeated same-spot inserts.
*/
import {
generateKeyBetween,
generateNKeysBetween,
} from '@/lib/fractional-indexing/fractional-indexing'
import { generateKeyBetween, generateNKeysBetween } from '@sim/utils/fractional-indexing'
/**
* Returns a key that sorts strictly between `a` and `b`. Pass `null` for an open
@@ -1,130 +0,0 @@
#!/usr/bin/env bun
/**
* Backfills the `order_key` column on `user_table_rows`.
*
* Row ordering is moving from the contiguous integer `position` to a fractional
* string `order_key` (O(1) insert/delete — no reshift/recompact). This script
* assigns each existing row a key derived from its current `position` order, so
* the new ordering matches today's once the `TABLES_FRACTIONAL_ORDERING` flag is
* flipped on.
*
* Per-table-atomic: each table is keyed inside one transaction holding the same
* per-table advisory lock the app uses for inserts, so a concurrent insert can't
* interleave. Idempotent: tables already fully keyed are skipped; a table with
* any NULL key is fully re-keyed from `position` order (deterministic, so a
* re-run after a partial failure is safe).
*
* Usage:
* DATABASE_URL=... bun run apps/sim/scripts/backfill-table-order-keys.ts
* DATABASE_URL=... bun run apps/sim/scripts/backfill-table-order-keys.ts --dry-run
*/
import { userTableRows } from '@sim/db/schema'
import { getErrorMessage } from '@sim/utils/errors'
import { asc, eq, isNull, sql } from 'drizzle-orm'
import { drizzle } from 'drizzle-orm/postgres-js'
import postgres from 'postgres'
import { nKeysBetween } from '@/lib/table/order-key'
/**
* Rows written per `UPDATE … FROM (VALUES …)`. One statement for a whole large table builds an
* enormous VALUES list that overflows the JS call stack while drizzle assembles it (and would
* exceed Postgres's 65535-bound-param limit at ~32k rows, 2 params/row). 5000 keeps ~10k params
* — well under both ceilings — while minimizing round-trips. Chunks share the one per-table
* transaction, so the table is still keyed atomically.
*/
const WRITE_CHUNK_SIZE = 5000
export async function runBackfill(): Promise<void> {
const dryRun = process.argv.includes('--dry-run')
const connectionString = process.env.DATABASE_URL ?? process.env.POSTGRES_URL
if (!connectionString) {
console.error('Missing DATABASE_URL or POSTGRES_URL')
process.exit(1)
}
const client = postgres(connectionString, {
prepare: false,
idle_timeout: 20,
connect_timeout: 30,
max: 5,
onnotice: () => {},
})
const db = drizzle(client)
const stats = { tables: 0, tablesKeyed: 0, rowsKeyed: 0, failed: 0 }
try {
// Tables that still have at least one un-keyed row.
const pending = await db
.selectDistinct({ tableId: userTableRows.tableId })
.from(userTableRows)
.where(isNull(userTableRows.orderKey))
console.log(
`Backfill starting — ${pending.length} table(s) with NULL order_key${dryRun ? ' [DRY RUN]' : ''}`
)
for (const { tableId } of pending) {
stats.tables += 1
try {
const keyed = await db.transaction(async (trx) => {
// Serialize with concurrent inserts on this table (same lock the app uses).
await trx.execute(
sql`SELECT pg_advisory_xact_lock(hashtextextended(${`user_table_rows_pos:${tableId}`}, 0))`
)
const rows = await trx
.select({ id: userTableRows.id })
.from(userTableRows)
.where(eq(userTableRows.tableId, tableId))
.orderBy(asc(userTableRows.position), asc(userTableRows.id))
if (rows.length === 0) return 0
const keys = nKeysBetween(null, null, rows.length)
if (dryRun) return rows.length
// Chunked UPDATE … FROM (VALUES …) mapping id → key (see WRITE_CHUNK_SIZE).
for (let start = 0; start < rows.length; start += WRITE_CHUNK_SIZE) {
const chunk = rows.slice(start, start + WRITE_CHUNK_SIZE)
const values = sql.join(
chunk.map((r, i) => sql`(${r.id}, ${keys[start + i]})`),
sql`, `
)
await trx.execute(sql`
UPDATE user_table_rows AS t
SET order_key = v.order_key
FROM (VALUES ${values}) AS v(id, order_key)
WHERE t.id = v.id AND t.table_id = ${tableId}
`)
}
return rows.length
})
stats.tablesKeyed += 1
stats.rowsKeyed += keyed
console.log(` ${tableId}: keyed ${keyed} rows`)
} catch (error) {
stats.failed += 1
console.error(` ${tableId}: FAILED — ${getErrorMessage(error)}`)
}
}
console.log('Backfill complete.')
console.log(` tables scanned: ${stats.tables}`)
console.log(` tables keyed: ${stats.tablesKeyed}`)
console.log(` rows keyed: ${stats.rowsKeyed}`)
console.log(` failed: ${stats.failed}`)
if (stats.failed > 0) process.exitCode = 1
} finally {
await client.end({ timeout: 5 }).catch(() => {})
}
}
if ((import.meta as { main?: boolean }).main) {
try {
await runBackfill()
} catch (error) {
console.error('Backfill aborted:', getErrorMessage(error))
process.exitCode = 1
}
}
@@ -20,7 +20,8 @@
* the original backfill also used — minting a fresh, evenly-spaced, distinct run
* with `nKeysBetween`.
*
* Distinct from `backfill-table-order-keys.ts`, which keys tables with NULL keys;
* Distinct from the `0001_backfill_table_order_keys` script migration
* (`packages/db/script-migrations/`), which keys tables with NULL keys;
* this one repairs tables that are fully keyed but bytewise-disordered. Run it
* AFTER migration 0228 so the re-key writes and sorts under `COLLATE "C"`.
*
@@ -41,7 +42,7 @@ import { drizzle } from 'drizzle-orm/postgres-js'
import postgres from 'postgres'
import { nKeysBetween } from '@/lib/table/order-key'
/** See backfill-table-order-keys.ts — keeps each VALUES list well under the param/stack ceilings. */
/** Keeps each VALUES list well under Postgres's 65535-bound-param and JS call-stack ceilings. */
const WRITE_CHUNK_SIZE = 5000
export async function runRepair(): Promise<void> {
@@ -0,0 +1,85 @@
import { getErrorMessage } from '@sim/utils/errors'
import { generateNKeysBetween } from '@sim/utils/fractional-indexing'
import type { ScriptMigration } from './types'
/**
* Rows written per `UPDATE … FROM unnest(...)` statement. The two unnest
* arrays are single bind parameters, so chunking only bounds per-statement
* work and wire-message size, not parameter count. Chunks share the one
* per-table transaction, so the table is still keyed atomically.
*/
const WRITE_CHUNK_SIZE = 5000
/**
* Backfills the `order_key` column on `user_table_rows`.
*
* Row ordering moved from the contiguous integer `position` to a fractional
* string `order_key` (O(1) insert/delete — no reshift/recompact). Each existing
* row gets a key derived from its current `position` order, so the fractional
* ordering matches today's once `TABLES_FRACTIONAL_ORDERING` is on.
*
* Per-table-atomic: each table is keyed inside one transaction holding the same
* per-table advisory lock the app uses for inserts, so a concurrent insert
* can't interleave. Idempotent: tables already fully keyed are skipped; a table
* with any NULL key is fully re-keyed from `position` order (deterministic, so
* a re-run after a partial failure is safe). Per-table failures are isolated —
* remaining tables still run, then the migration throws so it stays pending and
* retries (only still-unkeyed tables) on the next upgrade.
*/
export const backfillTableOrderKeys: ScriptMigration = {
name: '0001_backfill_table_order_keys',
async up(sql) {
const pending = await sql<{ table_id: string }[]>`
SELECT DISTINCT table_id FROM user_table_rows WHERE order_key IS NULL
`
console.log(`order_key backfill — ${pending.length} table(s) with NULL order_key`)
const stats = { tablesKeyed: 0, rowsKeyed: 0, failed: 0 }
for (const { table_id: tableId } of pending) {
try {
const keyed = await sql.begin(async (tx) => {
await tx`
SELECT pg_advisory_xact_lock(hashtextextended(${`user_table_rows_pos:${tableId}`}, 0))
`
const rows = await tx<{ id: string }[]>`
SELECT id FROM user_table_rows
WHERE table_id = ${tableId}
ORDER BY position ASC, id ASC
`
if (rows.length === 0) return 0
const keys = generateNKeysBetween(null, null, rows.length)
for (let start = 0; start < rows.length; start += WRITE_CHUNK_SIZE) {
const chunk = rows.slice(start, start + WRITE_CHUNK_SIZE)
const ids = chunk.map((r) => r.id)
const chunkKeys = keys.slice(start, start + chunk.length)
await tx`
UPDATE user_table_rows AS t
SET order_key = v.order_key
FROM (
SELECT unnest(${ids}::text[]) AS id, unnest(${chunkKeys}::text[]) AS order_key
) AS v
WHERE t.id = v.id AND t.table_id = ${tableId}
`
}
return rows.length
})
stats.tablesKeyed += 1
stats.rowsKeyed += keyed
console.log(` ${tableId}: keyed ${keyed} rows`)
} catch (error) {
stats.failed += 1
console.error(` ${tableId}: FAILED — ${getErrorMessage(error)}`)
}
}
console.log(
`order_key backfill done — tables keyed: ${stats.tablesKeyed}, ` +
`rows keyed: ${stats.rowsKeyed}, failed: ${stats.failed}`
)
if (stats.failed > 0) {
throw new Error(`order_key backfill failed for ${stats.failed} table(s); will retry next run`)
}
},
}
+77
View File
@@ -0,0 +1,77 @@
import type { Sql } from 'postgres'
import { backfillTableOrderKeys } from './0001_backfill_table_order_keys'
import type { ScriptMigration } from './types'
export type { ScriptMigration } from './types'
/**
* Ordered, append-only registry of script migrations. An entry may be deleted
* once a later SQL migration supersedes it (accepting that deployments which
* never ran it skip the backfill) — never renamed or reordered.
*/
export const scriptMigrations: readonly ScriptMigration[] = [backfillTableOrderKeys]
/**
* Applies pending script migrations in registry order, recording each by name.
* Runs inside the migration advisory-lock session, immediately after drizzle's
* SQL migrations succeed — so a script always sees the fully-migrated schema of
* the release that ships it. The tracking table is runner-managed (like
* drizzle's own `__drizzle_migrations`), not part of the app schema.
*
* Fails fast: a missing required env var or a throwing `up` aborts the run
* before the name is recorded, so the migration retries on the next upgrade.
*/
export async function runScriptMigrations(sql: Sql): Promise<void> {
const names = new Set<string>()
for (const migration of scriptMigrations) {
if (names.has(migration.name)) {
throw new Error(`Duplicate script migration name: ${migration.name}`)
}
names.add(migration.name)
}
/**
* The SQL-migration phase leaves `lock_timeout = '5s'` on this session (DDL
* must fail fast rather than queue a table-wide stall). Script migrations
* want the opposite: they block on app-held row/advisory locks (e.g. the
* per-table insert lock in the order_key backfill) and must wait them out —
* exactly like the standalone scripts they replace, which ran on fresh
* connections with the Postgres defaults. `SET` cannot be parameterized,
* hence `unsafe` with constants.
*/
await sql.unsafe('SET statement_timeout = 0')
await sql.unsafe('SET lock_timeout = 0')
await sql`
CREATE TABLE IF NOT EXISTS script_migrations (
name text PRIMARY KEY,
applied_at timestamptz NOT NULL DEFAULT now()
)
`
const appliedRows = await sql<{ name: string }[]>`SELECT name FROM script_migrations`
const applied = new Set(appliedRows.map((row) => row.name))
const pending = scriptMigrations.filter((migration) => !applied.has(migration.name))
if (pending.length === 0) {
console.log('No pending script migrations.')
return
}
for (const migration of pending) {
for (const key of migration.requiredEnv ?? []) {
if (!process.env[key]) {
throw new Error(
`Script migration ${migration.name} requires env var ${key}; set it on the migrations container.`
)
}
}
console.log(`Applying script migration ${migration.name}...`)
const startedAt = Date.now()
await migration.up(sql)
await sql`
INSERT INTO script_migrations (name) VALUES (${migration.name})
ON CONFLICT (name) DO NOTHING
`
console.log(`Script migration ${migration.name} applied in ${Date.now() - startedAt}ms.`)
}
}
+35
View File
@@ -0,0 +1,35 @@
import type { Sql } from 'postgres'
/**
* A run-once TypeScript data migration, applied by the migration runner
* (`scripts/migrate.ts`) after all pending SQL migrations succeed and recorded
* by name in the `script_migrations` table — the code-migration analogue of
* drizzle's `__drizzle_migrations` journal.
*
* Authoring rules:
* - **Idempotent and resumable.** The name is recorded only after `up`
* resolves; a crash mid-run means the whole migration re-runs from the top
* on the next upgrade. Guard with cheap preconditions (`WHERE x IS NULL`,
* `ON CONFLICT DO NOTHING`) so re-runs are near-free.
* - **db-level imports only** (`postgres`, `drizzle-orm`, `@sim/utils`). The
* migrations docker image ships `packages/db` + `utils` + `logger` and
* nothing from `apps/*`.
* - **Runs at whatever schema HEAD the release ships** — scripts are not
* interleaved with SQL migrations. A SQL migration must never drop or
* repurpose a column that a registered script still reads; delete the
* script from the registry in the same PR instead.
* - **Owns its transactions.** The runner deliberately does not wrap `up` in
* one: backfills commit per batch and cannot be rolled back wholesale.
*/
export interface ScriptMigration {
/** Unique stable identifier recorded in `script_migrations`; never rename after release. */
name: string
/** Env vars the migration needs; the runner throws before `up` if any is unset. */
requiredEnv?: readonly string[]
/**
* Applies the migration using the runner's session. The runner resets
* `statement_timeout` and `lock_timeout` to 0 first, so long backfills and
* blocking lock acquisitions wait instead of aborting with `55P03`.
*/
up(sql: Sql): Promise<void>
}
+20 -9
View File
@@ -4,6 +4,7 @@ import { backoffWithJitter } from '@sim/utils/retry'
import { drizzle } from 'drizzle-orm/postgres-js'
import { migrate } from 'drizzle-orm/postgres-js/migrator'
import postgres from 'postgres'
import { runScriptMigrations } from '../script-migrations/index'
/**
* Concurrent-index convention: plain `CREATE INDEX` write-blocks large/hot
@@ -111,6 +112,8 @@ try {
try {
await runMigrationsWithRetry()
console.log('Migrations applied successfully.')
await assertLockSessionHeld()
await runScriptMigrations(client)
await warnOnInvalidIndexes()
} finally {
await releaseMigrationLock()
@@ -179,17 +182,25 @@ async function acquireMigrationLock(): Promise<void> {
* Replays are safe: drizzle rolls the batch back on failure, and post-COMMIT
* CONCURRENTLY statements are idempotent by convention.
*/
/**
* Verify the session still holds the migration advisory lock: a changed
* backend pid means the connection was recycled and the lock silently dropped.
* Only sound on a direct connection — see `hasDirectMigrationUrl`.
*/
async function assertLockSessionHeld(): Promise<void> {
if (!hasDirectMigrationUrl) return
const [{ pid }] = await client`SELECT pg_backend_pid() AS pid`
if (pid !== lockSessionPid) {
throw new Error(
`Database session changed mid-run (backend pid ${lockSessionPid} -> ${pid}); ` +
'the migration advisory lock was lost. Aborting so a fresh runner can retry safely.'
)
}
}
async function runMigrationsWithRetry(): Promise<void> {
for (let attempt = 1; ; attempt++) {
if (hasDirectMigrationUrl) {
const [{ pid }] = await client`SELECT pg_backend_pid() AS pid`
if (pid !== lockSessionPid) {
throw new Error(
`Database session changed mid-run (backend pid ${lockSessionPid} -> ${pid}); ` +
'the migration advisory lock was lost. Aborting so a fresh runner can retry safely.'
)
}
}
await assertLockSessionHeld()
await client.unsafe('SET statement_timeout = 0')
await client.unsafe(`SET lock_timeout = '${DDL_LOCK_TIMEOUT}'`)
try {
+4
View File
@@ -34,6 +34,10 @@
"types": "./src/formatting.ts",
"default": "./src/formatting.ts"
},
"./fractional-indexing": {
"types": "./src/fractional-indexing.ts",
"default": "./src/fractional-indexing.ts"
},
"./object": {
"types": "./src/object.ts",
"default": "./src/object.ts"
@@ -9,11 +9,7 @@
* assertion and rows would display out of order. These tests would catch that.
*/
import { describe, expect, it } from 'vitest'
import {
BASE_62_DIGITS,
generateKeyBetween,
generateNKeysBetween,
} from '@/lib/fractional-indexing/fractional-indexing'
import { BASE_62_DIGITS, generateKeyBetween, generateNKeysBetween } from './fractional-indexing'
/** Bytewise (ASCII / UTF-16 code-unit) comparator — what `COLLATE "C"` and the library use. */
const byteCompare = (a: string, b: string): number => (a < b ? -1 : a > b ? 1 : 0)