Files
meshcore-analyzer/cmd/ingestor/async_migration.go
T
n30nex 7f2c999266 fix(ingestor): record cancelled background migrations separately (#2161)
Refs #1739. Depends on the small local-validation fix in #2160; merge
that first. This branch contains its prerequisite commit plus one
migration-status commit.

An interrupted background migration currently records `failed`, making
an operator shutdown look like a migration fault. Errors matching
`context.Canceled`, including wrapped errors, now record `cancelled` and
retain the error and completion time without logging `FAILED`. Ordinary
errors, deadline expiry and panics remain failures even when the context
is cancelled; a successful callback still records `done`.

Cancelled work follows the existing retry path on the next run.
Regression tests cover classification and clearing stale
completion/error fields before retrying, then successful completion. The
migration runbook documents the distinction.

This covers the existing ingestor bookkeeping portion of #1739. The
issue's server status mapper and banner depend on closed, unmerged #1735
and are absent from current master; this PR does not add that API/UI and
leaves #1739 open for the remaining scope. No new configuration,
customizer setting or schema migration is required.

Validation: direct and wrapped cancellation failed the new regression
before the change; focused tests and the race check pass afterward.
Native full server/ingestor suites and all 194 standalone frontend
suites pass with #2160. The browser smoke suite passed 148 checks, with
three fixture-dependent skips and two pre-existing disabled cases.
Windows-only filesystem capability skips are documented in #2160; Linux
CI retains the security assertions.

Performance impact: one error-chain check when a background migration
finishes, using the existing single status update. No packet-ingestion
loop, batching, polling or retention behavior changes, and no speedup is
claimed.
2026-10-10 21:20:57 +02:00

155 lines
6.1 KiB
Go

// Async migration helper — runs schema/backfill work that may take minutes on
// large prod tables WITHOUT blocking ingestor startup.
//
// MIGRATION ANNOTATION CONVENTION (read this before touching migrations):
//
// Sync schema/data migrations (CREATE INDEX, ALTER TABLE, UPDATE ... WHERE)
// that run inline during OpenStore() block the ingestor from accepting
// packets until they finish. On an empty dev DB they return in milliseconds;
// at prod scale (1.9M+ observations, 80K+ adverts) they can pin the boot
// for minutes and trigger restart loops. This regression class has bitten us
// repeatedly (#791 resolved_path backfill, #1483 obs_observer_ts_idx_v1).
//
// ANY new CREATE INDEX / ALTER TABLE / data-rewrite migration MUST EITHER:
// 1. Run via Store.RunAsyncMigration(...) below (preferred for backfills
// and any work that may touch >1K rows). The migration is recorded as
// `pending_async` immediately, returns to the caller (boot proceeds),
// and completes in a goroutine. Status flips to `done`, `cancelled`,
// or `failed` when fn returns; interruptions retain the error message.
// 2. Carry the preflight annotation comment immediately above the
// migration block, e.g.
// // PREFLIGHT: async=true reason="<one-line justification>"
// Use this for migrations that are genuinely cheap at any scale
// (e.g. ALTER TABLE ADD COLUMN, CREATE INDEX on a known-bounded
// table). The annotation is grepped by
// ~/.openclaw/skills/pr-preflight/scripts/check-async-migrations.sh
// — its absence on a touched migration block is a hard-fail gate.
//
// See MIGRATIONS.md in the repo root for the full policy and examples.
package main
import (
"context"
"database/sql"
"errors"
"fmt"
"log"
)
// ensureAsyncMigrationsTable creates the bookkeeping table used by
// RunAsyncMigration / AsyncMigrationStatus. Idempotent.
func ensureAsyncMigrationsTable(db *sql.DB) error {
_, err := db.Exec(`
CREATE TABLE IF NOT EXISTS _async_migrations (
name TEXT PRIMARY KEY,
status TEXT NOT NULL, -- pending_async | done | cancelled | failed
started_at TEXT NOT NULL DEFAULT (datetime('now')),
ended_at TEXT,
error TEXT
)
`)
return err
}
// RunAsyncMigration registers `name` as a pending async migration and
// schedules `fn` to run in a background goroutine. It returns to the caller
// immediately so the ingestor can keep booting.
//
// Contract (pinned by async_migration_test.go):
// - status is `pending_async` IMMEDIATELY after this returns.
// - fn runs in a goroutine; on success status becomes `done`. Errors matching
// context.Canceled become `cancelled`; other errors and panics become
// `failed`. Both interrupted states record the error and completion time.
// - Idempotent: if a row with the same name already exists in `done`
// state, fn is NOT re-run. If in `failed`, `cancelled`, or `pending_async`,
// fn IS re-scheduled (a previous run may have crashed mid-flight).
// - The caller's WaitGroup tracks the goroutine so tests/shutdown can
// wait via Store.WaitForAsyncMigrations().
func (s *Store) RunAsyncMigration(ctx context.Context, name string, fn func(context.Context, *sql.DB) error) error {
if err := ensureAsyncMigrationsTable(s.db); err != nil {
return fmt.Errorf("ensure _async_migrations: %w", err)
}
var existing string
row := s.db.QueryRow(`SELECT status FROM _async_migrations WHERE name = ?`, name)
switch err := row.Scan(&existing); err {
case nil:
if existing == "done" {
return nil // already complete, nothing to do
}
// pending_async, cancelled, or failed → reset and retry.
if _, err := s.db.Exec(`
UPDATE _async_migrations
SET status = 'pending_async', started_at = datetime('now'), ended_at = NULL, error = NULL
WHERE name = ?`, name); err != nil {
return fmt.Errorf("reset async migration %q: %w", name, err)
}
case sql.ErrNoRows:
if _, err := s.db.Exec(`
INSERT INTO _async_migrations (name, status) VALUES (?, 'pending_async')`,
name); err != nil {
return fmt.Errorf("register async migration %q: %w", name, err)
}
default:
return fmt.Errorf("lookup async migration %q: %w", name, err)
}
s.backfillWg.Add(1)
go func() {
defer s.backfillWg.Done()
var runErr error
defer func() {
if r := recover(); r != nil {
runErr = fmt.Errorf("panic: %v", r)
log.Printf("[async-migration] %q panic recovered: %v", name, r)
}
if runErr != nil {
status, label := "failed", "FAILED"
if errors.Is(runErr, context.Canceled) {
status, label = "cancelled", "cancelled"
}
if _, err := s.db.Exec(`
UPDATE _async_migrations
SET status = ?, ended_at = datetime('now'), error = ?
WHERE name = ?`, status, runErr.Error(), name); err != nil {
log.Printf("[async-migration] failed to record %s for %q: %v", status, name, err)
}
log.Printf("[async-migration] %q %s: %v", name, label, runErr)
return
}
if _, err := s.db.Exec(`
UPDATE _async_migrations
SET status = 'done', ended_at = datetime('now'), error = NULL
WHERE name = ?`, name); err != nil {
log.Printf("[async-migration] failed to mark %q done: %v", name, err)
return
}
log.Printf("[async-migration] %q done", name)
}()
log.Printf("[async-migration] %q starting (boot continues)", name)
runErr = fn(ctx, s.db)
}()
return nil
}
// AsyncMigrationStatus returns the current status of an async migration
// (one of "pending_async", "done", "cancelled", "failed") or sql.ErrNoRows if no such
// migration has been registered.
func (s *Store) AsyncMigrationStatus(name string) (string, error) {
if err := ensureAsyncMigrationsTable(s.db); err != nil {
return "", err
}
var status string
err := s.db.QueryRow(`SELECT status FROM _async_migrations WHERE name = ?`, name).Scan(&status)
return status, err
}
// WaitForAsyncMigrations blocks until all currently-scheduled async migrations
// finish. Intended for tests + graceful shutdown; production boot path does NOT
// call this (that's the whole point).
func (s *Store) WaitForAsyncMigrations() {
s.backfillWg.Wait()
}