world-postgres: startup recovery enqueues the entire active set on every boot — bound it with a recovery batch size

未关闭
#5,237 5 条评论 0 个 reaction 已指派 0 人 在 GitHub 查看

维护者通常 1 天内回复

还没有人认领这个 Issue。

评估

难度
4/5
预计耗时
3-5 天
新手友好度
62/100
Issue 类型
功能
描述清晰度
描述清楚
活跃度
活跃
技术栈
node.js, postgresql, typescript

调研方向

Read reenqueueActiveRuns() and the createRunsList pagination path in world-postgres, then check how recovery configuration is loaded and tested. Add a configurable per-boot recovery bound while preserving the current unbounded default. Done means startup enqueues no more than the configured number of active runs and leaves the rest eligible for recovery on a later boot.

由索引模型根据 Issue 内容生成。

描述

Summary

reenqueueActiveRuns() walks every non-terminal run on world.start(), enqueuing one Graphile job each, with no bound on how many are enqueued in a single boot. When the active set is large, a restart turns into a burst of thousands of jobs, which drives the process into memory pressure and, in our case, an out-of-memory crash loop the restart cannot escape.

Recovery already pages its read (createRunsList defaults to limit ?? 20) and awaits enqueue() one run at a time — so the loop is inherently a batch loop today. The only thing missing is a way to stop after N runs instead of draining every page.

Versions: @workflow/[email protected], @workflow/[email protected], [email protected], [email protected], Node 24.

Relationship to existing issues

Not a duplicate of the three open reports on this function; all three ask for something different:

  • #3119 — which runs get selected (parked runs are replayed; duplicate jobs accumulate).
  • #3968 — a producer-only World that does no startup recovery at all.
  • #5129 — an enqueue-only process (queueConcurrency: 0) that also does no startup recovery.

All three are binary: recover everything, or recover nothing. This issue is about recovering all of it, a batch at a time. A predicate fix (#3119) does not help here — the runs we hit were genuinely running with real progress records, so any correct runnable predicate still selects them. A recovery-off switch (#3968 / #5129) is also wrong for us: we want recovery, we just cannot afford 1,500 jobs in one boot.

What happens

Single-pod deployment, 4 GiB limit, FAT environment. A load test left 1,512 runs in status='running' plus a job backlog. Each restart then:

  1. world.start() → reenqueueActiveRuns() → lists all 1,512, enqueues 1,512 jobs.
  2. The process works through them; RSS climbs to ~2.8 GiB within a few minutes.
  3. At the old 1 GiB limit this became an OOM crash loop: process dies → pm2 restarts → recovery re-enqueues everything again → dies again. SIGKILL repeated ~58 times in 30 minutes, CPU pinned at ~200% while nothing user-visible was happening.

Raising the limit to 4 GiB stopped the crash, but each boot still pushes RSS from ~600 MiB to ~2.8 GiB before settling. On a smaller instance, or with a larger active set, the same shape returns.

The amplification is in the interaction with per-job retry budgets: every recovered run gets a fresh max_attempts (MAX_GRAPHILE_JOB_ATTEMPTS = 49 + 24 in queue.ts), so a run that fails deterministically never exhausts its budget and never leaves the active set. Each restart therefore re-creates the same burst rather than draining it. (Details of that in #3119 — mentioned only because it is what makes the active set stay large.)

Reproduction

  1. Insert N runs with status='running' in workflow.workflow_runs (any N well above a few hundred reproduces the burst).
  2. Start the World once and watch graphile_worker._private_jobs: it gains N jobs.
  3. Watch RSS across the first few minutes of the boot.

We observed N = 1,512 → ~1,790 jobs and RSS 600 MiB → 2.8 GiB.

Proposal

Bound the number of runs recovered per boot. Two shapes, either would work:

a) A batch size option — recover at most recoveryBatchSize runs, continue from the same cursor on the next boot:

await reenqueueActiveRuns(runs, queue.queue, 'world-postgres', namespace, {
  limit: config.recoveryBatchSize,   // default: unbounded (current behavior)
  // the remaining runs stay non-terminal and are picked up next start
});

Because the read is already paginated and enqueue() is already awaited per run, this is a small change inside the existing loop — stop when the counter reaches the limit, leave the cursor implicit (the next boot re-reads from the beginning).

b) Page-level pacing — process at most one page per boot, or pause between pages:

for (let page = 0; hasMore; page += 1) {
  if (page >= maxPagesPerBoot) break;
  ...
}

Either way the important property is: a bounded boot may leave runs un-recovered, and that is fine — they are still pending/running, so the next start() sees them again. Nothing is lost; the burst is spread across boots instead of landing on one.

A bounded recovery also composes with the existing per-job retry accounting: runs that do not get recovered in a given boot keep accumulating nothing, which (together with a durable enqueue identity, see #3119) would let the retry ceiling do its job.

Environment variable suggestion

Following the package's existing convention:

WORKFLOW_POSTGRES_RECOVERY_BATCH_SIZE   (default: 0 = unbounded, current behavior)

Defaulting to unbounded keeps the change backwards compatible.

Current workaround

We sweep workflow_runs before eve start and flip runs that have been replayed too many times to failed, so they stop being selected. That caps the eventual damage but does not bound a single boot: while the counter is below the threshold, every boot still recovers the whole active set at once. We also tried deferring a subset without persisting any marker — that does nothing, because world.start() selects purely on status, so the "deferred" runs are recovered anyway. A recovery limit is the only way to bound the burst from outside the package, and it needs to live inside reenqueueActiveRuns.

Happy to test a branch against our environment if that helps.

主要语言
TypeScript
星标
2.5k
派生
401
平均合并
1 天 21 小时
30 天内合并 PR
201

环境准备

  • 没有 Dockerfile 或 Docker Compose 文件
  • 有 Pull Request 模板
  • 没有贡献指南

从这里开始

  1. 先读完整个 Issue,再读项目的贡献指南。
  2. 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
  3. Fork 仓库,在一个分支上完成修改。
  4. 提交 Pull Request,并在描述里引用这个 Issue 编号。

vercel/workflow 的其他 Issue

查看 vercel/workflow 的全部 Issue

相似的 Issue

更多 TypeScript Issue

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。