feat(tasks): workers started with versions only claim their own tasks

OpenWorkflow 0.10.1 claims any run regardless of workflow name and fails
unregistered ones with 5s-5m backoff; filter the claim SQL by name@version.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
This commit is contained in:
rootandClaude Opus 5.5 committed 2026-09-27 22:12:04 +08:00
1 parent cda9255083
commit a9cee69bd2
3 files changed
+135 -2

No files matched your search

+52 -1
View File
@@ -59,6 +59,54 @@ export async function migrateEngineSchema(databaseUrl: string, schema = ENGINE_S
await backend.stop();
}
/**
* 包一层只认领指定 (任务名, 版本) 的 backend:与 BackendPostgres.claimWorkflowRun 同一条 SQL(截止过期处理 +
* SKIP LOCKED 认领),candidate 多一个 workflow_name/version 过滤。其余方法经原型链委托给原 backend。
*/
function claimFilteredBackend(backend: BackendPostgres, defs: readonly TaskDefinition[]): BackendPostgres {
// biome-ignore lint/suspicious/noExplicitAny: 访问 BackendPostgres 的私有字段(pg / namespaceId / workflowRunsTable)
const b = backend as any;
const keys = [...new Set(defs.map((d) => `${d.name}@${engineVersion(d.version)}`))];
const filtered = Object.create(backend) as BackendPostgres;
Object.defineProperty(filtered, "claimWorkflowRun", {
value: async (params: { workerId: string; leaseDurationMs: number }) => {
const table = b.workflowRunsTable();
const [claimed] = await b.pg`
WITH expired AS (
UPDATE ${table}
SET "status" = 'failed', "error" = ${b.pg.json({ message: "Workflow run deadline exceeded" })},
"worker_id" = NULL, "available_at" = NULL, "finished_at" = NOW(), "updated_at" = NOW()
WHERE "namespace_id" = ${b.namespaceId}
AND "status" IN ('pending', 'running', 'sleeping')
AND "deadline_at" IS NOT NULL
AND "deadline_at" <= NOW()
RETURNING "id"
),
candidate AS (
SELECT "id"
FROM ${table}
WHERE "namespace_id" = ${b.namespaceId}
AND "status" IN ('pending', 'running', 'sleeping')
AND "available_at" <= NOW()
AND ("deadline_at" IS NULL OR "deadline_at" > NOW())
AND ("workflow_name" || '@' || COALESCE("version", '')) = ANY(${keys}::text[])
ORDER BY CASE WHEN "status" = 'pending' THEN 0 ELSE 1 END, "available_at", "created_at"
LIMIT 1
FOR UPDATE SKIP LOCKED
)
UPDATE ${table} AS wr
SET "status" = 'running', "attempts" = "attempts" + 1, "worker_id" = ${params.workerId},
"available_at" = NOW() + ${params.leaseDurationMs} * INTERVAL '1 millisecond',
"started_at" = COALESCE("started_at", NOW()), "updated_at" = NOW()
FROM candidate
WHERE wr."id" = candidate."id" AND wr."namespace_id" = ${b.namespaceId}
RETURNING wr.*;`;
return claimed ?? null;
},
});
return filtered;
}
export class OpenWorkflowEngine implements TaskEngine {
private readonly ow: OpenWorkflow;
private readonly workers: Worker[] = [];
@@ -144,7 +192,10 @@ export class OpenWorkflowEngine implements TaskEngine {
async startWorker(opts: StartWorkerOptions = {}) {
const defs = this.opts.registry.select(opts.versions);
const workflows = defs.map((d) => this.toWorkflow(d)) as Workflow<unknown, unknown, unknown>[];
const worker = new Worker({ backend: this.backend, workflows, concurrency: opts.concurrency ?? 4 });
// 指定了 versions 的 worker 只认领这些任务:OpenWorkflow 0.10.1 认领不看任务名,认领到未注册的 run 会
// 以"未注册"失败并指数退避(5s→5m),分组 worker(io / agent)会互相拖慢对方的任务。
const backend = opts.versions?.length ? claimFilteredBackend(this.backend, defs) : this.backend;
const worker = new Worker({ backend, workflows, concurrency: opts.concurrency ?? 4 });
await worker.start();
this.workers.push(worker);
this.log.info("worker started", { tasks: defs.map((d) => `${d.name}@${d.version}`) });