firestore: BulkWriter.close() never resolves when a write retried after close() belongs to a batch cut for a same-document write

オープン
#9,444 コメント 0 件 リアクション 0 件 担当者 1 名 GitHub で見る

メンテナーはふだん 1 日以内に返信

@cherylEnkidu がすでに取り組んでいます。

2026年9月28日 から。

評価

難易度
4/5
見積もり時間
3〜5日
初心者へのやさしさ
72/100
issue の種類
バグ
明瞭さ
明確に書かれている
活発さ
活発
技術スタック
node.js, typescript
領域
backend, databases

調査の方向性

Start in build/src/bulk-writer.js by tracing _sendFn, _scheduleCurrentBatch, and _sendBatch through the provided reproduction, especially the retry after a batch is cut with flush=false. Verify the behavior with the emulator-based script and confirm that retries scheduled after close() or flush() are sent and that close() resolves.

索引モデルが issue の本文から書いたものです。

説明

api: firestore
Environment details
  • Package: @google-cloud/firestore 9.2.0 (latest); build/src/bulk-writer.js
  • Node.js: v24.21.0
  • OS: macOS 27.2 arm64
  • Backend: Firestore emulator v1.21.0 and v1.22.0 (both reproduce it deterministically)
Summary

BulkWriter.close() (and flush()) never resolves when a write gets a retryable error and it sits in a batch that _sendFn scheduled with flush=false, i.e. the batch that was cut because the next write targets a document already in it.

_sendFn does this when two writes target the same document:

if (this._bulkCommitBatch.has(op.ref)) {
    this._scheduleCurrentBatch();          // flush = false
}

Once that batch commits, _sendBatch calls if (flush) this._scheduleCurrentBatch(flush), which is skipped here. The failing op's onError calls sendFn again, which puts the retry into the new current _bulkCommitBatch. Nothing ever schedules that batch:

  • close() already ran,
  • the op is not flushed (only buffered ops get markFlushed()),
  • and its batch was not a flush batch.

_lastOp therefore never settles, and close() hangs forever. The docs for close() promise the opposite: "Any retries scheduled as part of an onWriteError() handler will be run before the close() promise resolves."

With the default error handler the same thing happens for any retryable code. We hit it through ABORTED, where two writes to one document in one BulkWriter raced each other. The repro below uses a custom onWriteError so it is deterministic.

Steps to reproduce
// npm i @google-cloud/firestore
// FIRESTORE_EMULATOR_HOST=127.0.0.1:8080 node bulkwriter-lost-retry.js
const { Firestore } = require('@google-cloud/firestore');
const db = new Firestore({ projectId: `bw-lost-retry-${Date.now()}` });
const t0 = Date.now();
const log = (...a) => console.log(`+${Date.now() - t0}ms`, ...a);

(async () => {
  const bw = db.bulkWriter();
  // Retry every error up to 3 attempts, NOT_FOUND included.
  bw.onWriteError(err => {
    log(`write error code=${err.code} attempt=${err.failedAttempts}`);
    return err.failedAttempts < 3;
  });
  const missing = db.doc('c/missing');
  const other = db.doc('c/other');
  bw.update(missing, { a: 1 }).catch(e => log('update rejected', e.code));
  bw.set(other, { a: 1 });
  // Second write to `other`: _sendFn schedules the current batch (update +
  // first set) with flush=false and starts a new batch for this write.
  bw.set(other, { a: 2 });
  await bw.close();
  log('close() resolved');
  process.exit(0);
})();
setTimeout(() => {
  log('BUG: close() still pending after 15s; the retried update was never sent');
  process.exit(1);
}, 15000);

Actual:

+64ms write error code=5 attempt=1
+15001ms BUG: close() still pending after 15s; the retried update was never sent

Control: remove the second bw.set(other, …), so no batch gets cut. The retries then run and close() resolves:

+64ms write error code=5 attempt=1
+1200ms write error code=5 attempt=2
+2927ms write error code=5 attempt=3
+2927ms update rejected 5
+2927ms close() resolved

With setLogFunction the last line is BulkWriter.errorFn ... shouldRetry: true, and no further request is sent.

Expected

A retry should be sent even when it belongs to a batch scheduled with flush=false, and close() should resolve once it has run, as documented.

A possible fix: have _sendBatch also reschedule the current batch after a commit when close()/flush() has already been called, for example by tracking a "flush requested" flag on the writer instead of relying on each batch's flush argument. Alternatively, _sendFn could schedule immediately when it re-enqueues a retry after a flush was requested.

Real-world impact

A service that writes one document twice in one BulkWriter (set-merge plus a follow-up update) hung forever in close(). Firestore emulator v1.22.0 started returning ABORTED Transaction lock timeout. for one of the two concurrent batchWrite calls (https://github.com/firebase/firebase-tools/issues/11160), and the retry of that ABORTED write was lost as described above. Real Firestore can also return ABORTED on contention, so production code is exposed too.

主要言語
TypeScript
スター
3.2k
フォーク
732
平均マージ
2日 11時間
マージ済み PR(30日)
99

環境構築

はじめの一歩

  1. issue を最後まで読み、次にプロジェクトのコントリビューションガイドを読みます。
  2. 着手することを issue にコメントします — 二人が同じ作業をするのを防げます。
  3. リポジトリをフォークし、ブランチを切って変更します。
  4. issue 番号を参照したプルリクエストを送ります。

googleapis/google-cloud-node のほかの issue

googleapis/google-cloud-node の issue をすべて見る

似ている issue

TypeScript の issue をもっと見る

新しい issue をメールで受け取る

初心者向けの GitHub issue を短くまとめたダイジェスト。