feat(queue): permanent versus transient consumer errors; undecodable messages to the DLQ; native `dead_letters` table
メンテナーはふだん 1 日以内に返信
まだ誰も着手していません。
評価
- 難易度
- 5/5
- 見積もり時間
- 1週間以上
- 初心者へのやさしさ
- 28/100
調査の方向性
Start with the queue handling paths in crates/thrust-worker/src/lib.rs and crates/thrust-adapters-memory/src/lib.rs, then read the EventConsumer contract in crates/thrust-ports/src/lib.rs and the ActivityConsumer mapping in crates/thrust-core/src/lib.rs. Use the listed contract tests and worker probe as the verification entry points. Done means permanent and undecodable messages are recorded once, transient errors retry with backoff, and the ADR, question, and docs reflect the policy.
索引モデルが issue の本文から書いたものです。
説明
Context
The Worker's queue consumer (crates/thrust-worker/src/lib.rs:166-187) decodes each message as an Envelope. Any handler error is retried after 5 s (:172-177); an undecodable body is logged, acked and dropped (:179-183). The port erases the error class: EventConsumer::handle returns Result<(), String> (crates/thrust-ports/src/lib.rs:179-181) and ActivityConsumer turns every CoreError into a string (crates/thrust-core/src/lib.rs:532-536), although ADR-0070's sketch already names a ConsumerError (line 53). The local wrangler.toml sets max_retries = 3 and no dead_letter_queue (crates/thrust-worker/wrangler.toml:39-43), so an exhausted message is deleted; only wrangler.stage2.toml:45-50 names a DLQ. Natively, MemoryQueue discards the consumer's result (crates/thrust-adapters-memory/src/lib.rs:318-322): no retry, no record.
Problem
Spike 0001 Stage 2 (research/0004 §10.2 S2-9, line 371; F4, lines 419-420): an event with no project_id, a permanent CoreError::Invalid (crates/thrust-core/src/lib.rs:434-436), was delivered 4 times, 5 to 6 s apart, then dead-lettered. The undecodable branch could not be driven, because the inject route only accepts a valid Envelope. Q-0006.5 (QUESTIONS.md:358) is open. ADR-0070 asks for retries with backoff and dead-letter handling (line 22; Confirmation "DLQ after N failures", line 73). The docs describe the gap (docs/src/architecture/infra.md:60-63, docs/src/flows/app.md:114-115). M2 adds the notification consumer, and every new consumer copies whatever policy exists.
Proposal
thrust-ports:enum ConsumerError { Permanent(String), Transient(String) }, andEventConsumer::handlereturnsResult<(), ConsumerError>. Core mapsInvalid,NotFoundand decode failures toPermanent, andDbError::Backend,QueueErrorandBlobErrortoTransient.- Worker:
Permanent: send the body and the error to the DLQ through a producer binding, thenack.Transient:retry_with_optionswith a delay that grows withmsg.attempts(). Undecodable: send the raw text to the DLQ, thenack.wrangler.tomlgets the DLQ producer binding anddead_letter_queueon the consumer, aswrangler.stage2.toml:50has. - Native: a
dead_letterstable (crates/thrust-adapters-postgres/migrations/0002_dead_letters.sql:id,received_at,topic,event_idnullable,bodytext,error,attempts), written through an untenantedSystemScopemethod (#35). The in-process queue records permanent failures at once and retries transient ones with backoff before recording them. - One view on every target: the Worker consumes its own DLQ into the same table, and a test route
GET /internal/dead-letters(test routes on, admin token) reads it, so one contract test runs on all four targets. - Fold the policy into ADR-0070 (round-29 note) and answer Q-0006.5 citing FR-21.
Acceptance criteria
-
permanent_consumer_error_is_dead_lettered_without_retry(contract suite, four targets): anEnvelopewithoutproject_idsent through/internal/inject-eventappears once in/internal/dead-letterswithattempts = 1. Control: mappingInvalidtoTransientmakes it fail; add that mutant toscripts/mutants.sh. - A
thrust-coreunit test shows a backendDbErrorfrom the memory adapter classifiedTransient, and a missingproject_idclassifiedPermanent. -
worker_undecodable_message_is_dead_lettered(crates/thrust-contract-tests/tests/worker_only.rs): aqueue-rawprobe incrates/thrust-worker/src/probes.rssends a non-JSON body, and the row appears with the raw text. Control: restoring ack-and-drop makes it fail. -
grep -n "dropping undecodable" crates/thrust-worker/src/lib.rsandgrep -n "let _ = c.handle" crates/thrust-adapters-memory/src/lib.rsreturn nothing. - ADR-0070 carries the round-29 note; Q-0006.5 is
answeredwith FR-21;infra.mdandflows/app.mddescribe the new behaviour.
References
- Review: .agents/plans/contrail/reviews/2026-10-19-fable-architecture-review.md (FR-21)
crates/thrust-worker/src/lib.rs:166-187;crates/thrust-ports/src/lib.rs:179-181;crates/thrust-core/src/lib.rs:434-436,532-536;crates/thrust-adapters-memory/src/lib.rs:318-322crates/thrust-worker/wrangler.toml:39-43;crates/thrust-worker/wrangler.stage2.toml:45-50research/0004-rust-on-workers-binding-matrix.md§10.2 (S2-9) and §10.4 (F4)- ADR-0070 (lines 22, 53, 73), ADR-0046; Q-0006.5
docs/src/architecture/infra.md:60-63;docs/src/flows/app.md:114-115
Blocked by: #43
Finding: FR-21
- 主要言語
- Rust
- スター
- 0
- フォーク
- 0
- 平均マージ
- 1時間 35分
- マージ済み PR(30日)
- 67
環境構築
- Dockerfile・Docker Compose ファイルなし
- プルリクエストのテンプレートなし
- コントリビューションガイドを読む
はじめの一歩
- 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。
- 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。
- リポジトリをフォークし、ブランチを切って変更します。
- issue 番号を参照したプルリクエストを送ります。
aRustyDev/thrust のほかの issue
-
area:process prio:p1 type:security
難易度 4/5 3〜5日 初心者へのやさしさ 48/100
メンテナーはふだん 1 日以内に返信
-
area:process bug prio:p1
難易度 4/5 3〜5日 初心者へのやさしさ 38/100
メンテナーはふだん 1 日以内に返信
-
area:process prio:p1
難易度 4/5 3〜5日 初心者へのやさしさ 55/100
メンテナーはふだん 1 日以内に返信
-
area:process prio:p1
難易度 4/5 1〜2日 初心者へのやさしさ 55/100
メンテナーはふだん 1 日以内に返信
-
area:process prio:p1
難易度 4/5 3〜5日 初心者へのやさしさ 56/100
メンテナーはふだん 1 日以内に返信
aRustyDev/thrust の issue をすべて見る
似ている issue
-
難易度 2/5 1〜3時間 初心者へのやさしさ 68/100
メンテナーはふだん 1 日以内に返信
-
bug
難易度 2/5 1〜3時間 初心者へのやさしさ 78/100
-
難易度 2/5 1〜3時間 初心者へのやさしさ 72/100
-
難易度 2/5 1〜3時間 初心者へのやさしさ 78/100
CakeRepository/agentshangout#995 ·
-
area:frontend kind:bug priority:p3 ready
難易度 2/5 1〜3時間 初心者へのやさしさ 88/100
メンテナーはふだん 1 日以内に返信