feat(jobs): exports and seeding as queue jobs with a `jobs` table (no workflow engine)

未關閉
#56 0 則留言 0 個 reaction 已指派 0 人 在 GitHub 檢視

維護者通常 1 天內回覆

還沒有人認領這個 Issue。

評估

難度
5/5
預估耗時
一週以上
新手友好度
20/100
Issue 類型
功能
描述清晰度
描述清楚
活躍度
活躍
技術堆疊
postgresql, rust

研究方向

Read ADR-0071, especially Option D and the round-29 note, alongside ADR-0046 and ADR-0073. Then inspect the migrations under crates/thrust-adapters-postgres/migrations/ and the worker contract test at crates/thrust-contract-tests/tests/worker_only.rs:203-226. Done means resumable chunked export and workspace seeding work on both worker-cf and native-pg, and the listed ADR, grep, and oracle checks pass.

由索引模型根據 Issue 內容生成。

描述

area:app prio:p1 review:fable-2026-10-19

Context

Long jobs were planned on a WorkflowEngine port: Cloudflare Workflows through the JavaScript shim crates/thrust-worker/entry.mjs:1-26 (a Workflow class that calls back into Rust over the SELF service binding, :20-23; bindings at crates/thrust-worker/wrangler.toml:74-83), and a Postgres-backed step runner in the native binary (ADR-0071 Option B, lines 33, 60, 116). The port exists only in the plan; the code has a probe, worker_bindings_workflow_via_wasm_bindgen_and_js_shim (crates/thrust-contract-tests/tests/worker_only.rs:211-212). ADR-0073 maps the exports (issue_export_task, analytic_plot_export) to Workflows and workspace_seed to "W (or Q)" (lines 72-73, 82); exports are CSV or JSON with links valid 24 h and a size ceiling starting at 100 MB (line 122; Q-0081.2, Q-0081.3, Q-0073.3). #43 removes the shim, the Workflow binding and SELF.

Problem

  • FR-05: the Workflows adapter is the only JavaScript in the Worker, and its exit plan waits on a workers-rs Workflow API with no date (ADR-0071 round 26, line 122). The self-hosted engine is net-new code where "we own correctness (timers, leader election, replay determinism)" (line 89).
  • At this scale a queue plus a state table covers the job catalogue. Workers allow 15 minutes of wall time per queue-consumer or cron invocation; CPU per request defaults to 30 s and can be raised to 5 minutes with limits.cpu_ms (workers/platform/limits). A cron that fires more often than hourly gets 30 s of CPU, so job steps belong in the consumer, not in the cron.
  • Seeding sits on the critical path earlier than M3: POST /api/workspaces/ dispatches workspace_seed, which creates a bot user and a sample project with members, states, labels and work items (research/0003 lines 397-401), and that handler (S14) is in the 48-route slice (line 796).

Proposal

  • A jobs table in crates/thrust-adapters-postgres/migrations/: id, workspace_id, kind (export.work_items_csv, workspace.seed, ...), state (queued, running, succeeded, failed), cursor (resume point), attempts, result_key (a BlobStore key), error, created_by, timestamps, expires_at.
  • Start: a use-case inserts the job row and a job.requested outbox envelope in one transaction (ADR-0046), so the existing relay, queue and sweeper deliver it on both targets.
  • Step: the consumer claims the job, processes one bounded chunk, writes the part to BlobStore, advances cursor and state in one transaction, then enqueues the next step. Steps are idempotent on (job_id, cursor); the sweeper re-enqueues running jobs whose updated_at is stale.
  • This is ADR-0071's Option D (queue-chained steps, state in the database), adopted in the ADR-0071 round-29 note; its main objection, matching Cloudflare Workflows semantics, goes away with the port. ADR-0073 rows move from W to Q.
  • Seeding: a workspace.seed job reproduces Plane's seed. Until it lands, #44 records whether the slice seeds synchronously or deviates.

Acceptance criteria

  • ADR-0071's reference workflow, restated as a job (export a 100k-work-item project to CSV in BlobStore, Confirmation line 77), completes on worker-cf and native-pg, with no step over the Worker's configured CPU limit on the real platform.
  • A contract test fails one export step with a transient error; the job resumes from its cursor, and the output equals that of an uninterrupted run (no missing or duplicated rows). Control: advancing the cursor outside the step's transaction makes it fail.
  • After a workspace create, the seeded projects and members match the oracle's shapes on the routes in the acceptance list, or the deviation is recorded in DEVIATIONS.md.
  • grep -rnE "WorkflowEngine|WORKFLOW|ThrustWorkflow" crates returns nothing.
  • ADR-0073's rows for exports and seeding read Q; Q-0073.3 records the ceiling.

References

  • Review: .agents/plans/contrail/reviews/2026-10-19-fable-architecture-review.md (FR-05)
  • crates/thrust-worker/entry.mjs:1-26; crates/thrust-worker/wrangler.toml:74-83; crates/thrust-contract-tests/tests/worker_only.rs:203-226
  • ADR-0071 (lines 33, 60, 77, 89, 96-99, 116, 122; round-29 note), ADR-0073 (lines 72-73, 82, 122), ADR-0046, ADR-0081
  • research/0003-compat-plane-inventory.md lines 397-401, 796
  • Cloudflare docs: workers/platform/limits (wall time and CPU per invocation type)
  • Q-0073.3, Q-0081.2, Q-0081.3, Q-0071.1

Blocked by: #43, #46

Finding: FR-05

主要語言
Rust
星號
0
分支
0
平均合併
1 小時 35 分鐘
30 天內合併 PR
67

環境準備

  • 沒有 Dockerfile 或 Docker Compose 檔案
  • 沒有 Pull Request 範本
  • 閱讀貢獻指南

從這裡開始

  1. 先讀完整個 Issue,再讀專案的貢獻指南。
  2. 在 Issue 下留言說明你要接手 —— 這能避免兩個人做同樣的事。
  3. Fork 儲存庫,在一個分支上完成修改。
  4. 送出 Pull Request,並在描述裡引用這個 Issue 編號。

aRustyDev/thrust 的其他 Issue

查看 aRustyDev/thrust 的全部 Issue

相似的 Issue

更多 Rust Issue

把新 issue 寄到你的電子郵件信箱

精選適合新手參與的 GitHub issue 摘要。