Checkpoints

Persist pipeline state mid-execution so long pipelines survive wall-clock / runner crashes.

v0.4.0concept

Checkpoints (v0.3+)

Persist pipeline state mid-execution so long pipelines survive wall-clock / runner crashes.

Use-case priority

Use case	Status
30s Cloudflare Workers wall-clock	MUST in v0.3 (otherwise long pipelines don’t run server-side at all)
Server / CI runner crash mid-execute	MUST (same mechanism)
User pauses run, resumes later	COULD (CLI ergonomic, can ship later)
Cross-execution continuity (multi-day)	COULD (separate scope)

Storage model

- id: long-phase
  checkpoint:
    on: stage_complete
    storage:
      type: r2
      path: checkpoints/{pipeline_id}/{exec_id}.json
    retention: 24h

Stage serialization shape: { stages: [progress], pending_stages: [...], resolved_inputs: {...], timestamp }.

On resume: deserialize → skip completed stages → continue from pending.

What gets checkpointed

NOT stage outputs (recomputable from inputs). Only:

Which stages completed (id + start timestamp)
Resolved template variables (avoid re-running expensive interpolation)
Current DAG position (UI progress display)
Operator signature so far (resume preserves audit trail)

Behavioral oracle

packages/pipeline-runtime/tests/checkpoints.test.ts — R10.2 checkpoints: serialize completedStageIds + pending DAG after each stage, restore on resume: true, skip already-completed stages
packages/server/tests/checkpoint-r10-4.test.ts — server-side R2 storage backend (checkpoint write/read/delete lifecycle)
packages/pipeline-runtime/tests/integration-r10-x.test.ts — end-to-end: long pipeline interrupted, resumed via checkpoint storage