Airflow／Dagster／Prefect

Agentに定期実行・データ依存・再試行・asset lineageを丸投げせず、非LLMのscheduler/data workflowを別の決定論的制御面として後続更新へ渡す。

non-LLM component / implementation / governance / benchmark / audit / cost / operability / provider mapping。

Airflow、Dagster、Prefectの公式Docsは、schedule/task/DAG、asset/lineage、Python flow/state/retry/deploymentを各々何と定義するか。Agent orchestrationと混同しない役割分担、各製品の運用限界・managed/OSS境界をどこまで確立できるか。

• source_class=official technical documentation; exact_title=Dags; author/org=Apache Airflow; canonical_url=https://airflow.apache.org/docs/apache-airflow/stable/core-concepts/dags.html publication/update/access_date=update date not stated / docs version 3.3.1 / accessed 2026-09-14 00:00 JST; version/release/tag/commit/API revision=Airflow 3.3.1; supporting_locator=lines 4-30（Dagがworkflow、schedule/tasks/dependencies/callbacks/retries/timeoutsを包む）、lines 34-76（Python declaration）、lines 141-174（manual/API/schedule trigger、Dag Run/data interval、parallel runs）、lines 741-756（dag.test simulated execution）; evidence_label=VERIFIED_SPEC+PRODUCT_VOLATILE; relation=support+limit。
• source_class=official technical documentation; exact_title=Overview; author/org=Dagster Labs / Dagster; canonical_url=https://docs.dagster.io/ publication/update/access_date=© 2026 / accessed 2026-09-14 00:00 JST; version/release/tag/commit/API revision=current Dagster Docs; supporting_locator=lines 109-126（data orchestrator、integrated lineage/observability、declarative programming、testability、asset definition/dependency例、Dagster Plus managed offering）; evidence_label=VERIFIED_SPEC+PRODUCT_VOLATILE; relation=support+context。
• source_class=official product/documentation page; exact_title=Modern Data Orchestrator Platform; author/org=Dagster Labs / Dagster; canonical_url=https://dagster.io/ publication/update/access_date=update date not stated / accessed 2026-09-14 00:00 JST; version/release/tag/commit/API revision=current public page; supporting_locator=lines 160-189（branch deployments、guardrails、hybrid deployment、observability、AI-ready developer experience）および lines 205-220（asset-centric vs Airflow task-centric、hybrid、Dagster core open source、Dagster+ enterprise features）; evidence_label=SOURCE_REPORT+PRODUCT_VOLATILE; relation=context+limit。
• source_class=official technical documentation; exact_title=Introduction; author/org=Prefect; canonical_url=https://docs.prefect.io/v3/get-started publication/update/access_date=update date not stated / accessed 2026-09-14 00:00 JST; version/release/tag/commit/API revision=Prefect 3 documentation; supporting_locator=lines 21-39（Python orchestration engine、state tracking、failure handling、monitoring、portable execution、events/human intervention、dynamic runtime、CI/CD）および lines 50-75（Prefect Cloud/self-hosted server、read-only MCP diagnostics、RBAC scope）; evidence_label=VERIFIED_SPEC+PRODUCT_VOLATILE; relation=support+limit。
• source_class=official technical documentation; exact_title=Flows; author/org=Prefect; canonical_url=https://docs.prefect.io/v3/concepts/flows publication/update/access_date=update date not stated / accessed 2026-09-14 00:00 JST; version/release/tag/commit/API revision=Prefect 3 documentation; supporting_locator=lines 21-50（Python flow、tracked state、retries、timeouts、remote deployment）および lines 55-90（flow run interruption、zombie flow run、schedule/event/API/deployment、logs/events for monitoring/auditing）; evidence_label=VERIFIED_SPEC+PRODUCT_VOLATILE; relation=support+limit。

• AIRFLOW ESTABLISHED: Airflow 3.3.1 DocsはDagをschedule、tasks、dependencies、callbacks、retry、timeoutを含むworkflow modelと定義し、manual/API/schedule trigger、Dag Run/data interval、parallel runs、Python declaration、simulated dag.testを記載する。task-centricな定期・依存実行の一次証拠になる。
• DAGSTER ESTABLISHED: Dagster Docsはdata orchestratorとしてlineage、observability、declarative programming、testability、asset dependencyを中心に説明する。Dagster公式ページはasset-centricとAirflow task-centricの差、Dagster core open source、Dagster+のmanaged/enterprise機能、hybrid deploymentを主張するため、Docsの機能記載と製品ページの主張を分ける。
• PREFECT ESTABLISHED: Prefect 3 DocsはPython functionをflowにし、state tracking、failure handling、retries、timeouts、deployments、arbitrary infrastructure、events、human intervention、monitoring/auditingを説明する。Prefect Docsはhost消失等でterminal stateに到達しない“zombie flow run”も定義し、運用限界を明記する。
• INFERENCE / PROPOSAL: Agentはboundedな判断・例外分類・tool選択に限定し、schedule、data dependency、retry、asset lineage、state recoveryをAirflow/Dagster/Prefect等へ寄せる構成は、責任境界と再現性を明確にする設計案である。3製品の公式Docsがこの統合構成を保証しているわけではない。

AirflowのDAG/task retry、Dagsterのasset/job/resource、Prefectのflow/task state/retryはいずれも外部side effectのexactly-onceやLLM品質を保証しない。Airflow scheduler/DAG parsing・backfill、Dagster asset materialization/lineage、Prefect zombie run・host/network・dynamic task、provider/plugin/version差が失敗包絡になる。公式ページの“reliable/trustworthy/production-grade”は製品説明であり、独立SLO/benchmarkではない。複数製品を一つの機能として合算しない。

Airflow 3.3.1 vs other releases、executor/worker/metadata DB/provider package、Dagster OSS vs Dagster+、asset checks/branch/hybrid/RBAC、Prefect 3 vs Cloud/self-hosted server、work pool/worker/infrastructure、MCP read-only scope、deployment/region/data residency、secrets/credentials、schedule/queue/concurrency、retry/timeout/retention。

Airflow/Dagster/Prefectの公式Docs/公式ページのみ確認。いずれもinstallation、DAG/asset/flow creation、scheduler/worker execution、MCP接続、external write、負荷・障害・replay試験は未実施。

Airflowはdag.test、Dagsterはtestability/asset checks導線、Prefectはmonitoring/stateを記載するが、共通dataset、ground truth、grader、threshold、cross-product performance、agent success rate、cost per accepted outcomeはない。製品のmonitoring/evaluation機能と測定結果を混同しない。UNKNOWN。

Security=executor/worker isolation、secrets、provider/plugin、RBAC、MCP read-only、network egress。Privacy=data assets、workflow inputs、logs、metadata DB、model/tool payload。Audit=DAG/asset/flow version、schedule/trigger、task/run/state、retry、operator/approval、lineage/log、admin change。Cost=compute/worker、managed control plane、metadata/storage、provider/API、model/tool、retry/backfill。Operability=dependency pinning、backfill/partition、idempotency、retry/dead-letter、zombie detection、asset freshness、scheduler/worker scaling、rollback/replay、log completeness。

「Airflow 3.3.1はschedule・task・dependency・retry・timeoutを持つDag実行、Dagsterはasset/lineage/observability中心のdata orchestration、Prefect 3はPython flowのstate・retry・timeout・deploymentを公式Docsで説明している。これらはAgentの確率的判断を支える非LLMの実行・データ制御面として分離できるが、exactly-once・LLM品質・全deploymentのSLOは別途検証が必要である。」

EA-156を三製品の比較棚として保持し、Airflow=task/DAG scheduling、Dagster=asset/lineage、Prefect=Python flow/stateというclaim relationを混同しない。将来本文ではAgent→bounded decision→deterministic workflow/data planeのhandoff、human approval、idempotency、retry/rollback、audit/costをTEST_PROFILEとして追加する。

current release/API parity、cross-product integration with agents/MCP、exact retry/idempotency semantics、managed/OSS feature gates、RBAC/data residency、SLO/benchmark、scheduler/worker failure, cost, public incidents。

Airflow・Dagster・Prefectを一つの製品・一つの性能値・一つのOSS/managed契約として扱わない。公式のdata orchestration説明をagent autonomyやexactly-once保証へ外挿せず、Agent連携はINFERENCE/PROPOSALとして明示する。 UNIT_CLOSE_STATUS: INDIVIDUAL_BOUNDED_CLOSE / EVIDENCE_RECORDED / UNKNOWN_PRESERVED ===== RUN CLOSE / HANDOFF ===== INDIVIDUAL_PACKET_COUNT: 6/6（EA-151..EA-156）

Zapier, Make, Workato, Pipedream, Temporal, Apache Airflow, Dagster, and Prefect official/author-maintained records with canonical URLs and locators are recorded; Temporal issue #8901 is retained as a bounded counterevidence report.

EA-151..EA-156 each has an independent topic, findings, limits/counterevidence/failure envelope, availability gates, implementation state, benchmark/eval boundary, safe wording, handoff, open unknowns, and stop condition. EA-156 keeps Airflow, Dagster, and Prefect subclaims separate.

Product documentation establishes documented feature statements at the recorded access/version boundary, not universal plan/region/tenant availability, production safety, quality, ROI, or SLO. Vendor metrics and customer/product claims are not promoted to independent benchmark evidence. LIVE_SIDE_EFFECTS: 0.