Workflow Engines (Durable Execution)

The "queue" question (see Queues) is "how do I run a job later?" The "workflow" question is "how do I run a 17-step process that takes 3 days, survives restarts, retries each step independently, and is observable end-to-end?" — that's durable execution.

TS / Node-friendly

• ★ Inngest — TS SDK, declarative steps (step.run, step.sleep, step.waitForEvent), great free tier; runs your code on your infra. The most-recommended pick for new TS projects.
• ★ Trigger.dev — open source + hosted; same niche as Inngest with its own ergonomic SDK; generous free tier; self-hostable.
• Restate — open source durable execution; SDK in TS, Java, Kotlin, Go, Rust, Python; deploy as a stateful service.
• DBOS (database-backed durable execution) — TS + Python SDKs, durable on Postgres; relatively new.
• Cloudflare Workflows — durable, multi-step workflows on Workers; native to Cloudflare.

Polyglot heavyweight

• Temporal — the gold standard for serious enterprise workflows; SDKs in TS, Go, Java, Python, .NET. Powerful, complex; self-host or Temporal Cloud.
• Cadence — Uber's predecessor to Temporal; less common today.
• Conductor (Netflix) — JSON workflow definitions; OSS.
• Camunda / Activiti — BPMN-based; for big-enterprise process modeling.
• Argo Workflows — Kubernetes-native.

Lighter / event-driven

• Mastra — TS framework; has workflows for AI agents; see AI/LLM.
• Zeebe — workflow engine often paired with Camunda.
• Prefect / Dagster — Python-flavored workflow tools (data pipeline focused).

Patterns to know

• Determinism — workflow code must be deterministic; non-determinism (random, time, network) goes through SDK helpers (step.run, ctx.now).
• Idempotent steps — required so retries don't double-charge / double-send.
• Side-effect isolation — wrap calls to external systems in steps; the engine remembers the result.
• Wait-for-event — pause until "user clicked link" or "stripe webhook arrived"; survives weeks.
• Compensation / sagas — if step 5 of 7 fails, run rollback steps 4..1.

When to use a workflow engine vs. a queue

• Queue: "send this email later." Queue + retry is enough.
• Workflow: "send email, wait 3 days for click, if no click send reminder, after another 7 days mark dormant." Reach for durable execution.
• State machine in your DB + cron: works up to ~5 steps; falls over fast.

Pick this if…

• Default new TS project, hosted free tier: Inngest or Trigger.dev.
• Self-host, OSS, polyglot: Restate or Temporal.
• Enterprise scale, mature ecosystem: Temporal.
• All-Cloudflare: Cloudflare Workflows.
• Postgres-backed, simplest possible: DBOS.