Source: Inngest docs: how functions are executed, step.waitForEvent and usage limits

Inngest's step-memoization execution model (checkpointing vs standard orchestration), event-based waits for approvals, and its published limits on steps, state and run length.

Contents

Summary

An Inngest function is triggered by an event and split into named steps. Each successful step’s result is saved; on a retry or later request the SDK re-invokes the handler, returns saved results for completed steps (memoization keyed by a hash of the step ID), and runs only the next or failed step. Waits (sleep, waitForEvent) pause the run without consuming compute, which makes event-driven approvals cheap.

Key claims

  • TypeScript SDK v4 defaults to “checkpointing” (steps run back to back in one handler invocation, results saved as they go); it falls back to standard orchestration (one HTTP request per step) on retries and parallel steps. Python always uses standard orchestration.
  • Default retries: four after the initial attempt, counted per step (each step.run() has its own counter, so a failing step does not re-run completed ones).
  • Code outside step.run() can run again on every replay; put DB/API calls inside steps.
  • A step can succeed externally (e.g. a refund) before its result is saved and then be retried, so side-effecting steps should be idempotent.
  • step.waitForEvent(id, {event, timeout, match | if}) resolves to the matching event or null on timeout; match (dot-path equality) and if (CEL) are mutually exclusive. A losing waitForEvent in a parallel race is not cancelled and keeps the run “Running” until its timeout.
  • Limits (accessed 2026-11-14): 1,000 steps per function; 32 MiB run state; 4 MiB per step output; sleeps up to 1 year; max run length 30 days (Free), 90 (Pro), 366 (Business); event payload 256 KiB (Free) / 3 MiB (paid); step timeout up to 2 hours subject to the host.

Feeds

Durable agent execution, Human-in-the-loop

Suggest changes →