Serializing an AbortSignal into a step adds ~0.5s fixed latency per step invocation on world-local

[AndrewBarba]

Summary

Passing an AbortSignal (or AbortController) into a step's arguments adds ~0.5 s of fixed latency to every invocation of that step on world-local — a ~7× slowdown for a trivial step (84 ms → ~600 ms). The cost is entirely on the step side: creating the controller in the workflow is free, and the same workflow with the signal omitted from the step input runs at full speed.

On a world that has accumulated stream chunks, the same signal-bearing step degrades far past the fixed cost — see the companion issue about the world-local streamer's polling (readdir-per-tick) behavior, filed separately.

This makes the documented cancellation pattern (durable AbortController, signal threaded into long-running steps) unusable in practice on world-local: we had to pull it out of a feature branch after a 100-turn session e2e went from ~0.9 s/turn to a 240 s replay timeout.

Versions observed

• @workflow/core@5.0.0-beta.26
• @workflow/world-local@5.0.0-beta.22 (via createLocalWorld; in-process handler, WORKFLOW_TURBO=0)
• Node v24, macOS (also reproduced on Linux CI)

Numbers

20 sequential trivial steps (append one line to a file, return):

WITHOUT signal: total=4025ms  per-step deltas ≈ 84ms
WITH signal:    total=13026ms per-step deltas ≈ 600ms   (~+515ms per step, flat)

Same shape inside our product's turn loop: per-turn latency 385 ms → 910 ms the moment the turn's one step carries a signal; bisected across five variants (hook only: free; controller created but signal not passed: free; signal passed: +~520 ms).

Repro

import { appendFile } from "node:fs/promises";

export async function timedStep(input: {
  logPath: string;
  index: number;
  abortSignal?: AbortSignal;
}) {
  "use step";
  await appendFile(input.logPath, `step ${input.index} ${Date.now()}\n`);
  return input.index;
}

export async function signalCostWorkflow(input: {
  count: number;
  withSignal: boolean;
  logPath: string;
}) {
  "use workflow";
  const controller = new AbortController();
  for (let index = 0; index < input.count; index++) {
    await timedStep({
      abortSignal: input.withSignal ? controller.signal : undefined,
      index,
      logPath: input.logPath,
    });
  }
  return "done";
}

Run once with withSignal: false and once with true; diff the timestamps.

Where the time appears to go

The serialized signal carries { streamName, hookToken }. On every step revival, the step host opens a tail reader on the abort stream for the step's entire duration (the real-time abort delivery path — gx(...) in the serialization chunk: new ox(runId, streamName).getReader() raced against a per-step teardown controller, pushed into the step context's ops). The ~0.5 s fixed cost tracks that reader's open/tail/teardown lifecycle against a stream that has no chunks and, in the common case, never will.

Two adjacent observations:

• The controller's abrt_… system hook is registered per controller and never disposed, so long-lived sessions creating one controller per unit of work accumulate live system hooks.
• The reader polls the chunk store every 100 ms with a full directory listing — that is the scaling half of this problem, filed separately with numbers.

Expected

A signal in step arguments should cost roughly what a hook read costs — ideally near-zero when never aborted: e.g. subscribe lazily/in-process first, or don't tie a polling stream tail's lifecycle to every step invocation.

Actual

Every invocation of a signal-bearing step pays ~0.5 s on world-local even in a fresh world, growing without bound as the world accumulates stream chunks.

Related

• Companion scaling issue (world-local streamer readdir polling) — filed alongside this one.
• A hook_received while a step attempt is in flight re-dispatches the step; both attempts run concurrently in-process #2780, resumeHook racing hook disposal is journaled after hook_disposed; the owner's replay then diverges into CorruptedEventLogError #2781, Recreating a hook token after dispose() in the same run self-conflicts: suspension processing validates the new hook_created before the pending hook_disposed persists #2777, Hook token claim is not released in order with hook_disposed: a fast next-run claimant gets HookConflictError against an already-disposed hook #2778 — same downstream investigation (turn cancellation).