Fan-Out / Fan-In — Senior¶

At senior level, focus on this question:

When one of N fanned-out workers fails, what should happen to the other N-1, and to the overall operation?

Prerequisite: middle.md.

Fail-fast: cancel everything the instant one fails¶

with concurrent.futures.ThreadPoolExecutor(max_workers=4) as executor:
    futures = [executor.submit(risky_task, c) for c in chunks]
    try:
        results = [f.result() for f in futures]  # first exception propagates
    except Exception:
        for f in futures:
            f.cancel()  # attempt to cancel remaining work
        raise

Fail-fast is appropriate when the overall operation's semantics require all sub-tasks to succeed for the result to be meaningful at all (e.g. "process every row in this batch, all-or-nothing") — continuing the other workers wastes resources on work that will be discarded anyway.

Partial success: collect what succeeded, report what failed¶

results = {"succeeded": [], "failed": []}
for future in concurrent.futures.as_completed(futures):
    try:
        results["succeeded"].append(future.result())
    except Exception as e:
        results["failed"].append(e)

Partial success is appropriate when sub-tasks are genuinely independent and useful on their own (e.g. "resize these 10 images" — 9 succeeding is still valuable even if 1 fails) — matching this page's choice to junior.md's independence requirement for fan-out itself.

🎯 Senior takeaway: the fail-fast-vs-partial-success decision should be made deliberately based on whether the fanned-out sub-tasks' results are only meaningful together (fail-fast) or independently valuable (partial success) — this is the exact same all-or-nothing-vs-independent- value distinction that determines saga design (per the Compensating Transaction reliability pattern) applied at the in-process, single-operation scale instead of a distributed transaction's scale.

Test yourself¶

1. Give an example where fail-fast is clearly the right choice for a fanned-out operation, and one where partial success is clearly right.
2. Why does continuing to run other workers after one has already failed waste resources specifically in the fail-fast scenario?
3. Design the failure-handling strategy for a fan-out operation resizing 100 images, where a small percentage of images might have corrupted data.

Continue to professional.md to design fan-out with bounded concurrency at production scale.