Skip to content

CANDIDATE.md

Latest commit

 

History

History
60 lines (45 loc) · 2.88 KB

CANDIDATE.md

File metadata and controls

60 lines (45 loc) · 2.88 KB
Raw

Technical Screen: rlfan — Fan-out Command Runner

Time: ~50 minutes of coding. You won't finish everything — that's expected. Language: Python 3.9+. AI assistants / autocomplete: <ALLOWED / NOT ALLOWED — interviewer will confirm>. SDK docs: https://github.com/runloopai/api-client-python (see api.md and README.md)

The problem

At Runloop, customers constantly need to run the same command across a fleet of devboxes — think "run this test suite on 20 fresh sandboxes" or "check the installed CUDA version everywhere." You're going to build the small CLI tool that does this.

Build rlfan:

python fanout.py --count 5 --command "python --version" [--timeout 60] [--mock]

Core requirements (in priority order)

  1. Provision --count devboxes and wait until each is running.
  2. Execute --command on all of them concurrently — total wall-clock time should be roughly the time of the slowest box, not the sum.
  3. Collect results and print a summary at the end: per-devbox ID, exit status, duration, and the first line of stdout/stderr. The process exit code should be 0 only if every command succeeded.
  4. Tear down cleanly. No devbox should be left running when the tool exits — including when provisioning fails partway through, a command errors, or the run times out. (Leaked devboxes cost customers money. We check.)

Failure semantics

  • One box failing must not prevent results from the others, and must not prevent any box from being shut down.
  • --timeout bounds the whole run. On timeout, report what you have and clean up.

Stretch goals (only if time remains — interviewer will pick)

  • --retries N: retry a failed provision or command on a fresh devbox.
  • --max-concurrent N: cap in-flight provisioning (the API rate-limits).
  • Stream results as they complete instead of only at the end.
  • Structured --output json mode suitable for piping into other tools.

What we give you

  • starter/fanout.py — CLI skeleton, client setup, and a working single-devbox example (run_one) showing the exact SDK calls you need. Start by reading it.
  • starter/mock_runloop.py — a drop-in fake client (--mock flag) with simulated latency and injectable failures, so you can develop and test without touching the network. RLFAN_MOCK_FAIL_RATE=0.3 makes ~30% of operations fail — use it to exercise your error paths. Runs against the real API too if RUNLOOP_API_KEY is set and you drop --mock.

What we're evaluating

Working code first; polish later. We care about correct concurrency, honest error handling, and guaranteed cleanup far more than clever abstractions. Talk through your tradeoffs as you go — sync threads vs. asyncio, where retries belong, what "clean teardown" means when the process is killed. It's fine (encouraged) to ask questions; the spec is intentionally a little underspecified.