Session 14. Deploy and operate the capstone — Thu 01 Oct

Quick quiz (ungraded)

Nothing; the second call is the more stable measurement
It is stable and it is about a warm process. The question was what the first caller pays.
The 2.4 seconds the first caller waits, every time the process has to start
Correct. The first call pays for the process starting. Report the first call, and keep it separate from the verdict.
The p95, which needs at least twenty runs
A percentile is a different measurement with a different method. The cold start is one number: the first one.

malformed_rejected: true — it handled the request without crashing
It handled nothing. It answered a body it could not have understood, which is the failure, not the recovery.
malformed_rejected: false, and healthy: false
Correct. Validation has stopped running. A 200 for garbage is indistinguishable from working at every level except the content, so the report is the only place it can surface.
Nothing; the smoke test should skip malformed inputs and measure the happy path
A boundary you never probe is a boundary you are assuming. Send the garbage on purpose.

It is not worse; a crash is loud and somebody will notice
Somebody notices the crash and still has no report about the deployment, which is what they came for.
The report never gets produced, at the one moment it mattered
Correct. A smoke test that dies on a broken deployment tells you nothing about it. Read the status first, and never reach into a body you have not checked.
It hides the cold start
The cold start is measured before that, on the first call. The loss here is the whole report.

We would roll back quickly if the error rate looks bad.
An intention with no action anybody can run, and no number. 'Looks bad' is a feeling.
Return the service to the last good state as soon as possible.
It names a state, not a way to reach it, and 'as soon as possible' is not a duration.
Stop the service and restart it from the previous tag: git checkout v0.3.1 && uv run python serve.py; answering again inside 2 minutes.
Correct. A named action, a version, and a duration — the same standard an architecture decision's reversal trigger has to meet.

That it works, since the concept page describes it
A description is not a deployment. The page states in its own words that the flow is not merged and not deployed.
That it is planned, that nothing here depends on it, and that it is a claim to check rather than a step to follow
Correct. Same standard as a recorded fixture in session 13: evidence of what existed when it was recorded, and of nothing else.
That local deployment is a workaround for it
The local path is the real one. Every failure this session teaches — cold start, resource limit, malformed request, rollback — is visible on a port on your own machine.