Skip to content

I0922 — The memory hypothesis is tested, not asserted

Run the production image under the intended memory limit, measure resident memory through a representative request sequence, and let the evidence decide the memory the deployment declares.

BLOCKED wave 6 · p0 · implementation profile · parallel safe

Part of fly-deployment — A deployment is a projection of the repository model, and a reachable URL is its evidence.

Blocked. This issue cannot start until I0930 is done. The status is derived from that, not declared.

Objective

Run the production image under the intended memory limit, measure resident memory through a representative request sequence, and let the evidence decide the memory the deployment declares.

Why

The cheap profile rests on a guess about memory. Raising it because a run failed once is how a hosting bill quadruples; keeping it because it looks tidy is how a Machine gets killed under load. Only a measurement settles it.

Current state

The canonical object states a memory hypothesis. Nothing has been run under any limit.

Desired state

Resident memory is measured at startup and after each representative request, the high-water mark is recorded, and the declared memory is the smallest value the evidence supports.

Scope

  • .ai/repo/benchmarks/rust

Out of scope

  • Raising the limit before diagnosing what consumed the memory
  • A load test; this is a resource envelope, not a throughput study

Dependencies

What waits on this

Acceptance criteria

  • The image runs under the declared limit and stays alive through the whole representative sequence
  • Resident memory is measured after startup and after a landing, documentation, API, OpenAPI and cockpit request, and the high-water mark is recorded
  • If the declared limit does not hold, the cause is diagnosed and the waste addressed before the limit is raised, and the raise is recorded with its evidence
  • The measurement is recorded through the existing benchmark infrastructure and becomes a regression budget

Validation

  • majordomus bench --check
  • bash test/run.sh

Evidence required

  • rss_measured
  • runs_at_limit
  • budget_recorded

Evidence

None recorded. Every token above needs a command or an artifact behind it before this issue can be completed; narrative is refused.

Risk

A single machine's measurement is not a guarantee on another host. The high-water mark needs headroom, and the headroom needs to be stated.

Timeline

started
—
verified
—
completed
—

Those three fields, the evidence above and the state of the dependencies are all the status is made of. There is no status field to disagree with them.

Canonical record: .ai/repo/project/issues/I0922.yaml. Read it back with majordomus plan show I0922.