bedrock-kv / bedrock / 12776d422a8e4313b36a1a15d80b06815c91a3f9

82%

push

github

Make the foreman's health verdict independent of worker order (#214)

Closes bedrock-287.

`compute_health_from_worker_info/1` was a pairwise fold whose clause
`{:ok, _}, _ -> :starting` overwrote an already-recorded failure, so the
verdict depended on the order workers arrived in:

| input | old | new |
|---|---|---|
| `[ok, bad, bad]` | `{:failed_to_start, _}` | `{:failed_to_start, _}` |
| `[bad, bad, ok]` | `:starting` | `{:failed_to_start, _}` |
| `[ok, bad]` | `:starting` | `{:failed_to_start, _}` |
| `[bad]` | `:starting` | `{:failed_to_start, _}` |

`recompute_health/1` folds `Map.values(t.workers)`, and small Erlang
maps iterate in term order — so whether a wedged worker was reported
came down to how worker ids sort. Exhaustively: over every multiset of
size 1–3 drawn from the four reachable health shapes, **18 distinct
multisets change verdict under permutation**. The last row is the
starkest — a foreman hosting exactly one worker, and that worker failed
to start, reported `:starting`.

A second bug shared the clause list: `:stopped` matched only the
catch-all, so a still-starting worker yielded `:unknown`.

### Approach
State the verdict as precedence over the whole collection rather than as
a pairwise fold, so order-independence is structural instead of
something to re-derive clause by clause.

### No caller can observe a difference in what anything actually reads
The `:ok` / not-`:ok` partition is **identical** before and after: the
old fold could only reach `:ok` via `{:ok, _}, :ok` and never re-entered
it once it left, so both versions return `:ok` exactly when every worker
is running. Both consumers (`do_wait_for_healthy/2`,
`notify_waiting_for_healthy/1`) match `:ok` literally.

`:unknown` is kept and is **not** dead code — `Foreman.report_health/3`
accepts `Worker.health()`, which includes `{:error, :timeout |
:unavailable}`, and that reaches this field.

### Verification
2803 tests, 0 failures. Credo and dialyzer clean.

Adversariall... (continued)

9 of 10 new or added lines in 1 file covered. (90.0%)

6802 of 8655 relevant lines covered (78.59%)

1046.31 hits per line