floatdrop / moq-go / 31764328097
89%

DEFAULT BRANCH: draft-19
Ran
Jobs 1
Files 102
Run time 1min
Badge
Embed ▾
README BADGES
x

If you need to use a raster PNG badge, change the '.svg' to '.png' in the link

Markdown

Textile

RDoc

HTML

Rst
coverage: 88.357% (-0.03%) from 88.386%
31764328097

push

github

web-flow 
test(relay): stop the stitch harness dying silently on a transient upstream error (#81)

TestFetch_DescendingCappedUpstreamFallsBackToWholeUnknown failed once in about
ten full `go test -coverpkg=./pkg/... ./pkg/...` runs, reporting only "the relay
never served a stitched FETCH (served=false objects=0)" after exhausting its
10s budget.

served=false for the whole budget means the relay rejected every downstream
FETCH, which it does before any upstream interaction — so its cache never
filled, so the live subscription never delivered, so the upstream goroutine had
stopped. That goroutine answers BOTH the on-demand SUBSCRIBE that fills the
cache and the stitch FETCH, and it returned on any error without recording one.
A momentary hiccup became a ten-second wait and a message describing the
symptom rather than the cause.

Two changes, both worth making whatever the trigger was:

  - Transient stream-credit exhaustion is waited out rather than treated as
    fatal. sessiontest hands opened streams to the peer through a bounded queue
    (capacity 4) and reports a full one as ErrNoStreamCredit — the in-process
    stand-in for a peer that has not raised MAX_STREAMS yet. The harness pushes
    five subgroups back to back, so a relay slow to drain can produce it, and
    the condition clears on its own. The original
    TestFetch_StitchesEvictedRangeFromUpstream harness has the same burst and
    the same silent return, so it uses the helper too.
  - Every other upstream-side failure is recorded and printed with the timeout,
    so the next occurrence names its cause.

Being straight about the evidence: the credit-exhaustion hypothesis is NOT
confirmed. An instrumented build did not reproduce it in 8 full-package runs or
14 targeted ones, and the original failure came from the heavier all-packages
run. 16 clean all-package coverage runs followed this change, against a
baseline of roughly one failure in ten — consistent with a fix, short of proof.
What is cert... (continued)

9304 of 10530 relevant lines covered (88.36%)

2.01 hits per line

Coverage Regressions

Lines Coverage ∆ File
2
81.48
 -3.7% pkg/moqt/session/control.go
1
92.91
 -0.71% pkg/moqt/session/datastream_out.go
Jobs
ID Job ID Ran Files Coverage
1 31764328097.1 102
88.36
 GitHub Action Run
Source Files on build 31764328097