Skip to content

fix(egress): fall back to IPv4 when the IPv6 path hangs (Happy Eyeballs) - #849

Merged
fclairamb merged 2 commits into
feat/todos-2026-11-11from
feat/todos-2026-11-12
Nov 12, 2026
Merged

fclairamb merged 2 commits into
feat/todos-2026-11-11from
feat/todos-2026-11-12

Conversation

@fclairamb

Copy link
Copy Markdown
Owner

Stacked on #848: merge it first.

Races IPv6 and IPv4 in the enforcing egress dialer (second family starts after 300 ms or on first failure), so dual-stack targets with a lossy IPv6 path no longer time out.

Not done (needs a spec): degraded status when only one family fails, reporting the answering family, v4/v6 sub-probes.

Closes fclairamb/solidping-business#69

…target (fclairamb/solidping-business#69)

Under an enforcing egress guard, a dual-stack target whose IPv6 path drops
SYNs looked down: DialContextWith (internal/egress/egress.go) dialed the
pinned IPs one after the other, each with the 30 s dial timeout, so the
IPv4 address was never tried within a probe's budget. It now dials them
Happy-Eyeballs style like net.Dialer (happyeyeballs.go): the first family in
order, the other raced in after FallbackDelay (300 ms default), first
connection wins, the loser is closed. TestEnforcingDialFallsBackToIPv4WhenIPv6Hangs
fails before the fix (v6 error after the probe budget) and passes after.

Closes fclairamb/solidping-business#69

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@fclairamb fclairamb added autopilot frozen Batch PR under review: takes only its own fixes, new work goes to the next batch labels Nov 11, 2026
@fclairamb
fclairamb merged commit 0c5ccaf into feat/todos-2026-11-11 Nov 12, 2026
3 checks passed
@fclairamb
fclairamb deleted the feat/todos-2026-11-12 branch November 12, 2026 08:06
fclairamb added a commit that referenced this pull request Nov 12, 2026
…ack family (#852)

Stacked on #848: merge it first.

## What changed
- HTTP and Prometheus checks now return `warning` (rolled up as
degraded) when a dual-stack target only answered on its fallback family,
e.g. IPv6 unreachable but IPv4 up (answers the "which family failed"
part of the issue, the IPv4 fallback itself shipped in #849). The result
carries `output.ip_fallback` and the dashboard shows it. Pinned-family
checks never warn.

## Decided by autopilot
- Spec fclairamb/solidping-business#69 follow-up, Q1: warn only when the
first family failed or was still pending 1 s after its start
(recommended answer).
- Q2: no separate v4/v6 sub-probes (recommended answer).

## Notes
- With the 1 s threshold and the 300 ms race, a silently dropped v6 SYN
with instant v4 will not warn; tune `FamilyFallbackThreshold` if wanted.
- An HTTP check on `localhost` can now warn when `::1` is refused but
`127.0.0.1` answers.
- Batch gate: build, lint, Go tests, e2e typecheck/lint green;
Playwright not run locally (left to CI).

Closes fclairamb/solidping-business#69

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

automerge autopilot autorelease frozen Batch PR under review: takes only its own fixes, new work goes to the next batch

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant