tari-project / tari / 34224064744

76%

Build Type

push

github

Committed by web-flow

Commit Message

fix: decouple liveness ping timeout from ping interval (#7992)

`LivenessService` reused `auto_ping_interval` as the per-ping in-flight
TTL, so configuring a faster ping rate silently tightened the failure
deadline by the same factor. Three consecutive expiries hard-disconnect
the peer, so a short interval (3s in the cucumber suite, 100ms in a sync
test helper) turned liveness into a trigger-happy peer killer under load
- including tearing down the peer an in-progress sync was using.

The inversion was that `MAX_INFLIGHT_TTL` (30s) only applied when
auto-ping was *disabled*: turning the feature on at a short interval
made the failure detector up to 30x more aggressive than leaving it off.

- Add `LivenessConfig::max_inflight_ttl`, defaulting to
`MAX_INFLIGHT_TTL` (30s), and use it at both ping-send call sites. Rate
and timeout are now independent knobs. `MAX_INFLIGHT_TTL` moves to
`config.rs`; the public path
`tari_p2p::services::liveness::MAX_INFLIGHT_TTL` is unchanged.
- `disconnect_failed_peers` now skips strongly held connections,
mirroring the `!conn.is_strongly_held()` guard the connectivity manager
already applies in `reap_inactive_connections` and
`minimize_connections`. Liveness was the one reaper that bypassed it,
which is what allowed it to abort a running sync.
- Tests: a short `auto_ping_interval` must not shorten the TTL or
increment the failed-ping counter, plus a strongly-held-connection guard
test, each with a control proving the harness can observe the
disconnect.


<!-- greptile_comment -->

<h3>Greptile Summary</h3>

This PR separates the liveness ping timeout from the automatic ping
interval and protects strongly held connections from liveness-driven
disconnection.
- Adds a 30-second default `max_inflight_ttl` and applies it to manual
and automatic pings.
- Preserves the public `MAX_INFLIGHT_TTL` re-export.
- Skips failed-peer disconnection while another subsystem strongly holds
the connection.
- Adds regression and control tests ... (continued)

Coverage Stats

123 of 126 new or added lines in 2 files covered. (97.62%)

58 existing lines in 11 files now uncovered.

84442 of 135598 relevant lines covered (62.27%)

259338.17 hits per line