prisma-risk / tsoracle / 26367782141

95%

Build Type

push

github

Committed by web-flow

Commit Message

fix(client): evict cached channel on transport-class RPC failure (#239) (#292)

PR #290 evicts ChannelPool entries on a failed *dial*, but a channel that
dialed successfully, got cached, then had a later RPC fail with a transport
error was never removed. A tonic static `Endpoint` resolves its address once
at connect and never re-resolves, so after pod replacement (new IP behind the
same endpoint string) the cached channel's background reconnect loop hammers a
dead address forever and is reused on every future resolution onto that string.
The leader cache's `leader_ttl` ages out the leader *pin* but nothing aged out
the *channel* — a slow resource leak and staleness hazard in rolling-restart
topologies.

`attempt` now keeps the channel's `OnceCell` handle for the duration of the RPC
and, on a transport-class failure, hands that exact cell to a new
`ChannelPool::evict_if_current` so the dead channel is dropped (forcing the next
attempt to re-dial and re-resolve) without racing a concurrent re-dial — the
same `Arc::ptr_eq` identity guard the dial-failure path already uses. Transport
class is `Unavailable`, transport errors, and `DeadlineExceeded` (a half-open
connection that black-holes until the per-attempt deadline rather than failing
fast); application errors such as `Internal` leave the channel cached because
the connection is healthy.

The backoff and channel-eviction policies share one trigger today, so the
classification is factored into `is_transport_failure` with `should_backoff`
delegating to it; divergence would have to be a deliberate edit. `client` is
split into `client_with_cell` (returns the cell handle for the retry loop) plus
a test-only wrapper.

Coverage Stats

178 of 196 new or added lines in 3 files covered. (90.82%)

2 existing lines in 1 file now uncovered.

10996 of 11602 relevant lines covered (94.78%)

504253.34 hits per line