prisma-risk / tsoracle / 26353705474

95%

push

github

fix(client): single-flight ChannelPool dials via per-endpoint OnceCell (#286)

ChannelPool::client checked the cache without the lock, dialed, then
re-locked to insert. Two concurrent first-callers to the same endpoint
both missed the cache and both ran connect(); the second insert clobbered
the first, wasting a full connect's latency on the orphaned channel.
The driver currently funnels RPCs serially so this is unreachable today,
but it activates on any future concurrency change.

Replace Mutex<HashMap<String, Channel>> with
Mutex<HashMap<String, Arc<OnceCell<Channel>>>>. client() now looks up or
inserts the per-endpoint cell under the parking_lot mutex (HashMap ops
only, no await), drops the lock, then dials through
cell.get_or_try_init(..). Concurrent first-callers to one endpoint share
the cell and run a single connect(); distinct endpoints still dial in
parallel. A failed dial leaves the cell uninitialized so the next caller
retries rather than caching a poisoned error. The connect.duration and
failure metrics move into the init closure so they fire once per real
dial.

65 of 66 new or added lines in 1 file covered. (98.48%)

10675 of 11266 relevant lines covered (94.75%)

507452.08 hits per line