prisma-risk / tsoracle / 26300061678

95%

Build Type

push

github

Committed by web-flow

Commit Message

fix(client): bound coalescing driver waiters and stream chunk delivery (#115)

* fix(client): bound coalescing driver waiters and stream chunk delivery

The driver's secondary VecDeque<Waiter> was unbounded: the in-flight
select! drained rx.recv() into it without backpressure, so the documented
4096-entry waiter bound did not hold under a slow server. run_chunks also
accumulated every chunk's Vec<Timestamp> before any delivery, so a slow
first chunk amplified retained memory and tail latency.

Split QUEUE_CAPACITY evenly between the mpsc buffer and the secondary
VecDeque, gating rx.recv() on queue.len() < QUEUE_CAPACITY/2 in both
select sites that pull from rx. Move deliver() inside run_chunks so each
chunk's outcome streams to its waiters before the next chunk's RPC
starts; the only state crossing chunk boundaries is the Option<ClientError>
used for fail-fast. Update the per-waiter footprint estimate (~256 bytes,
~1 MB at the bound) to account for oneshot heap allocations.

* test(client,server,tests): cover driver backpressure end-to-end

Three layers of additional coverage for the coalescing driver fix:

- `StallableDriver` test fixture in `tsoracle-server::test_fakes` wraps
  `InMemoryDriver` with an indexed stall knob on `persist_high_water`,
  so integration tests can simulate a slow consensus driver mid-flight.
  Stalled calls bump their persist target to `now + 60 s` on resume so
  the allocator's post-extension window check still serves the in-flight
  GetTs — what a real driver recovering from a transient slowdown would
  do.
- Integration tests in `tsoracle-tests/tests/client_backpressure.rs`
  exercise the driver through the real gRPC stack: an 8192-caller smoke
  test for the new `select!` gates, a per-chunk delivery test that
  stalls the second chunk's persist and asserts the first response
  arrives without waiting, and an `#[ignore]`-tagged soak that mixes
  count distributions over 512 concurrent tasks × 200 iterations.
- proptest in... (continued)

Coverage Stats

183 of 185 new or added lines in 2 files covered. (98.92%)

2 existing lines in 1 file now uncovered.

7088 of 7382 relevant lines covered (96.02%)

684437.23 hits per line