Repository navigation
relay: a re-REQ can cancel the read that replaced it #950
Description
Correction after recon on job 010: this is not a regression from the TypeScript capture. The capture has no per-id table, cleans up by object identity, and does not abort on a plain re-REQ. The silent stop is a shape the Rust port introduced. Anyone fixing this should not "restore" capture behaviour — there is none to restore here.
Measured: a probe deterministically reads Err(RecvError) where a live read should be. That is the deciding check — one assertion in the shape of the existing connection test, not a scenario.
Where
crates/nostream-relay/src/connection.rs:213-241 — the replacement path begins a new read for an id that already has one; the old read's end_read then tears down the new read.
Ownership idiom
No epoch or token idiom exists to reuse: AtomicU64 appears only in a test double, and tokio watch handles have no PartialEq, so a per-read token cannot be derived from the receiver. A (u64, Sender) pair under the existing map lock is the cheap fit — the lock is held while the previous sender is signalled, so a token allocated before the lock would change the lock discipline and a counter incremented under it does not.
Acceptance
- A re-REQ for a live id delivers frames from the new read.
- The old read's termination does not cancel the new one.
- A regression test that fails at
740ef69and passes after.
Activity
chappie-daemon commentedon Oct 24, 2026
Build started — factory job
010-relay-concurrency-fixes(spec stage done, builder running onfactory/010-relay-concurrency-fixes).The spec's requirement for this issue (R1–R5): the in-flight table records a connection-wide monotone generation beside each
watch::Sender,end_readtakes the read's own handle by value and removes the entry only when the stored generation still matches, andbegin_readreturnsInFlightRead { subscription_id, generation, receiver }. The deciding test isa_finished_read_does_not_remove_the_read_that_replaced_it, which fails at740ef69onhas_changed()→Err(RecvError(()))(probeA2). Re-REQmust still reachEOSE.fix(relay): read ownership, cancellation before the mint, and a bound…
chappie-daemon commentedon Oct 24, 2026
Landed in #962 (merge commit
bca0f6f) onrewrite/main.Cost — job 010, actual vs expected: $2.34 actual against $250 expected (0.9%). Measured by
factory-spans.tsover 17 sessions, 802 requests, 6.0 h summed wall, 3.3M in / 0.7M out / 85.2M cache-read. The job fixed all three defects in one branch; pro-rata by changed source lines this issue's share is about $0.82 (read ownership (generation guard on end_read): connection.rs 97 + handlers/subscribe.rs 54 changed lines), which is an attribution and not a separate bill — the job is the billing unit.The estimate was the defect, not the work: the model is being replaced with one built from measured jobs (see
.factory/estimates.md).