bus-conformance.md still listed the example_demo root count and
session_over_one_router as not fixed here. Both are fixed on main now, so
the rows say what each test does instead: the example waits for the
producer's hold release before reading the counts, with its printed
output unchanged, and the integration renderer is held until the
publisher's twentieth receipt has returned, so its coalescing is forced
and the assertions are order, freshness, acceptance under a stalled
spectator and the replacement accounting, with the before and after
counts.
The delivery check in that test compared (step, sequence) against
(step, step), whose first element can never fail. It compares sequence
against step.
pending_connections_are_bounded_and_hello_expires gave the router 20 ms
to register a connection and then slept 80 ms past a 40 ms hello timeout
before reading the count once. Both are claims about how fast this box
is, not about the router: the first failed once in 20 whole-crate runs
beside four busy loops. Both now wait for the event they are about, under
the suite's ordinary ten-second bound, so the test still fails if a
pending connection never registers or a pending Hello never expires.
session_over_one_router asserted that a latest subscriber must drop
snapshots. bus-v1 section 7 says a latest subscription replaces an
undelivered value; it lets a consumer miss values, it does not oblige it
to. Under contention the publisher was slow enough that the renderer kept
up, saw all twenty snapshots and the test failed on conforming behaviour:
16 of 40 runs beside four busy loops, and 9 of 20 whole-crate runs.
The renderer is now held until the publisher's completion is observed
rather than until a timer expires, so the coalescing is forced instead of
raced for: the subscription keeps the one delivery in flight and one
replaceable queued value, and the renderer receives snapshots 1 and 20 of
20. The assertions are the guarantees that hold -- what arrives is in
publication order, the last value is the latest published, the stalled
spectator never refuses a publication or drops out of the fan-out, and
every snapshot the renderer missed is counted as a replacement to the
publisher at admission and to the renderer on delivery, so nothing is
lost silently.
Three more tests in the crate asserted the same kind of race:
- latest_replay_is_ordered_ahead_of_a_racing_publish demanded the
non-coalesced outcome of a race section 7 allows either way ("bounded
mode preserves that order, while latest mode may coalesce it"). It now
puts one racing publication to a bounded and a latest subscription at
once: bounded must deliver the replay and then the publication, and the
latest branch is chosen by that publication's own replaced count.
- collection_waits_for_every_retained_owner, and the two disconnect
cleanup tests beside it, read the store directory for the unlink that
follows the registry update outside the router lock. They use
settle_files, like every other unlink check in the suite.
- The demo example printed a root count taken before the producer's own
release had reached the router, so the line the guide quotes was a
race. It waits for the release, the same way it already waits for
collection; the printed output is unchanged.
Nothing under flybus/src is touched: no routing defect was found. The
conformance rows for the credit/queue split, the replaced count, the
replay ordering and the latest spectator that cannot refuse a publication
now cite what each rewritten test actually varies.
caller_disconnect_cleanup_works_before_and_after_consumption failed 37 times
in 240 runs under load: the reply raced the router's own teardown. op_reply
reports routed:false only once disconnect has marked the call detached, and
the router runs that when its connection task reads EOF, while Client::close
waits for this client's reader only. The reply could therefore reach a
connection that was still closing and be routed. Nothing escaped -- teardown
releases those roots -- but the flag was read one step early.
The test now settles on the caller's connection being gone before asserting
the section 6 sentence, which is the poll the rest of the file already uses.
360 runs after the fix, 0 failures.
bus-conformance.md records the mechanism, the counts and the two other
intermittent failures seen in the same sweep, which are left to the bus slice.
An independent review of 4f44894 accepted the audit with two required evidence
changes, neither of which needed a code change.
"Classification is not a topic-name heuristic" cited a lane-order test that
never varies a topic name. The row is now conforms-by-construction, naming the
four scheduler functions that read only the queue an item sits in, and
a_topic_named_like_a_notice_is_still_classified_as_topic_data publishes on
topics called call.failed, route.removed and subscription.closed and asserts
each arrives as a topic.message delivery with its payload intact.
"A latest subscriber never causes BACKPRESSURE" cited a bounded-mode test and
the ignored measurement. a_latest_subscriber_never_refuses_a_publication floods
one unconsumed latest slot with 100 publications of 60 KB, six times the
per-client bounded byte pool, and asserts none is refused, 98 coalesce, the
queue stays at one and the credit never comes back. That row also underwrites
the byte-budget deviation and the first contradiction.
Both tests run over both transports.
The coordinator resolved both contradictions in the spec rather than the code,
so bus-v1 carries them as dated amendments with one line of reason each: the
section 9 table row becomes "Per-client ordinary bounded queued envelope bytes"
with a new "Latest subscription slots | subscriptions x 64 KiB" row, and the
section 2 sketch drops the budget argument and shows the deadline at the caller.
With the amended wording the byte-budget row conforms, so the report's counts
are 178 / 9 / 1 fixed / 7, and the crate README's differences 4 and 8 say they
no longer differ.
The guide's first deliverable is one small program with a counter RPC, a pub/sub
observer and a frame artifact held past its message object. examples/demo.rs was
already that; its body moves into run(), which returns the lines it prints, so
tests/example_demo.rs can assert all seven of them. The counter now uses the
spec's own example.counter / Counter.Increment / {"amount": 1} and pins the
registration it discovered, and the example prints the router's root count while
the frame is held and after the last handle goes, so the lifetime it
demonstrates is visible rather than implied. cargo run -p flybus --example demo
prints the same lines.
tests/perf.rs reports what BUS-03 and bus-v1 section 11.7 actually ask for.
Allocate, the producer's copy into staging, the router's seal copy, publish
admission, the RPC round trip and a consumer's readback are six separate
percentile lines instead of one. The router gets its own two-thread runtime
whose threads carry a distinct name, and per-thread CPU is sampled from /proc by
that name, so router CPU is separable from the clients' in the same process.
Store bytes, outstanding roots and queue lengths are reported live as well as
peak.
One named test per implementation-guide bullet that the existing suites did not
already cover: a lost result that leaks no roots while the endpoint's own cache
still replays it, a retransmission of one domain body under a fresh call id, no
failover onto a replacement registration, a status RPC answered while another
handler is delayed, and a disconnect that reclaims logical ownership without
touching a file the consumer still has open. Each runs over both transports.
BUS-01 also asks for equivalent behaviour traces from the two transports, so
tests/common/mod.rs gains a Trace recorder whose record() panics on any
router-issued id: a trace can hold methods, payload fields, counts, sequences,
credits, cancel states and error codes, and nothing operational. One RPC
scenario and one pub/sub-with-artifacts scenario produce 29 events, identical in
memory and over a Unix socket. FLYBUS_TRACE=1 prints them.
The sixteen bullets the older suites already prove are cited in the conformance
report rather than duplicated here.
The write gate was one mutex the writer held across every synchronous transport
poll, and close_conn took that same mutex before reclaiming a connection's
owners. Under a transport that accepts a byte per poll the writer re-acquired it
hundreds of times while teardown queued behind it, so teardown could be starved
for a whole frame and the delivery completed just before its owner was
reclaimed. bus-v1 section 9 puts release and route-health control out of reach of
telemetry starvation; tests/sol_rereview_regressions.rs's poll-gate regression
failed 6 release runs out of 6 alone, and about one debug run in five.
The gate now separates "teardown has begun" from "a poll is in progress": a
mutex plus a condvar. begin_close sets closing once without waiting, then waits
for the poll already in progress and records whether it left a frame half
written. The writer brackets each poll with enter_poll (refused once closing)
and leave_poll(bytes) instead of holding a lock across it, so teardown's window
is one poll rather than a frame, and no byte follows it. close_conn no longer
holds the gate across reclamation. Every item is pub(crate): no public
signature changed.
The regression test no longer synchronises with a sleep. The shutdown thread
announces itself on a channel, and what follows is structural: teardown cannot
pass the gate until the held poll returns, which only release() allows. Its
delivery is 50 KB now, which a writer resuming a byte per poll cannot finish
inside the window, so the test observes the ordering instead of a coin toss and
asserts the two legal shapes of the stream: cut short with nothing appended, or
never begun with exactly the closing notices following.
