smarm::causal is feature-gated, so examples/load_profile.rs no longer
compiles under the default feature set and broke `cargo build --examples`.
Declare it with required-features = ["smarm-causal"] so the default build
skips it and `--features smarm-causal` still builds it.
Splits the committable half out of the working-tree Cargo.toml; the
`smarm = { path = "../smarm_full" }` dev pin stays uncommitted and is
restored to a tag in the release commit.
ROADMAP gains a v0.8 entry: the smarm 415effb lifetime change as root cause,
the description/instantiation split, the tree shape, the accepted per-call
resolution cost, and what stays open (RegistryName newtype, PubSub's
unprotectable const, dynamic session actors, hammer.sh's missing feature
matrix). The v0.7 "unreproduced test failure" open item is closed out and
pointed at it — it was this hang, hidden because hammer.sh builds with default
features while the failure needs --all-features load.
README and the two module headers still taught the v0.5 rules: in-runtime-only
construction, the non-static Arc<OnceLock<..>> cell, and "a relay must never
hold a PubSub clone". None of those are true any more; they are replaced by
what actually holds now, with a note on what changed for anyone who learnt the
old shape.
session.rs also gains an honest note that its session actors are the last
lifetime in urus implied by a drop rather than stated — they are dynamic, so a
fixed ChildSpec list cannot hold them, and the trigger is at least a command
now rather than a refcount.
clippy::byte_char_slices, pre-existing and unrelated to the refactor in the
previous commit — split out so that diff stays scoped. Restores a clean
`cargo clippy --all-features --all-targets`.
smarm 0.7 (415effb, "lifetime is the actor's — refs are addresses") removed the
rule these three actors were built on: a GenServerRef no longer owns the server,
the loop holds its own inbox sender, and the inbox never closes when the last ref
drops. urus's pubsub table, channel hub bus and session registries were still
governed by that deleted rule — PubSub::new() spawned the table and the handle
owned its life — so nothing commanded them to stop. What still terminated a run
was the root-exit sweep, racing the drain: shutdown_with_open_chat_terminates and
channels_wire::shutdown_with_open_channel_terminates failed 4 times in 25
--all-features runs with "serve did not return: ... outlived the drain: Timeout".
Zero in 10 full-suite runs after this change.
The fix is not a supervisor wrapped around the old shape. Every gotcha in this
area descended from constructors that spawn: PubSub::new(), ChannelHub::new() and
PrefixRouter::channel_session() all started actors, which forced in-runtime-only
construction, which forced the Arc<OnceLock<..>> lazy-init from the first handler,
which forced the "cell must not be static" and "a relay must never hold a PubSub
clone" rules. Five documented rules propping up one inverted dependency. So:
description is separated from instantiation.
- PubSub<M> is a name, not a GenServerRef: const-constructible, Copy, spawns
nothing, valid outside the runtime and in a static. Operations resolve through
the registry per call, so a table restarted by its supervisor is reached
transparently (one lookup per broadcast — bench before caching a ref, which
would go stale across exactly the restart the supervisor exists to perform).
PubSub::new() is gone; PubSub::new(name) + PubSub::child() replace it.
- ChannelHub::new(bus, router) returns (hub, Vec<ChildSpec>) — the bus table plus
one registry per session route. Returning both is the point: a hub whose
children were never started compiles and fails on the first join, so the vec is
not left behind a method you can forget to call. #[must_use].
- channel_session gains a registry name; each session registry is separately
named and separately supervised.
- serve_with/serve_with_shutdown take a Vec<ChildSpec> of app children and build
the root as RestForOne[..app children, endpoint]. They start before the
endpoint and, shutdown being ordered in reverse, stop after it has drained, so
a request still in flight can reach the bus. RestForOne because a bus crash
leaves live sockets addressing a table that no longer knows them.
- Deleted: the Arc<OnceLock> idiom from both examples and both test pipelines,
and the module rules that existed only to hand-manage a refcount.
Known cost, not fixed here: channel_session("session:*", "chat-sessions", f) puts
two unrelated string literals side by side and nothing catches a transposition —
a RegistryName newtype is the obvious follow-up.
Tests: 111 lib + 50 integration + 2 doc green, clippy clean, 10/10 full-suite
runs. Unit tests poll for name binding before use — smarm's start-order-is-not-
start-readiness gap; real apps don't hit it, since a handler only runs once a
connection has been accepted.
Full account of the refactor: the tree shape, why the endpoint spawns its
own listener sup (supervisor start order != start readiness), the config
split, what was deleted and the smarm properties re-probed to justify
deleting it. Also:
- PubSub discovery note: the non-static Arc<OnceLock> rule is now
serve*-only; an app that owns its tree starts the table as a supervised
sibling and addresses it by name (what crud does).
- Icebox 'operator introspection via typed names': partly delivered — the
endpoint is a named gen_server answering Call::ConnCount; per-listener
visibility and richer stats remain open.
- Open after this cycle: the one unreproduced hammer failure, and that
endpoint() returns impl Fn() so a second invocation clashes on the name.
- crud is now the demonstrator: app owns smarm runtime + root supervisor,
store actor and urus::endpoint as ordered siblings (store first, so
reverse-order shutdown drains HTTP before stopping the store),
Shutdown::Infinity on the endpoint, shutdown via
rt.handle().request_shutdown(root_sup) from the stdin thread.
Deletes two kludges the old shape forced:
* static OnceLock<Sender> spawn-on-first-use -> a supervised child
that self-registers a typed Name; handlers use smarm::send per
request and turn 'between incarnations' into a 503 instead of
panicking on a dropped store.
* static SHUTTING_DOWN AtomicBool + 250ms recv_timeout poll in the
store loop and in the SSE ticker -> a plain park; the tree stops
both. Smoke-tested live: CRUD round-trips, SSE stream, clean drain
with the stream open, port closed after.
- Other examples stay short and on serve*, updated for the split config
(serve_with(cfg, smarm::Config, pipe) / serve_with_shutdown(..., signal)).
plain_serve's URUS_SCHED_THREADS now builds a smarm::Config.
ws_chat's doc block explains the OnceLock is a serve*-only workaround
and points at crud for the clean shape.
- README: new 'Your Own Supervision Tree' section (endpoint as the real
API), graceful shutdown reframed as the serve*-only path, Config table
loses scheduler_threads and gains name, PubSub rule 1 notes the
supervised-sibling alternative.
111 lib + 50 integration + 2 doc tests green; clippy clean.
The v0.3 shape from the spec: an app owns its runtime and root supervisor
and places urus in it as one ordered child among its own.
your root sup
└── ChildSpec(Permanent, urus::endpoint(cfg, pipeline)?) <- Endpoint
└── listener_sup OneForOne over N listeners
└── plain connection actors
- src/conn_registry.rs -> src/endpoint.rs. The registry gains the listener
pool it registers for and becomes the Endpoint gen_server; ConnRegistry
-> Endpoint. It runs inline as the ChildSpec's actor
(NamedGenServerBuilder::run), so supervisor shutdown arrives as
handle_shutdown and a restart re-runs init on the same still-open fds.
- Endpoint spawns its OWN listener sup in init rather than being its
sibling: smarm's supervisor start order is not start *readiness* (spawn
is fire-and-forget), so a sibling listener could whereis the name before
the registry actor ran. Registrar-spawns-consumers makes it program
order inside one init. Gap filed in smarm ROADMAP (readiness ack);
making spawn blocking would only shrink the window, not close it —
'has begun executing' is not 'has bound its name'.
- Listener sup is monitored: death outside shutdown = panic (loud, the
app's supervisor decides) instead of a zombie on a dead port. Death
during shutdown is the 'no new connections' barrier.
- DELETED: the shutdown AtomicBool, LISTENER_TICK (250ms wake per listener
per tick, now an untimed wait_readable park), SHUTDOWN_POLL (100ms root
poll — the root parks on the signal channel now), Restart::Transient
(listeners are Permanent: they only exit by supervisor action, so a
self-exit always means breakage). Verified against current smarm:
request_stop unwinds an untimed wait_readable park, is no longer lossy
against a QUEUED actor, and supervisor shutdown joins in ~200us.
- Config: scheduler_threads/max_actors removed (runtime knobs an
endpoint-as-child cannot honour) -> serve_with(cfg, smarm::Config, pipe)
and serve_with_shutdown(cfg, smarm::Config, pipe, signal). Added
Config.name (default 'urus'): the endpoint's registry name, unique per
endpoint, and the introspection handle via endpoint::whereis(name).
- serve* keep their meaning as the batteries-included path: they build a
one-child tree around endpoint() with Shutdown::Infinity. Handle stays
(a serve* caller has no RuntimeHandle to reach for) and now backs a real
park instead of a poll.
Tests: 4 drain tests ported onto a real endpoint (bound socket, supervised
child, request_shutdown driven); new integration test boots an app tree
with an ordered sibling and asserts serve-then-drain, reverse-order
teardown and a closed port. 106 lib + 50 integration green.
The drain protocol moves wholesale into the registry (smarm >=0.7:
trap_exit + handle_shutdown + gen_server timers + StopHandle):
- handle_shutdown: flip draining, stop idle conns, arm a drain_timeout
timer, Continue; with no conns, Exit immediately.
- ConnIdle while draining stops the conn (unchanged); ConnStarted while
draining stops it on arrival (unchanged); ConnEnded that empties the
set while draining = StopHandle::stop() — the registry's own normal
exit is now the 'every connection is gone' barrier.
- handle_timer (deadline): one force-stop sweep. The old re-sweep-every-
10ms loop existed to catch late registrants; stop-on-arrival already
covers every post-sweep entry path, so one sweep suffices.
- Cast::{BeginDrain,ForceStopConns} deleted (internal now); ConnCount
stays as the introspection call. start() takes drain_timeout.
- serve.rs: the root's whole drain/poll block collapses to
registry.shutdown() (graceful, monitors until the server has stopped
itself). SHUTDOWN_POLL + the listener flag are untouched here; they go
with the endpoint refactor.
- Known residual window documented in the module docs: a conn spawned
by a dying listener that has not yet registered can outlive an
already-empty registry; it is collected by smarm's root-exit sweep.
Tests (in-lib, request_shutdown driven): empty-set immediate exit;
idle-now/busy-at-deadline ordering with exit-after; stop-on-idle
mid-drain; stop-on-arrival mid-drain. 35x hammer subset green.
Discriminates herd-contention vs stranding for the ka low-c latency
floor: sweeps URUS_SCHED_THREADS {1,2,4,8} x CONNS {4,8} over the plain
matrix (PLAIN_ONLY; close rides along as control). Per-cell audit
asserts the knob reached the server via plain_serve's stderr line.
Summary parser (req/s + p50/75/90/99, unit-normalized) validated
against the 96b40ad5 sweep output.
E1 (herd-vs-stranding discrimination) sweeps smarm scheduler count at
fixed low concurrency. Strict parse — a malformed value panics instead
of silently reverting to default, and the effective value is logged to
stderr so every bench cell's server.log records it (same per-cell
verification discipline as mode-verify).
Server-only discovery tool: real hot path, no planted bottleneck, no
in-process loadgen. wrk (or any external generator, pinned to other
cores) drives GET /json/:id; the sweep runs the causal_bench plan over
the five lib sites. Trust gates only (traffic flowed, ledger printed) —
no known-answer verdict, this one asks the question instead.
examples/causal_bench.rs (required-features smarm-causal): the Phase 2
RFC 007 test case — a real urus server replacing the synthetic demo as
validation workload.
Shape: 16 plain-OS-thread clients hammer GET /order/:id over blocking
loopback keep-alive TCP (OS threads can't absorb injected virtual delay
— the established trick, so only server code is delayable). The handler
calls a single store actor (crud's once-cell pattern, serialization
structural) burning 400µs of calibrated *work* under
causal_site!("store") — work-shaped, not timed, or injection reads as a
no-op. 50µs handler render lands under the enclosing pipeline site.
Known answer: store serialized + saturated => ceiling 2500 rps; speedup
p => x1/(1-p): +33% @25, +100% @50. The five lib sites are tens of µs
and parallel across conns => ~0. Verdict mirrors the demo: store @25 >
+15%, others < +10%, SKIPPED under 4 cores; magnitudes trusted to ~±15%
per the smarm-side validation notes, ranking is the hard check.
1-core sandbox smoke (verdict SKIPPED but numbers indicative): store
+26.2% @25 / +74.0% @50, all other sites within ±2%. Unlike the demo
this workload keeps its bottleneck structure on one core — everything
but the store is IO-parked — and the theory shortfall there is plain
core contention (render/parse share the store's CPU), which the 24-core
sweep should close.
RFC 007 causal-site coverage for the Phase 2 bench (and any downstream
profiling): parse (head parsing attempts in read_head), router (dispatch
matching only — the winning handler and the next fall-through run outside
the guard, so the site measures dispatch, not what it dispatches to;
call() restructured match-then-dispatch for that, semantics preserved
incl. first-path+method-match-wins and the 405 two-pass), pipeline (the
plug chain inside catch_unwind; nested sites like router take over
attribution during their span, so this reads as chain overhead + handler
code outside inner sites), serialize (response head+bytes-body
serialisation; chunked stream framing is not covered — it interleaves
with writes in pump_stream and the bench doesn't stream), and
socket-write (all of write_all, writability parks included: a park
inside a site is exactly what the park-gated resume credit attributes).
progress!("responses") fires once per response fully on the wire, at
both completion points (the common path and the HTTP/1.0 EOF-stream
early return).
Site guards are #[cfg(feature = "smarm-causal")]-gated: the macros are
no-ops featureless, but binding the unit expansion would trip clippy's
let_unit_value. progress! is bare — its featureless expansion is an
empty block, free and lint-clean. Featureless build byte-behavior is
unchanged; both configs clippy-clean, full suite green both ways.
Mirrors the smarm-trace precedent: urus itself gains no causal code
paths yet — this just lets downstream binaries (the Phase 2 causal
bench lives in examples/) flip smarm's instrumentation on through the
urus dep. Whole suite is green with the feature enabled: the causal
runtime is behaviorally inert while no experiment is active.
Pre-existing, surfaced by the toolchain bump (rust-version is 1.95; the
sandbox gates with stable 1.97). All four are cargo clippy --fix output
with the mechanical-collapse indentation hand-tidied to house style;
the parser change is semantics-preserving (a non-100-continue Expect
value now falls to the _ arm instead of an empty if — headers.append
still runs after the match either way). No fmt pass: rustfmt would
clobber the aligned-assignment style, so only the touched lines moved.
- gen_server rename (RFC 015, 3e31606): ServerRef/ServerCtx/ServerBuilder
-> GenServerRef/GenServerCtx/GenServerBuilder across conn_actor,
conn_registry, serve, pubsub, channels::session.
- type Timer = () on the three GenServer impls (RFC 015, 57eadb5);
handle_timer/handle_idle/tick_every stay defaulted — opt-in later.
- Watcher and GenServerCtx are now generic over the server type: watcher
fields typed Watcher<Table<M>> / Watcher<Registry<P, K>>; Registry's
struct bounds strengthened to its GenServer impl bounds
(P: Encode + Decode + Send + Sync + 'static, K: SessionKey) so the
Watcher field's G: GenServer bound is satisfiable at the declaration.
- RFC 014 (a866e34) registry: register is (Name<M>, Sender<M>), self-only
— a name is a typed messaging endpoint, not a pid tag. The
introspection-only urus.server / urus.listener.{i} bindings are dropped
rather than faked with unit channels; the whereis integration test is
deleted; a proper messageable-name design is icebox'd in ROADMAP.md.
- Audit vs smarm 6c2b7e9 (queued messages dropped when Receiver drops):
pubsub's prune-on-send-failure retain still holds — send Errs once
receiver_alive is false, so a dropped rx prunes on next broadcast,
exactly what subscriber_count's doc already promised. Freeing stranded
Arc<M> broadcasts is strictly good. No change needed.
- The 7x E0283 in channels/mod.rs were cascade fallout of the generics
changes; dissolved with the port, as discovery predicted.
Suite: 90 lib + 45 integration + 2 doc, green under default, smarm-trace,
phoenix, and all-features. 3 pre-existing clippy lints (conn.rs,
parser.rs, conn_actor.rs; clippy 1.97 strictness) deferred to a follow-up
chore commit to keep this diff pure.
examples/channels_chat.rs (required-features phoenix): ephemeral
room:* and persistent session:* side by side over phoenix.js V2 JSON,
stdin-driven graceful shutdown (the ws_chat pattern). Smoke-tested
live: rejoin after transport drop lands on the same instance (state
counter persists across three transports), broadcasts carry the new
join generation's ref.
README: channels section (feature split, core trait walkthrough,
session persistence semantics incl. the re-entrant-join contract).
ROADMAP: v0.6 marked done with the full as-landed decision list,
chunked as committed.
ChannelSession<P> (src/channels/session.rs): a channel actor that
outlives its transport, buffering outbound broadcasts as the same
Arc<Broadcast<P>>s the relay path carries (zero re-allocation) until
rejoin, buffer cap, or TTL. Registered per pattern via
PrefixRouter::channel_session::<S>(pattern, factory) /
channel_session_default::<C>(pattern); one registry gen_server per
registration, monomorphic over S::Key (no type erasure).
Machinery: deploy() computes the key on the conn actor and casts the
join handshake to the registry; the registry maps key -> (pid, control
sender), forwards reattaches, spawns-and-monitors fresh actors, and
prunes on Down (pid-guarded against replaced entries). The session
actor selects [inbound, bus, ctl] attached and select_timeout([ctl,
bus], ttl-remaining) detached; detach triggers are the closed inbound
arm and ws send failure (the failed broadcast is the buffer's first
entry — nothing lost). Drain re-encodes under the new join generation.
As-landed decisions (each in module docs):
- ChannelFactory gained a #[doc(hidden)] deploy() seam (default =
ephemeral linked actor, the chunk-1 behavior verbatim); JoinHandshake
and ChannelInbox are opaque pub structs. Channel construction moved
from run_channel into deploy — same linked blast radius.
- Every attach calls ch.join() again (rejoin acks need a payload and
channels get to re-auth); session channels must treat join as
re-entrant.
- A rejected rejoin ENDS the session (no zombie sessions held open for
unauthorized clients); next join is a cold start.
- A second transport EVICTS the first (best-effort phx_close); a
session is single-transport.
- Explicit leave ends the session, not just the attachment.
- TTL per detach episode; subscribe once, on first successful join;
cap.max(1) semantics (cap 0 = first buffered message tears down,
but an empty buffer still waits out the TTL).
- Key: Clone beyond the ratified Eq+Hash+Send — the registry keeps a
pid -> key reverse index for Down pruning.
- The registry spawns eagerly at registration (channel_session is
in-runtime-only, same law as ChannelHub::new): a lazy OnceLock would
block std-sync under green threads on first-join races.
- Shutdown composes WITHOUT links: conns die -> router Arcs drop ->
registry inbox closes -> registry exits -> ctl senders drop ->
detached sessions wake on the closed ctl arm, terminate, exit.
shutdown_with_detached_session_terminates is the proof (ttl 30s vs
5s budget — only the chain can reap it).
Tests: 6 session unit tests (buffer-and-drain ordering across
reconnect, TTL expiry, cap teardown, leave, eviction incl. stale
conn-entry self-prune, rejected rejoin) + the shutdown integration
test. Shared toy codec + conn harness extracted to
channels/testkit.rs (mod.rs tests now import it; no behavior change).
Hammer: 35x lifecycle subset (+channels +session) + 3x full (phoenix)
+ 1x full (phoenix+smarm-trace) + featureless + channels-only, all
green; dbg-grep clean. Suite: 90 unit + 46 integration + 2 doc under
--features phoenix.
Json<T> newtype codec (src/channels/phoenix.rs), feature "phoenix":
speaks the Phoenix V2 wire format [join_ref, ref, topic, event, payload]
for any T: Serialize + DeserializeOwned. Replies encode as phx_reply
with the {"status", "response"} wrapper; payload-less control frames
encode payload as {}.
Deviation from roadmap sketch (documented in module docs): the roadmap
sketched a blanket Encode/Decode impl on bare P: Serialize. That is
coherence poison — with additive features, enabling "phoenix" anywhere
in the crate graph would forbid every user-written Encode impl. The
blanket therefore lives behind the Json<T> newtype: zero-boilerplate
(Deref/DerefMut/From), but opt-in per payload type.
Encode failure panics and rides the linked-teardown contract (channel
actor is linked to the conn; a payload that cannot serialize is a
programming error, not a protocol event). Decode strictness is exactly
T's Deserialize — use serde defaults or Json<serde_json::Value> for
lenient channels.
Integration (tests/integration.rs, cfg(feature = "phoenix")):
- channels_join_heartbeat_event_broadcast_leave: join ack ordering,
conn-side heartbeat, unrouted-topic error, join rejection, broadcast
incl. sender, correlated echo reply, leave -> ok + phx_close, and
post-leave isolation.
- channels_codec_garbage_closes_1002: undecodable text frame closes
the socket with protocol-error status.
- shutdown_with_open_channel_terminates: drain force-stop tears down
the linked channel actor and the runtime reaches AllDone in budget.
Hammer: 35x lifecycle subset (+channels filter) + 3x full (phoenix) +
1x full (phoenix+smarm-trace) + featureless + channels-only, all green.
Suite: 84 unit + 45 integration + 2 doc under --features phoenix.
Feature 'channels' (zero new deps): wire-neutral ChannelFrame/FrameRef +
Encode/Decode codec traits (one method each, whole-envelope — the codec
owns the wire format end to end); Channel trait; ChannelFactory +
TopicRouter + shipped PrefixRouter (exact map + head:* prefix map);
ChannelSocket (reply/push/broadcast/broadcast_from/broadcast_to);
ChannelHub (router + PubSub<Broadcast<P>>, in-runtime + non-static per
the v0.5 pattern, hub.upgrade(conn) entry, server-side hub.broadcast).
Topology as ratified: one channel actor per (socket, topic), linked to
the conn actor, select(&[&inbound, &bus_rx]) with inbound at index 0;
subscription pinned to the channel actor pid. Subscribe-before-ok-reply
makes the join ack first-class (the v0.5 on_open race, answered
structurally). Cleanup is the sender-drop chain (conn handler drops ->
inbound closes -> closed arm wakes -> terminate); bus arm dropped from
the select set once observed closed (closed arms stay ready forever).
Heartbeat answered conn-side on topic 'phoenix'; rejoin replaces the
old generation with a phx_close to the old join_ref; lazy prune of dead
channel entries on first failed forward (the pubsub discipline).
As-landed deviations from the ratified sketch (module docs, veto by
diff): join takes &mut self (trait-object callable; construction moved
to ChannelFactory which still gets the full raw topic); reply(status,
payload) with the ref threaded invisibly (the sketch listed both);
broadcast gained the event parameter (unencodable without it); route()
returns Arc not Box. Outbound payloads encode from borrows (FrameRef)
so bus fan-out never clones P; payload-less control frames
(phx_close, error/heartbeat replies) encode as the codec's empty
payload rather than conjuring a P.
11 runtime-backed unit tests with a toy pipe codec (core provably
JSON-free). Suites: 78 w/ feature, 67 featureless, both green.
User-ratified 2026-06-12. Resolves the v0.6 fork left by the v0.5
handoff: dep #4 is serde/serde_json, but ONLY behind an opt-in
'phoenix' feature (V2 codec + JsonCodec). The 'channels' core takes
zero new deps — payload generic over urus Encode/Decode traits.
Channel trait shape, ChannelSocket, TopicRouter/PrefixRouter, and
opt-in ChannelSession persistence (zero-copy Arc<M> buffer drain)
all specified in the v0.6 section.
WsHandler::on_open(&mut self, sender) — defaulted, non-breaking,
added mid-chunk for veto-by-diff (the v0.3 "push" precedent for
defaults-level decisions): without it a listen-only ws client can
never be subscribed, since on_message never fires for a client that
doesn't send. Runs exactly once, first, in the conn actor, before any
buffered pipelined frame is decoded. Panic contract identical to
on_message: check_cancelled re-raise dance, else 1011 and no
on_close.
DISCOVERY 1 (documented, not fixed — inherent): the 101 reaches the
client BEFORE on_open runs, so "is my subscription live yet" is a
race client-side. App-level ack is the answer (any reply proves
on_open completed; callbacks are sequential). The integration tests
hit this immediately (first read of a join notice flaked into a 5s
read-timeout) and use a sync/synced ack; same reason Phoenix joins
reply.
DISCOVERY 2 (the structural one): PubSub::new() is in-runtime only,
but pipelines are built pre-runtime. crud's static-OnceLock bootstrap
DOES NOT COMPOSE with serve_with_shutdown here: a static pins the
table actor's last ServerRef forever, its inbox never closes, the
gen_server never exits, AllDone is unreachable — serve hangs (the
cross-thread-unpark limitation closes the workaround routes). The
pattern that works: NON-static Arc<OnceLock<PubSub<M>>> captured by
the route closure. Drain stops conns+listeners -> last Arc<Pipeline>
drops IN-runtime -> cell+handle drop -> inbox closes -> table exits.
Corollary: relays hold the Receiver only, NEVER a PubSub clone
(relay holds table's inbox open, table holds relay's receiver open:
mutual keepalive, shutdown hangs). Both rules in module docs, README,
and enforced by the new shutdown_with_open_chat_terminates test:
two open chat sockets + live relays + live table, shutdown must
return within 5s.
examples/ws_chat.rs: rooms as topics (room:{name}), on_open
subscribes (conn-actor pid: the monitor scopes cleanup to the
session) + spawns the relay (rx -> WsSender clone; exits on prune or
WsClosed), on_message broadcast_from(self_pid()) so the speaker is
never echoed, on_close broadcasts the leave notice and relies on the
monitor for unsubscribe.
Integration: ws_chat_broadcast_reaches_other_client_not_sender
(no-echo proven orderingly, no timeout reads: B speaks after A, A's
next frame must be B's) + shutdown_with_open_chat_terminates.
hammer.sh default subset now includes ws_ (v0.4 ran it ad hoc; ws IS
conn-lifecycle). Hammered: 35x subset (incl. both new tests) + 3x
full + 1x full under smarm-trace, all green. dbg!-grep clean. smarm
PRISTINE. README pub/sub + on_open sections; ROADMAP v0.5 as-landed.
Suite: 67 unit + 42 integration + 2 doc.
urus::pubsub, phoenix_pubsub-shaped, local node only. Independent of
HTTP (imports smarm only); crate-extraction candidate.
Design as ratified (handoff Q1-Q5, user said "push" on the leans):
- Q1 generic: PubSub<M> per instance, zero-cost, no downcasts.
Heterogeneous events = app-side enum M. v0.6 channels build on it.
- Q2 Receiver: subscribe(topic) -> Receiver<Arc<M>>, fresh channel per
subscription, subscriber owns its loop. Fan-in via a Sender-passing
subscribe_with is a compatible later addition, deliberately not v1.
subscribe_as(pid, topic) pins the subscription lifetime to an
explicit pid for relay patterns (session actor subscribes, spawned
relay receives) — without it the monitor would watch the wrong actor.
- Q3 unbounded: broadcast never blocks the table (bounded-block =
head-of-line across ALL topics; bounded-drop = silent loss). Memory
risk documented; TWO cleanup paths: monitor Down on subscriber death
(eager) + prune-on-send-failure for dropped receivers (lazy, on next
broadcast). Bench before sharding.
- Q4 user-owned ServerRef wrapper. DISCOVERY: the ratified optional
register-by-name helper is NOT implementable against smarm's
registry — it maps name -> Pid, and a Pid cannot be turned back into
a ServerRef (the ref is the inbox sender). Needs smarm support or a
global type-erased map; deferred, noted in module docs. pid()
exposed for introspection.
- Q5 unique per (pid, topic): HashMap<Topic, HashMap<Pid, Sender>>,
subscribe idempotent (replaces sender; stale receiver's channel
closes). One monitor per live pid (monitored: HashSet<Pid>, retired
in handle_down so slot-reuse pids re-monitor). handle_down does the
full-scan cleanup as ratified; pid->topics reverse index is the
documented optimization if deaths ever measure hot.
Subscribe is a call (table updated before return: a broadcast issued
right after by the same caller is seen); unsubscribe/broadcast are
casts. subscriber_count(topic) added as a call — needed by the tests
to observe async cleanup, legitimate API anyway.
8 runtime-backed unit tests (smarm::run + collect-outside-assert-after
pattern from smarm's own gen_server tests), including: Arc payload
ptr-equality across subscribers, monitor-path pruning with NO
broadcast issued (isolates it from the lazy path), broadcast_from
self-skip, resubscribe replacement.
Suite: 67 unit + 40 integration + 2 doc, green. smarm PRISTINE.
Topology (b) as agreed: ONE actor per connection via smarm RFC 008 fd
arms. After the 101, ws::duplex::run_duplex replaces the HTTP loop:
try_select(&[&outbound_rx, &FdArm::readable(fd)]) per iteration,
outbound at index 0 (priority order = owed writes drain before reads;
honest backpressure). Accepted gap documented: mid-write of a large
outbound frame the actor isn't reading.
Handler API (deferred questions, as answered this session):
- WsHandler { on_message, on_close } runs IN the conn actor's select
loop; concurrency = spawn an actor with a WsSender clone (the SSE
producer pattern). Conn::upgrade(self, handler) flat signature;
WsUpgrade grows from marker to Box<dyn WsHandler> payload.
- WsSender: Clone; send/text/binary/ping/close -> Err(WsClosed).
Unbounded channel like SSE; slow client bounded by write_timeout.
- Control frames invisible v1: auto-pong inline (5.5.2), peer close
auto-echoed (5.5.1; server closes TCP first per 7.1.1) surfacing
only as on_close(Some(code), reason); wordless endings (EOF, write
failure, drain timeout) = on_close(None, ""). on_ping/on_pong can
land later as default methods, non-breaking.
- Server-initiated close (WsSender::close, FrameError->1002/1009/1007,
handler panic->1011): close out, bounded drain (one write_timeout
budget) for the peer echo, data discarded (1.4). Handler panics
re-raise smarm's stop sentinel first (the pipeline catch_unwind
dance, replicated).
- Caps: Config.max_frame_payload (1 MiB) / max_message_bytes (4 MiB).
Conn actor: 101 branch now hands fd + leftover buf (pipelined first
frame carries over; tested) + boxed handler to run_duplex; registry
entry stays Busy for the ws lifetime, so graceful shutdown force-stops
the conn out of the select park at the drain deadline (tested — the
in-runtime request_stop wake covers select parks too).
Tests: chunk-1 EOF test rewritten into a 9-test duplex suite (echo,
pipelined first frame, ping/pong, both close directions, 1002 unmasked,
1009 header-cap, fragmentation with interleaved ping, shutdown
force-stop). Suite 59u+40i+2d. Hammer: 35x lifecycle+ws subset + 3 full
+ 1 full under smarm-trace, all green; no AlreadyExists out of
try_select (the fresh eager-cleanup path held). Validated against
python websocket-client (echo, pong payload, close 1000).
examples/ws_echo.rs: /echo in-actor + /clock producer-spawning.
1.0 defaults to close: honoring 'Connection: Keep-Alive' without echoing
it back left spec-following clients waiting for an EOF that never came
(ab -k deadlocked to its poll timeout on response 1).
As agreed:
- serialise_response echoes 'connection: keep-alive' when keep_alive
is on AND version is 1.0. 1.1 keep-alive stays implicit.
- Error statuses (>=400) on 1.0 force keep_alive off in the conn actor
(joins the existing 1.0-stream force-close): we don't advertise reuse
on a 4xx/5xx to a protocol generation where reuse is opt-in.
- HEAD needs nothing: the echo is framing-neutral.
- Parse-error responses already carried 'connection: close'.
Tests: unit pair (1.0 echo / 1.1 implicit) + integration pair (socket
actually reused on 1.0; 404 forces close + EOF). Validated against the
original repro: ab -k -n 5000 -c 50 GET /users -> 5000/5000 keep-alive,
0 failed, ~63k rps (~18.9k without keep-alive).
Default run encodes the pre-commit ritual: 35x the lifecycle-sensitive
integration subset (shutdown timeout reaped slowloris streaming chunked
sse stalled), fail-fast, then 3x the full suite. -n/-f override counts,
'-- <words>' overrides the filter. Failing run's output lands in
target/hammer-fail.log.
Found benching with ab -k: a 1.0 request with Connection: Keep-Alive is
kept alive (second pipelined request answered on the same socket) but
the response never echoes the header. 1.0 clients only reuse on explicit
server opt-in, so they wait for EOF that never comes; ab -k deadlocks to
apr_pollset_poll timeout on the first response. Recorded with repro,
fix direction (echo header in serialise_response on live 1.0 conns, or
force-close 1.0), and the no-keep-alive baseline: GET /users, release,
loopback, logging silenced, c=50 -> ~18.9k req/s, p50 3ms / p99 5ms,
0 failed. Fix deliberately not started: scope (error responses, HEAD)
is a user call.
Pure module (src/ws/frame.rs): bytes in, frames out, zero io/actor
machinery, so the whole RFC 6455 §5 edge surface lives in fast unit
tests ahead of the chunk-3 conn-lifecycle wiring.
- decode(buf, require_masked, max_payload) -> Ok(Some((frame,
consumed))) | Ok(None: incomplete) | Err(fatal). Server mode requires
client masking (§5.1); the flag keeps the codec reusable for a client
mode. Enforced: RSV=0, reserved opcodes, control frames FIN+<=125,
MINIMAL length encodings (16/64-bit), 64-bit MSB clear. The payload
cap is checked from the header BEFORE the payload is buffered — a
hostile 8-byte length can't make the conn actor allocate toward it.
- Frame::encode is unmask-only (server MUST NOT mask); encode_masked
exists for the client side of tests + future client mode.
- Close payloads: parse_close_payload validates the §7.4 wire-code
ranges (1004/1005/1006/1015 and <1000/1012-2999/>=5000 rejected,
3000-4999 app space allowed), 1-byte payload rejected, UTF-8 reason
required; close_payload(code, reason) truncates the reason to the
125-byte control cap on a char boundary.
- Assembler (data frames only; control frames interleave at the caller
per §5.4): orphan continuation and mid-fragmentation data frames are
protocol errors, message cap spans fragments (and the unfragmented
fast path), text UTF-8 validated once on completion (whole-message
delivery makes incremental validation pointless), reusable after
every message/error.
- FrameError -> close_code mapping: Protocol=1002, TooLarge=1009,
BadUtf8=1007.
Tests pin all five §5.7 worked examples (masked Hello bytes verbatim),
round-trip the 125/126 and 65535/65536 boundaries, and prove decode
returns None at EVERY proper prefix of a frame. 57 unit tests total.
Design (as agreed in the v0.4 pass):
- Dep #3 spent: sha1_smol 1.0.1 (zero transitive deps) instead of a
vendored SHA-1 — user call: no hand-rolled crypto. Consequence noted
in ROADMAP: the v0.6 wire-format JSON question now costs dep #4.
- Base64 stays in-tree but ENCODE-ONLY (RFC 4648 vectors tested): it is
an encoding, not crypto, and the decode direction (where parsing bugs
live) is deliberately not implemented.
- src/ws/{mod,handshake}.rs: validate() implements §4.2.1 (GET, 1.1,
Host, upgrade/connection token lists case-insensitively, key shape =
base64 of exactly 16 bytes, version 13). Version mismatch -> 426 +
sec-websocket-version: 13 (§4.2.2); everything else -> 400. Rejected
upgrades stay plain HTTP with keep-alive intact (tested: 400 then 101
on the same connection).
- Conn grows pub upgrade: Option<WsUpgrade> (opaque marker; chunk 3
turns it into the duplex/handler handoff payload — shape deliberately
uncommitted while the handler API is open). Conn::upgrade(self) is the
commit point: 101 + accept key + marker + halt, or the rejection.
- conn_actor: upgrade marker + status 101 -> write head, leave the HTTP
loop. Until chunk 3 that drops the fd (clients see 101 then EOF);
status check is defensive against a post-handler plug clobbering 101.
- serialise_response: 1xx no longer get content-length injected
(RFC 7230 §3.3.2). 204 left alone on purpose — separate conversation.
Accept-key path pinned to the §1.3 worked example end-to-end (unit +
wire). Suite: 34 unit + 33 integration + 2 doc.
README: write_timeout in the Config sample, read/write timeout-split
semantics spelled out, new Streaming Responses and Server-Sent Events
sections (producer contract, 1.0 fallback, heartbeat liveness), roadmap
summary bumped. ROADMAP: v0.3 marked done with the as-landed decisions
(pull shape per the user's fork answer; Q3 timeout split as proposed)
and verified exit criteria. crud example: GET /ticker SSE route — one
event/second, producer exits on SseClosed or the example's shutdown
flag (so an open ticker doesn't block AllDone past the drain).
New src/sse.rs. Conn::sse() (and sse_with_heartbeat) turns the response
into a text/event-stream: status 200 if unset, content-type +
cache-control: no-cache, and a RespBody::Stream whose heartbeat is wired
to the new StreamBody.heartbeat field. Returns (Conn, EventSender);
EventSender is Clone (stream ends when the LAST clone drops) with
send(event, data) / data(data) / comment(text), all returning
Err(SseClosed) once the conn actor has dropped the stream — the producer
exit signal. Multi-line data becomes one data: line per line (pure
format_event, unit-tested).
Heartbeat lives in the pump, as the roadmap leaned: with
StreamBody.heartbeat set, pump_stream waits recv_timeout(interval) and on
expiry writes the payload (": keep-alive" comment chunk) and keeps
waiting. Liveness inversion is deliberate: no request clock governs an
SSE response (request_timeout is read-phase only since chunk 1); a dead
client is detected when an event or heartbeat write stalls past
write_timeout, and a draining shutdown force-stops the recv park like any
stream.
Tests: 4 unit (3 framing, 1 sse() wiring) + 2 integration (event round
trip over decoded chunked stream incl. named + unnamed events and clean
0-chunk EOS on sender drop; >=2 heartbeats observed across 450ms of
producer silence at a 100ms interval).
parser: Transfer-Encoding: chunked no longer 411s — it sets
ParsedHead.chunked and the conn actor decodes. Two hard rejections at
parse time, both Malformed/400: chunked together with Content-Length
(the request-smuggling ambiguity; RFC 7230 §3.3.3 permits rejection)
and chunked on HTTP/1.0 (TE is a 1.1 mechanism). ParseError::Unsupported
is now unconstructed; kept for future framings.
conn_actor: read_chunked_body decodes incrementally from buf[head_len..],
reading on the SAME request deadline the head came in under (a stalled
chunked body dies at request_timeout exactly like a stalled CL body).
Chunk extensions are ignored; trailers are consumed and discarded.
Bounds: decoded size capped at max_body_bytes -> 413 the moment the cap
would be crossed (the CL pre-check can't see a chunked body's size up
front); size lines capped at 128 bytes and the trailer section at 8 KiB
-> 400, so framing spam can't grow the read buffer unboundedly.
Returns (decoded, raw_consumed_past_head): the keep-alive drain at the
loop bottom now drops head_len + RAW framing length (not decoded length),
so pipelined requests behind a chunked body land exactly — covered by
the trailers+pipelining test.
Tests: 3 parser unit (flagging, CL+TE reject, 1.0 reject; the old
chunked-is-Unsupported test replaced) + 6 integration (decode with
extensions, trailers + pipelined next request, 413 over decoded cap,
400 malformed size line, 400 CL+TE on the wire, stalled chunked body
killed at request_timeout with silent close).
Design decisions (per handoff Q1 user answer + Q3 proposal):
- Producer shape is PULL: the handler returns a smarm Receiver<Vec<u8>>
(RespBody::Stream / StreamBody, From<Receiver<Vec<u8>>> for ergonomics);
the producer is an actor the handler spawned. The conn actor pumps in
pump_stream(): it keeps sole ownership of the socket and of write
deadlines, and the recv() park between chunks is stoppable by the
draining registry's request_stop (stop sentinel unwinds out of
park_current; fd + registry guards clean up) — so an infinite stream is
force-stoppable at the drain deadline like any in-flight request, with
no polling. End of stream = every Sender dropped = terminating 0-chunk.
Producer contract: a send() error means the conn died; exit.
- Framing: HTTP/1.1 gets transfer-encoding: chunked and the connection
stays reusable after the terminator (keep-alive after chunked). HTTP/1.0
has no chunked TE: bytes go raw, keep-alive is forced off, EOF delimits.
User-set content-length/transfer-encoding headers are DROPPED for stream
bodies — we own the framing, and CL+chunked is a smuggling vector.
Empty producer chunks are skipped (a 0-length chunk would terminate the
framing early).
- request_timeout stays READ-phase only (unchanged). NEW Config/ConnLimits
field write_timeout (default 30s) gives every response write a per-write
budget: the fixed head+body write, each streamed chunk, and the error
paths (413/100-continue/4xx) all go through the now deadline-bounded
write_all (wait_writable_timeout, mirroring read_some). Each chunk gets
a FRESH budget — streams may outlive any whole-response clock; a single
stalled write may not (write-side slowloris). Naming/default were
flagged as a user call in the handoff: veto here if write_timeout(30s)
isn't it.
- RespBody loses derive(Clone) (Receiver isn't Clone; Clone was unused)
and gets a manual Debug.
Tests: 3 serialiser unit tests (chunked head shape, user-framing-header
stripping, 1.0 fallback) + 5 integration (chunked round-trip with decoder,
keep-alive after chunked, 1.0 EOF-delimited, shutdown force-stops an
infinite stream at drain deadline, stalled reader killed at write_timeout
with producer observing the closed channel).
Registration:
- Each listener registers "urus.listener.{i}" from inside its ChildSpec
factory. On a restart the stale binding points at a dead pid; smarm's
registry evicts those lazily on register/whereis, so re-registering
the same name is safe. Result ignored — a registry hiccup must not
take the listener down.
- "urus.server" -> supervisor pid, registered via the JoinHandle's
.pid() immediately after spawn rather than inside the closure: the
binding exists before the supervisor runs an instruction, so an early
whereis can't observe None.
Test: server_and_listeners_are_registered probes whereis from a route
handler — whereis must run inside the runtime, and the test thread is a
foreign OS thread with no runtime in its TLS.
Docs:
- README: Config documented in full (timeout semantics: keep_alive =
idle-before-first-byte, request = Instant deadline first-byte ->
head+body, slowloris-proof), Handle/shutdown_handle/serve_with_shutdown
section with the 4-step shutdown sequence, named-actors note, test
coverage list refreshed, Roadmap section now points at ROADMAP.md.
- ROADMAP: v0.2 chunks 1-4 marked landed; design detail stays in the
chunk commit bodies.
Verified: cargo build --features smarm-trace clean; full suite (32
tests) green 3x; debug-grep clean.
This closes v0.2. Exit criteria all verified across chunks 1-3: crud
drains on stdin-Enter (~100ms), idle keep-alive reaped at
keep_alive_timeout, panicking listener restarts under load.
Two wall-clock budgets in the conn read path, both via
smarm::wait_readable_timeout:
- keep_alive_timeout: idle budget between requests — how long we park
waiting for the FIRST byte of a request (including the first request on
a fresh connection). Expiry closes silently; nothing is owed to a
client that isn't talking.
- request_timeout: per-request budget as an Instant deadline from the
first byte of a request until the request (head + body) is fully read.
Pipelined leftovers in the buffer at cycle start count as a started
request. The deadline is returned from read_head and threaded into
read_body so head and body share one budget. Pipeline run time is NOT
covered. Each read wait uses whichever budget is currently active;
deadlines are Instants, so EAGAIN retries don't reset the clock.
Expiry mid-head gets a best-effort 408 via try_write_once: a single
non-parking write syscall — a client that stalls its read side must not
defeat the timeout by making the 408 write park forever. Expiry mid-body
just closes.
examples/crud.rs gains stdin-Enter graceful shutdown (plain OS thread on
read_line -> handle.shutdown(); no signal crate). Doing so surfaced a
pattern worth knowing: crud's lazily-spawned store actor parked forever
in recv(), which blocks smarm's AllDone, so serve_with_shutdown never
returned (gdb: scheduler idle in poll_wake, store the only live actor).
It can't be messaged awake from the stdin thread either — a cross-thread
send's unpark is a no-op without runtime TLS, the same limitation behind
SHUTDOWN_POLL. Fix: store_loop recv_timeout(250ms) + a SHUTTING_DOWN
AtomicBool set by the stdin thread before handle.shutdown(). This poll
dies with the cross-thread-unpark limitation; recorded in smarm docs.
Tests: idle keep-alive conn reaped at a small keep_alive_timeout;
slowloris partial-head stall killed at request_timeout with the 408
observed. Timeout+shutdown subset hammered 30x clean; full suite 3x;
smarm-trace build and suite clean.
Shutdown is drain-then-force-stop: flag.store(true) -> sup_h.join() (the
no-more-accepts barrier) -> cast BeginDrain -> poll ConnCount every 50ms
until 0 or drain_timeout; past the deadline each tick casts ForceStopConns
and re-sweeps, so conns that registered after the deadline are still
caught.
Listener shutdown is redesigned vs chunk 1: Restart::Permanent ->
Transient, wait-failure return -> panic (Transient restarts it), and a
shared Arc<AtomicBool> flag checked at loop top with a 250ms
wait_readable_timeout tick instead of request_stop — no request_stop on
the supervisor or listeners anywhere. Historical note: the redesign was
originally forced by smarm's then-lossy stop against QUEUED actors (fixed
upstream in 7bab4d2); it is kept because flag-based listener shutdown is
simpler and stop-semantics-free.
Conns self-register with the registry (ConnStarted from the conn actor,
not a listener cast): per-sender FIFO ordering is the only thing that
prevents ConnEnded overtaking ConnStarted; see conn_registry module docs.
conn_actor: in the catch_unwind(pipeline.run) Err branch,
smarm::preempt::check_cancelled() runs BEFORE composing the 500 — smarm's
stop is an undowncastable panic_any(StopSentinel), so the catch_unwind
would otherwise swallow a stop and 500-and-keep-running. The stop flag is
persistent, so check_cancelled re-raises cleanly.
Handle.shutdown() wakes the root via a 100ms try_recv poll
(SHUTDOWN_POLL): a cross-thread Sender::send's unpark is a
try_with_runtime no-op without runtime TLS on the sending thread (still
true on smarm 8e5b754; cross-thread unpark is a recorded smarm roadmap
candidate — this poll dies with it).
Two smarm bugs were found during this chunk and fixed upstream: lossy
stop against QUEUED actors (7bab4d2) and the terminal-wake shutdown
stall/hang (eddf3fe); post-mortem in artefact smarm-bug-terminal-wake.md.
OneForOne supervisor on the root actor, Restart::Permanent per listener.
ChildSpec factory owns its fd via Arc<OwnedFd> (OwnedFd gains Sync);
restarts reuse the fd — no re-dup, no fd-less window. wait_readable
failure now exits-and-restarts instead of being fatal. Conn actors stay
unsupervised bare spawns. Test hook INJECT_LISTENER_PANICS panics before
accept so a pending connection must survive into the restarted listener.
Roadmap: chunk 1 marked done; chunk 3 fork resolved to smarm-native
timed waits (RFC 008 landed at 393cdd0, reaper option removed); chunk 2
drain-then-stop decided; v0.4-3(b) unblocked.
urus builds clean and passes 24/24 against smarm master; the refresh is
OTP adoption, not breakage. v0.2 chunks: supervised listener pool,
graceful shutdown via request_stop, enforce the (currently decorative)
timeouts, registry naming. Then: streaming bodies + SSE (v0.3),
WebSocket upgrade (v0.4), PubSub (v0.5), Channels (v0.6). HTTP/2 stays
iceboxed per spec §2.2. Two decision forks flagged inline: timeout
mechanism (reaper vs smarm timed fd waits) and ws duplex topology
(dup'd fds vs fd-arms-in-select) — both hinge on smarm RFC 008.
Claude