zenkey_fleet/bus/query.rs
1//! The fan-in query discipline (RFC 05 §2.1) — moved verbatim from
2//! zenctl's `bus.rs`; this stays the single chokepoint for fleet GETs.
3
4use std::time::Duration;
5
6use crate::{Error, Result};
7use zenkey::{RegistrySlice, parse_slice};
8use zenoh::Session;
9use zenoh::qos::Priority;
10use zenoh::query::{ConsolidationMode, QueryTarget};
11
12use crate::bus::session::Fleet;
13use crate::report::ValueSource;
14
15/// How a producer answered a procedure call.
16///
17/// `Clone` is a refcount bump on the payload, not a copy — which is what lets
18/// a GUI hold an answer in widget state without paying for it.
19#[derive(Debug, Clone)]
20pub enum Answer {
21 /// A value reply — RFC 05 §3: "a reply always indicates success".
22 /// Carried as zenoh's refcounted buffer: cloning is a refcount bump,
23 /// and consumers decode via `reader()`/`to_bytes()` (a `Cow` — it
24 /// copies only when the payload arrived fragmented). Report §14's
25 /// zero-copy discipline: the old `to_bytes().to_vec()` double copy per
26 /// reply is retired.
27 Value(zenoh::bytes::ZBytes),
28 /// An error reply (`reply_err`), carrying the `{error, message}` envelope
29 /// when it parses. RFC 05 §3: "an error always indicates failure".
30 Error { name: String, message: String },
31}
32
33/// One host's answer, attributed to the origin that actually replied.
34#[derive(Debug, Clone)]
35pub struct FleetAnswer {
36 pub origin: String,
37 /// The reply's **own** key expression — what `origin` was derived from, and
38 /// the concrete key a follow-up must be addressed to.
39 ///
40 /// Empty for an error reply, which zenoh gives no sample and therefore no
41 /// key. Carried because `origin` is lossy by design: the attribution helper goes
42 /// through the grammar and yields `"?"` for any key that does not parse
43 /// under `base`, and a caller that must still *name* the responder (RFC 09
44 /// §5.1 O1 — a non-conforming key is a fact) has nowhere else to look.
45 pub key: String,
46 /// The reply's declared encoding, when it carried one.
47 ///
48 /// A caller that speaks a specific wire (`@blob`'s postcard replies, say)
49 /// needs to tell "answered in a dialect we do not speak" from "did not
50 /// answer": the first is an observation, the second is silence, and RFC 09
51 /// §5.1 O4 forbids rendering them alike.
52 pub encoding: Option<String>,
53 /// The reply's attachment, when it carried one (refcounted, like the
54 /// payload). `None` on an error reply is the only truth available:
55 /// zenoh's `ReplyError` carries no attachment — a fact about the wire,
56 /// not an unobserved field.
57 pub attachment: Option<zenoh::bytes::ZBytes>,
58 pub answer: Answer,
59}
60
61/// What one GET may vary — everything RFC 05 §2.1 does **not** fix.
62///
63/// The §2.1 triple is not a knob and deliberately has no field here: it is
64/// applied by `disciplined_get` to every GET this crate issues. What a
65/// caller does choose is the timeout, the request body, an attachment riding
66/// beside it (#126), the query's priority (RFC 04 §3, RFC 07 §2.6) and
67/// whether replies from *outside* the selector are accepted.
68///
69/// A spec struct rather than named sibling functions: `fleet_get_at`
70/// (priority) and `fleet_get_call` (attachment) used to be those siblings, and
71/// `_at` had come to mean two things — this axis, and the injected-clock
72/// convention (`ingest_at`, `ZrecWriter::new_at`). The greps the siblings
73/// bought survive as setter greps: "who issues bulk GETs?" is
74/// `grep '\.priority('`, "who sends attachments on queries?" is
75/// `grep '\.attachment('`.
76#[derive(Debug, Clone)]
77pub struct GetOpts {
78 timeout: Duration,
79 payload: Option<Vec<u8>>,
80 attachment: Option<Vec<u8>>,
81 priority: Priority,
82 accept_any: bool,
83 max_replies: usize,
84 /// What the bound cost, filled in by the GET (#339). Shared rather than
85 /// returned — see [`GetOpts::elided`].
86 elided: std::sync::Arc<std::sync::atomic::AtomicU64>,
87}
88
89/// How many replies a GET keeps unless the caller says otherwise (#339).
90///
91/// Every fan-out here was unbounded: `collect_answers`, `fetch_timed` and
92/// `admin_get` pushed every reply into a `Vec`, each holding a refcounted
93/// payload, so a `**` sweep against a router with a large storage was
94/// unbounded memory in a tool that bounds everything else it accumulates.
95///
96/// 4096 is chosen against what the fan-out *means*: a fleet GET is one reply
97/// per producer per key, and a fleet with four thousand replying entities on
98/// one selector is past what any of these renderers show anyway. A caller
99/// that genuinely wants more says so, and hears what the last bound cost.
100pub const DEFAULT_MAX_REPLIES: usize = 4096;
101
102impl GetOpts {
103 /// A plain GET, bounded by `timeout`.
104 ///
105 /// [`Priority::DEFAULT`] is `Priority::Data` — byte-identical to setting
106 /// no priority at all, which is what every un-annotated GET did before
107 /// this type existed.
108 pub fn new(timeout: Duration) -> Self {
109 GetOpts {
110 timeout,
111 payload: None,
112 attachment: None,
113 priority: Priority::DEFAULT,
114 accept_any: false,
115 max_replies: DEFAULT_MAX_REPLIES,
116 elided: std::sync::Arc::new(std::sync::atomic::AtomicU64::new(0)),
117 }
118 }
119
120 /// The request body, when there is one. `None` is the common case and
121 /// costs nothing to say.
122 pub fn payload(mut self, payload: Option<Vec<u8>>) -> Self {
123 self.payload = payload;
124 self
125 }
126
127 /// A query attachment (#126), verbatim — never schema-encoded. The encode
128 /// ladder is for bodies; an attachment is outside the registry's
129 /// vocabulary (#117), on a query exactly as on a publish.
130 pub fn attachment(mut self, attachment: Option<Vec<u8>>) -> Self {
131 self.attachment = attachment;
132 self
133 }
134
135 /// State the query's priority (RFC 04 §3, RFC 07 §2.6).
136 ///
137 /// Replies inherit the *query's* QoS — a server-side setter is a no-op —
138 /// so a bulk plane's priority can only be decided here. RFC 07 §2.6 makes
139 /// that a caller obligation rather than a suggestion: `@blob` GETs MUST
140 /// ride at [`Priority::DataLow`], or one operator fetching a debug bundle
141 /// starves the telemetry and alerts sharing the link.
142 pub fn priority(mut self, priority: Priority) -> Self {
143 self.priority = priority;
144 self
145 }
146
147 /// Accept replies on keys **outside** the selector
148 /// ([`zenoh::query::ReplyKeyExpr::Any`]) — the querying-subscriber
149 /// pattern.
150 ///
151 /// The `@adv` cache replies with the cached sample on the *sample's* own
152 /// key, outside a `<key>/@adv/**` selector, and zenoh drops such replies
153 /// unless the caller opts in. Harmless on a rung whose replies sit inside
154 /// the selector anyway.
155 pub fn accept_any(mut self) -> Self {
156 self.accept_any = true;
157 self
158 }
159
160 /// The bound this GET runs under.
161 pub fn timeout(&self) -> Duration {
162 self.timeout
163 }
164
165 /// Keep at most `max` replies (#339). Zero is clamped to one: a GET that
166 /// kept nothing would report silence, and silence is never a verdict
167 /// (RFC 05 §3.1).
168 pub fn max_replies(mut self, max: usize) -> Self {
169 self.max_replies = max.max(1);
170 self
171 }
172
173 /// The reply bound in force.
174 pub fn reply_bound(&self) -> usize {
175 self.max_replies
176 }
177
178 /// **What the bound cost**: replies that arrived and were not kept,
179 /// across every GET run under these options (RFC 13 §3 O6 — a bound that
180 /// hides data must say so).
181 ///
182 /// It rides here, on the object that *states* the bound, rather than in
183 /// the return type, for the reason every other bounded structure in this
184 /// crate keeps its own ledger (`StatsTable::evicted`,
185 /// `Retention::evicted`, `BoundedLru::admit`): the thing that owns the
186 /// ceiling owns the count of what the ceiling refused. A caller reads it
187 /// beside the answers it just got:
188 ///
189 /// ```ignore
190 /// let opts = GetOpts::new(timeout);
191 /// let answers = fleet_get(&fleet, key, &opts).await?;
192 /// if opts.elided() > 0 { /* say so — never render this as "all of them" */ }
193 /// ```
194 ///
195 /// The count is exact: past the bound the replies are still drained, they
196 /// are simply not kept. Draining is what makes the number honest; *keeping*
197 /// is what was unbounded.
198 pub fn elided(&self) -> u64 {
199 self.elided.load(std::sync::atomic::Ordering::Relaxed)
200 }
201
202 /// Forget what earlier GETs under these options cost — for a caller that
203 /// reuses one `GetOpts` and reports per GET rather than per run.
204 pub fn reset_elided(&self) {
205 self.elided.store(0, std::sync::atomic::Ordering::Relaxed);
206 }
207
208 /// Add to the ledger — for the drains that live in another module
209 /// ([`crate::admin_get_within`]) and keep their own reply shape.
210 pub(crate) fn note_elided(&self, n: u64) {
211 if n > 0 {
212 self.elided
213 .fetch_add(n, std::sync::atomic::Ordering::Relaxed);
214 }
215 }
216}
217
218/// **The** `session.get` of this crate (RFC 05 §2.1) — no other module issues
219/// one, which is what makes the discipline checkable by grep rather than by
220/// review.
221///
222/// Two of the three things §2.1 requires are set here, once:
223///
224/// 1. **target = All.** The default `BestMatching` short-circuits to a single
225/// queryable the moment any matching one is declared `complete` — "one
226/// storage config away from silently collapsing the fleet to one reply".
227/// 2. **consolidation = None.** Default consolidation keeps one reply *per
228/// reply key*; belt-and-braces against a producer that wrongly echoes the
229/// wildcard selector instead of replying on its own concrete key.
230///
231/// The third — **attribution by the reply's own key**, never by the key we
232/// asked on — belongs to whoever drains the channel, and lives in
233/// `answer_of` for the [`FleetAnswer`] path.
234///
235/// The error is the middleware's own, unwrapped: every caller has a better
236/// sentence to wrap it in than this function does.
237pub(crate) async fn disciplined_get(
238 session: &Session,
239 selector: &str,
240 opts: &GetOpts,
241) -> Result<zenoh::handlers::FifoChannelHandler<zenoh::query::Reply>> {
242 let mut builder = session
243 .get(selector)
244 .target(QueryTarget::All)
245 .consolidation(ConsolidationMode::None)
246 .priority(opts.priority)
247 .timeout(opts.timeout);
248 if let Some(body) = opts.payload.clone() {
249 builder = builder.payload(body);
250 }
251 if let Some(att) = opts.attachment.clone() {
252 builder = builder.attachment(att);
253 }
254 if opts.accept_any {
255 builder = builder.accept_replies(zenoh::query::ReplyKeyExpr::Any);
256 }
257 builder.await.map_err(|e| Error::bus("get", "", e))
258}
259
260/// Call a procedure and collect **every** reply, attributed by origin.
261///
262/// The RFC 05 §2.1 fan-in, end to end: `disciplined_get` sets target `All`
263/// and consolidation `None`, and `answer_of` attributes each reply by the
264/// reply's *own* key — which is what makes `*`-origin fan-out legible.
265///
266/// Silence is deliberately *not* interpreted here (RFC 05 §3.1: "no reply" is
267/// not one condition). Callers that need a verdict join this against the
268/// liveliness roster; see `cmd::doctor`.
269/// Bounded at [`GetOpts::reply_bound`], and what the bound cost is on
270/// [`GetOpts::elided`] (#339).
271pub async fn fleet_get(fleet: &Fleet<'_>, key: &str, opts: &GetOpts) -> Result<Vec<FleetAnswer>> {
272 let replies = disciplined_get(fleet.session(), key, opts)
273 .await
274 .map_err(|e| Error::bus("query", key.to_string(), e))?;
275 let (answers, elided) = collect_answers(fleet.base(), replies, opts.max_replies).await;
276 opts.note_elided(elided);
277 Ok(answers)
278}
279
280/// Drain a reply channel into attributed answers — the shared back half of
281/// [`fleet_get`] and [`RepeatingQuery`]: one implementation of reply-key
282/// attribution and the RFC 05 §3 error envelope, however the query was issued.
283///
284/// Returns what it kept and **how many it did not** (#339). Past `max` the
285/// replies are still drained — the channel is being emptied either way — they
286/// are simply not retained, so the count is exact and the memory is bounded.
287/// The two are different facts: draining is the fan-in finishing, keeping is
288/// what used to be unbounded.
289async fn collect_answers(
290 base: &str,
291 replies: zenoh::handlers::FifoChannelHandler<zenoh::query::Reply>,
292 max: usize,
293) -> (Vec<FleetAnswer>, u64) {
294 let mut out = Vec::new();
295 let mut elided = 0u64;
296
297 while let Ok(reply) = replies.recv_async().await {
298 if out.len() >= max {
299 elided += 1;
300 continue;
301 }
302 out.push(answer_of(base, reply));
303 }
304 (out, elided)
305}
306
307/// One reply, attributed — the per-reply half of [`collect_answers`], shared
308/// with the timed drain in [`RepeatingQuery::fetch_timed`] so attribution and
309/// the RFC 05 §3 error envelope have exactly one implementation.
310fn answer_of(base: &str, reply: zenoh::query::Reply) -> FleetAnswer {
311 match reply.result() {
312 Ok(sample) => FleetAnswer {
313 origin: origin_of(base, sample.key_expr().as_str()),
314 key: sample.key_expr().as_str().to_string(),
315 encoding: Some(sample.encoding().to_string()),
316 attachment: sample.attachment().cloned(),
317 answer: Answer::Value(sample.payload().clone()),
318 },
319 Err(err) => {
320 // The error envelope is `{ "error": "<name>", "message": "…" }`
321 // (RFC 05 §3), with reserved names like `error/not-found`. If it
322 // does not parse we still surface the bytes — an unreadable
323 // refusal is still a refusal.
324 let bytes = err.payload().to_bytes();
325 let (name, message) = match serde_json::from_slice::<serde_json::Value>(&bytes) {
326 Ok(v) => (
327 v.get("error")
328 .and_then(|e| e.as_str())
329 .unwrap_or("error/unparsed")
330 .to_string(),
331 v.get("message")
332 .and_then(|m| m.as_str())
333 .unwrap_or_default()
334 .to_string(),
335 ),
336 Err(_) => (
337 "error/unparsed".to_string(),
338 String::from_utf8_lossy(&bytes).to_string(),
339 ),
340 };
341 // An error reply has no sample, so no concrete key to attribute
342 // by; zenoh does not surface the responder here.
343 FleetAnswer {
344 origin: "?".to_string(),
345 key: String::new(),
346 encoding: None,
347 attachment: None,
348 answer: Answer::Error { name, message },
349 }
350 }
351 }
352}
353
354/// A **declared** querier carrying the same RFC 05 §2.1 discipline as
355/// [`fleet_get`] (target `All`, consolidation `None`, attribution by reply
356/// key), for fetches that re-ask the **same key expression** — watch loops,
357/// the schema cache's re-asks, registry sweeps, doctor. Declaring once lets
358/// the network keep routing state warm instead of rebuilding it per GET
359/// (report §12's zenoh-1.9 adoption row).
360///
361/// When to use which:
362/// - recurring, same keyexpr → declare a `RepeatingQuery` and `fetch` many
363/// times (parameters and payload ride **per get**, never in the declared
364/// keyexpr — a `?params` suffix in `key` is a bug here);
365/// - genuinely one-shot, or an ad-hoc key → [`fleet_get`].
366///
367/// Liveliness sweeps ([`crate::bus::roster::roster()`]) are a different API
368/// (`session.liveliness().get()`) with no querier equivalent and stay
369/// undeclared.
370pub struct RepeatingQuery {
371 querier: zenoh::query::Querier<'static>,
372 base: String,
373 /// Replies kept per fetch, and what the bound has cost across all of them
374 /// (#339) — the same ledger [`GetOpts`] carries, for the declared path.
375 max_replies: usize,
376 elided: std::sync::atomic::AtomicU64,
377}
378
379/// Declare a repeating query on `key` (a full wire keyexpr, no `?params`).
380///
381/// The §2.1 discipline is fixed at declaration: target `All`, consolidation
382/// `None`, `timeout` for every subsequent fetch.
383pub async fn declare_repeating(
384 fleet: &Fleet<'_>,
385 key: &str,
386 timeout: Duration,
387) -> Result<RepeatingQuery> {
388 declare(fleet, key, timeout, false).await
389}
390
391/// As [`declare_repeating`], additionally accepting replies **outside** the
392/// declared keyexpr (`ReplyKeyExpr::Any`) — the querying-subscriber pattern
393/// the `@adv` cache rung needs. A separate constructor because this axis is
394/// part of the querier's identity: never reuse one querier across both modes.
395pub async fn declare_repeating_any(
396 fleet: &Fleet<'_>,
397 key: &str,
398 timeout: Duration,
399) -> Result<RepeatingQuery> {
400 declare(fleet, key, timeout, true).await
401}
402
403async fn declare(
404 fleet: &Fleet<'_>,
405 key: &str,
406 timeout: Duration,
407 accept_any: bool,
408) -> Result<RepeatingQuery> {
409 let mut builder = fleet
410 .session()
411 .declare_querier(key.to_string())
412 .target(QueryTarget::All)
413 .consolidation(ConsolidationMode::None)
414 .timeout(timeout);
415 if accept_any {
416 builder = builder.accept_replies(zenoh::query::ReplyKeyExpr::Any);
417 }
418 let querier = crate::bus::teardown::declared("declare querier", &key, builder).await?;
419 Ok(RepeatingQuery {
420 querier,
421 base: fleet.base().to_string(),
422 max_replies: DEFAULT_MAX_REPLIES,
423 elided: std::sync::atomic::AtomicU64::new(0),
424 })
425}
426
427impl RepeatingQuery {
428 /// The declared key expression.
429 pub fn key(&self) -> &str {
430 self.querier.key_expr().as_str()
431 }
432
433 /// One fetch on the declared keyexpr, every reply attributed by its own
434 /// key — [`fleet_get`]'s contract, minus the per-call declaration.
435 pub async fn fetch(&self) -> Result<Vec<FleetAnswer>> {
436 self.fetch_with("", None).await
437 }
438
439 /// As [`fetch`](Self::fetch), with selector parameters and/or a request
440 /// payload riding this one get.
441 pub async fn fetch_with(
442 &self,
443 params: &str,
444 payload: Option<Vec<u8>>,
445 ) -> Result<Vec<FleetAnswer>> {
446 let mut builder = self.querier.get();
447 if !params.is_empty() {
448 builder = builder.parameters(params);
449 }
450 if let Some(body) = payload {
451 builder = builder.payload(body);
452 }
453 let replies = builder
454 .await
455 .map_err(|e| Error::bus("query", self.key(), e))?;
456 let (answers, elided) = collect_answers(&self.base, replies, self.max_replies).await;
457 self.note_elided(elided);
458 Ok(answers)
459 }
460
461 /// Keep at most `max` replies per fetch (#339). Zero is clamped to one.
462 pub fn max_replies(mut self, max: usize) -> Self {
463 self.max_replies = max.max(1);
464 self
465 }
466
467 /// The reply bound in force.
468 pub fn reply_bound(&self) -> usize {
469 self.max_replies
470 }
471
472 /// Replies this querier's bound refused, across every fetch (RFC 13 §3
473 /// O6). See [`GetOpts::elided`] for why the count lives with the bound.
474 pub fn elided(&self) -> u64 {
475 self.elided.load(std::sync::atomic::Ordering::Relaxed)
476 }
477
478 fn note_elided(&self, n: u64) {
479 if n > 0 {
480 self.elided
481 .fetch_add(n, std::sync::atomic::Ordering::Relaxed);
482 }
483 }
484
485 /// As [`fetch`](Self::fetch), stamping each reply with how long after the
486 /// GET it arrived (issue #52).
487 ///
488 /// This exists because a fan-out call's *call* duration is the time until
489 /// the slowest answer, so attributing it to every origin would report a
490 /// fast responder's latency as the fleet's worst. Timing each reply where
491 /// it is drained is the only place the distinction is available — and it
492 /// keeps the RFC 05 §2.1 chokepoint intact rather than forking a second
493 /// GET path to measure with.
494 pub async fn fetch_timed(&self) -> Result<Vec<(FleetAnswer, Duration)>> {
495 let started = std::time::Instant::now();
496 let replies = self
497 .querier
498 .get()
499 .await
500 .map_err(|e| Error::bus("query", self.key(), e))?;
501 let mut out = Vec::new();
502 let mut elided = 0u64;
503 while let Ok(reply) = replies.recv_async().await {
504 let at = started.elapsed();
505 if out.len() >= self.max_replies {
506 elided += 1;
507 continue;
508 }
509 out.push((answer_of(&self.base, reply), at));
510 }
511 self.note_elided(elided);
512 Ok(out)
513 }
514
515 /// Undeclare, telling the network to drop the routing state. The crate's
516 /// idiom: teardown is explicit and awaited, never left to `Drop`.
517 pub async fn undeclare(self) -> Result<()> {
518 self.querier
519 .undeclare()
520 .await
521 .map_err(|e| Error::bus("undeclare querier", "", e))
522 }
523
524 /// Whether any queryable currently matches **this querier** — "someone
525 /// serves what *we* ask", a routing fact about the querier this process
526 /// declared (RFC 12 §9's allowed half). `false` is not a fleet verdict:
527 /// it never means "nobody serves this key" (RFC 05 §3.1).
528 pub async fn matching_status(&self) -> Result<bool> {
529 self.querier
530 .matching_status()
531 .await
532 .map(|s| s.matching())
533 .map_err(|e| Error::bus("matching status", "", e))
534 }
535
536 /// Event-driven matching changes for this querier — same honesty bounds
537 /// as [`matching_status`](Self::matching_status).
538 pub async fn matching_events(&self) -> Result<crate::bus::write::MatchingEvents> {
539 crate::bus::write::MatchingEvents::for_querier(&self.querier).await
540 }
541}
542
543/// The origin chunk of a wire key, via the grammar (never by index — RFC 03
544/// §1.1: positions are relative to the configured base).
545fn origin_of(base: &str, key: &str) -> String {
546 zenkey::grammar::parse_full(base, key)
547 .map(|k| k.origin.chunk().to_string())
548 .unwrap_or_else(|| "?".to_string())
549}
550
551/// Discover every live producer's registry slice **from the bus**, with nothing
552/// compiled in (RFC 08 §6: "generic explorer tooling … needs no compiled-in
553/// registry").
554///
555/// Every producer MUST serve its registry slice as TOML on
556/// `@rpc/<producer>/introspect`. This fans one wildcard-producer `introspect`
557/// GET across the fleet — `<base>/v1/*/@rpc/*/introspect` — and parses each
558/// reply. It is the same introspect+`parse_slice` path `doctor` walks, minus
559/// the compiled-in diff: here the served slice *is* the answer.
560///
561/// A reply that does not parse is reported to stderr and skipped, never fatal:
562/// one malformed producer must not blind the tool to every other producer's
563/// slice. The tuple's first element is the producer (or service) base name the
564/// slice declares (`slice.name`), matching the compiled path's producer column.
565///
566/// A verbatim service origin is unmatchable by the `*` of a fleet selector
567/// (grammar property D4), so the wildcard sweep cannot enumerate services.
568/// The well-known `@catalog` identity service (RFC 06 §5) is therefore asked
569/// by name, exactly as [`crate::bus::roster::roster()`] does for its alive token; other
570/// service origins remain reachable only via local registry files
571/// (`doctor --registry` asks each declared `service_origin` by name).
572pub async fn fleet_registry(
573 fleet: &Fleet<'_>,
574 timeout: Duration,
575) -> Result<Vec<(String, RegistrySlice)>> {
576 Ok(fleet_registry_raw(fleet, timeout)
577 .await?
578 .into_iter()
579 .map(|(slice, _)| (slice.name.clone(), slice))
580 .collect())
581}
582
583/// As [`fleet_registry`], additionally yielding each reply's raw TOML text
584/// (the artifact the slice cache persists).
585pub async fn fleet_registry_raw(
586 fleet: &Fleet<'_>,
587 timeout: Duration,
588) -> Result<Vec<(RegistrySlice, String)>> {
589 let repeating = RepeatingRegistry::declare(fleet, timeout).await?;
590
591 let slices = repeating.fetch().await?;
592
593 repeating.undeclare().await?;
594
595 Ok(slices)
596}
597
598/// The registry sweep as a **declared** pair of queriers (#37) — for callers
599/// that re-run the sweep (`--watch topic list`, doctor's second pass, a GUI
600/// refresh). One-shot callers keep [`fleet_registry`].
601///
602/// Two queriers, not one: the wildcard-producer fan-out plus `@catalog` by
603/// name (a `*` never matches a verbatim origin, D4 — the two cannot
604/// double-count; same reasoning as [`fleet_registry`]).
605pub struct RepeatingRegistry {
606 wildcard: RepeatingQuery,
607 catalog: RepeatingQuery,
608}
609
610impl RepeatingRegistry {
611 pub async fn declare(fleet: &Fleet<'_>, timeout: Duration) -> Result<Self> {
612 // This session is un-namespaced on purpose (RFC 09 §5), so it must
613 // spell the base itself — exactly as `service call` composes its key.
614 let wildcard = fleet.wire(zenkey::selector::rpc(
615 zenkey::selector::Scope::fleet(),
616 zenkey::selector::Producers::all(),
617 &["introspect"],
618 ));
619 let catalog = fleet.wire(zenkey::selector::service_rpc(
620 &zenkey::ServiceOrigin::catalog(),
621 &["introspect"],
622 ));
623 Ok(RepeatingRegistry {
624 wildcard: declare_repeating(fleet, &wildcard, timeout).await?,
625 catalog: declare_repeating(fleet, &catalog, timeout).await?,
626 })
627 }
628
629 /// One sweep: every parsed slice with its raw TOML. A reply that does not
630 /// parse is logged and skipped, never fatal — one malformed producer must
631 /// not blind the tool to every other producer's slice.
632 pub async fn fetch(&self) -> Result<Vec<(RegistrySlice, String)>> {
633 let mut slices = Vec::new();
634 for q in [&self.wildcard, &self.catalog] {
635 for answer in q.fetch().await? {
636 let Answer::Value(bytes) = answer.answer else {
637 continue;
638 };
639 let served_toml = String::from_utf8_lossy(&bytes.to_bytes()).to_string();
640 match parse_slice(&served_toml) {
641 Ok(slice) => slices.push((slice, served_toml)),
642 Err(e) => tracing::warn!(
643 origin = %answer.origin,
644 "introspect reply did not parse, skipping: {e}"
645 ),
646 }
647 }
648 }
649 Ok(slices)
650 }
651
652 /// Undeclare both queriers, acknowledged.
653 ///
654 /// Both, even when the first refuses (#346): the wildcard sweep and the
655 /// `@catalog` ask are one teardown, and leaving the second declared
656 /// because the first would not go is the half-torn-down state
657 /// [`crate::Monitor::shutdown`] refuses. Failures are reported together.
658 pub async fn undeclare(self) -> Result<()> {
659 crate::bus::teardown::drain_undeclare(
660 vec![
661 ("wildcard introspect".to_string(), self.wildcard),
662 ("@catalog introspect".to_string(), self.catalog),
663 ],
664 RepeatingQuery::undeclare,
665 )
666 .await
667 }
668}
669
670/// One state sample from a snapshot GET.
671#[derive(Debug, Clone)]
672pub struct StateSample {
673 /// Full wire key.
674 pub key: String,
675 /// HLC timestamp, when the deployment stamps samples (RFC 04 §4
676 /// requires it for LWW to be meaningful — its absence is itself a
677 /// doctor-grade observation).
678 pub timestamp: Option<zenoh::time::Timestamp>,
679 pub payload_len: usize,
680}
681
682/// GET the current state under a selector with the fan-in discipline
683/// (target All, consolidation None) — the doctor's freshness check
684/// (RFC 04 §1.2) consumes the timestamps. Same chokepoint posture as
685/// [`fleet_get`]: no subcommand issues a raw `session.get`.
686///
687/// `max` bounds the samples **drained** (`doctor --sample N`): the loop
688/// stops reading at the cap, so a bounded sweep is cheaper, not merely
689/// quieter. `None` drains every reply.
690pub async fn state_snapshot(
691 session: &Session,
692 selector: &str,
693 timeout: Duration,
694 max: Option<usize>,
695) -> Result<Vec<StateSample>> {
696 let replies = disciplined_get(session, selector, &GetOpts::new(timeout))
697 .await
698 .map_err(|e| Error::bus("state snapshot", selector, e))?;
699 let mut out = Vec::new();
700 while let Ok(reply) = replies.recv_async().await {
701 if max.is_some_and(|m| out.len() >= m) {
702 break;
703 }
704 let Ok(sample) = reply.result() else { continue };
705 out.push(StateSample {
706 key: sample.key_expr().as_str().to_string(),
707 timestamp: sample.timestamp().copied(),
708 payload_len: sample.payload().len(),
709 });
710 }
711 Ok(out)
712}
713
714/// One fetched value with its provenance.
715#[derive(Debug, Clone)]
716pub struct FetchedValue {
717 /// The concrete key the value arrived on.
718 pub key: String,
719 pub payload: zenoh::bytes::ZBytes,
720 pub encoding: String,
721 pub timestamp: Option<zenoh::time::Timestamp>,
722 /// The value's attachment, when the sample carried one (#117).
723 pub attachment: Option<zenoh::bytes::ZBytes>,
724 pub source: ValueSource,
725}
726
727/// The outcome: a value, or an attributed nothing.
728#[derive(Debug, Clone)]
729pub enum FetchOutcome {
730 Value(FetchedValue),
731 /// Every rung was tried and none answered — a non-verdict, stated with
732 /// exactly what was asked (RFC 05 §3.1: silence never becomes a claim
733 /// that no value exists).
734 None {
735 attempted: [&'static str; 3],
736 },
737}
738
739/// Fetch ladder bounds.
740#[derive(Debug, Clone, Copy)]
741pub struct FetchSpec {
742 /// Per-GET timeout (two GETs happen: concrete key, then `@adv` cache).
743 pub get_timeout: Duration,
744 /// The final subscribe-window rung's duration.
745 pub window: Duration,
746}
747
748impl Default for FetchSpec {
749 fn default() -> Self {
750 FetchSpec {
751 get_timeout: Duration::from_secs(2),
752 window: Duration::from_millis(1500),
753 }
754 }
755}
756
757/// Fetch one concrete key's current value **on demand** — the value half of
758/// the lazy-observation contract (issue #84): a selection retrieves one
759/// value; nothing is prefetched, nothing stays subscribed.
760///
761/// The ladder, each rung bounded:
762/// 1. GET the concrete key (storages answer; RFC 04 §3.2's "a plain GET does
763/// not reach publisher caches" is exactly why rung 2 exists);
764/// 2. GET `<key>/@adv/**?_max=1` — zenoh-ext's AdvancedPublisher cache
765/// declares its queryable there and replies with the cached sample on its
766/// own concrete key (`@adv` is verbatim, so no data selector ever collides
767/// with it — RFC 03 §4 D2 working in our favor);
768/// 3. a brief callback subscription on the key, first sample wins.
769///
770/// Several answers on a rung (multiple storages) resolve by latest HLC
771/// timestamp; unstamped answers lose to stamped ones (RFC 04 §1.2's LWW).
772pub async fn fetch_value(session: &Session, key: &str, spec: FetchSpec) -> Result<FetchOutcome> {
773 // Rung 1 + 2: bounded GETs.
774 if let Some(v) = fetch_stored(session, key, spec.get_timeout).await? {
775 return Ok(FetchOutcome::Value(v));
776 }
777
778 // Rung 3: a window. The subscriber is explicitly undeclared afterwards —
779 // the window closes, provably.
780 let (tx, rx) = tokio::sync::oneshot::channel::<FetchedValue>();
781 let tx = std::sync::Mutex::new(Some(tx));
782 let subscriber = crate::bus::teardown::declared(
783 "window subscribe",
784 key,
785 session.declare_subscriber(key).callback(move |sample| {
786 if let Some(tx) = tx.lock().expect("fetch window lock").take() {
787 let _ = tx.send(FetchedValue {
788 key: sample.key_expr().as_str().to_string(),
789 payload: sample.payload().clone(),
790 encoding: sample.encoding().to_string(),
791 timestamp: sample.timestamp().copied(),
792 attachment: sample.attachment().cloned(),
793 source: ValueSource::Window,
794 });
795 }
796 }),
797 )
798 .await?;
799 let caught = tokio::time::timeout(spec.window, rx).await;
800 subscriber
801 .undeclare()
802 .await
803 .map_err(|e| Error::bus("window undeclare", key, e))?;
804 if let Ok(Ok(v)) = caught {
805 return Ok(FetchOutcome::Value(v));
806 }
807
808 Ok(FetchOutcome::None {
809 attempted: ["get", "@adv cache", "subscribe window"],
810 })
811}
812
813/// The **stored** half of the [`fetch_value`] ladder, on its own: GET the
814/// concrete key (rung 1 — storages answer), then GET the `@adv` cache
815/// (rung 2). No subscriber is ever declared, so this is two bounded GETs
816/// and nothing on the data plane — the shape `zenctl why`'s default run
817/// needs (issue #214), where the subscribe window is an explicit opt-in.
818///
819/// `Ok(None)` is silence, and silence is never a verdict (RFC 05 §3.1): it
820/// means neither a storage nor a publisher cache *answered*, not that no
821/// value exists.
822pub async fn fetch_stored(
823 session: &Session,
824 key: &str,
825 get_timeout: Duration,
826) -> Result<Option<FetchedValue>> {
827 for (selector, source) in [
828 (key.to_string(), ValueSource::Storage),
829 (format!("{key}/@adv/**?_max=1"), ValueSource::Cache),
830 ] {
831 if let Some(v) = get_latest(session, &selector, source, get_timeout).await? {
832 return Ok(Some(v));
833 }
834 }
835 Ok(None)
836}
837
838async fn get_latest(
839 session: &Session,
840 selector: &str,
841 source: ValueSource,
842 timeout: Duration,
843) -> Result<Option<FetchedValue>> {
844 // `accept_any`: the @adv cache replies with the cached sample on the
845 // sample's OWN key — outside the `<key>/@adv/**` selector.
846 let replies = disciplined_get(session, selector, &GetOpts::new(timeout).accept_any())
847 .await
848 .map_err(|e| Error::bus("get", selector, e))?;
849 let mut candidates = Vec::new();
850 while let Ok(reply) = replies.recv_async().await {
851 let Ok(sample) = reply.result() else { continue };
852 candidates.push(FetchedValue {
853 key: sample.key_expr().as_str().to_string(),
854 payload: sample.payload().clone(),
855 encoding: sample.encoding().to_string(),
856 timestamp: sample.timestamp().copied(),
857 attachment: sample.attachment().cloned(),
858 source,
859 });
860 }
861 Ok(pick_latest(candidates))
862}
863
864/// The winner among several answers on one rung — RFC 04 §1.2's LWW, as a
865/// pure function.
866///
867/// Latest HLC wins; a **stamped** answer beats an unstamped one whatever the
868/// order they arrived in (a storage that does not stamp cannot outrank one
869/// that does, and RFC 04 §4 is why an unstamped deployment is a doctor-grade
870/// observation rather than a tie-break rule here). Ties keep the first
871/// answer, which is the arrival order the channel gave us — arbitrary, but
872/// stated.
873///
874/// Extracted from [`get_latest`] because a rule this quiet is exactly the
875/// kind that stops being true: as a loop over a live reply channel it was
876/// unreachable from a test.
877fn pick_latest(candidates: impl IntoIterator<Item = FetchedValue>) -> Option<FetchedValue> {
878 let mut best: Option<FetchedValue> = None;
879 for candidate in candidates {
880 best = Some(match best.take() {
881 None => candidate,
882 Some(cur) => match (cur.timestamp, candidate.timestamp) {
883 (Some(a), Some(b)) if b > a => candidate,
884 (None, Some(_)) => candidate,
885 _ => cur,
886 },
887 });
888 }
889 best
890}
891
892#[cfg(test)]
893mod tests {
894 use super::*;
895
896 fn stamp(secs: u64) -> zenoh::time::Timestamp {
897 zenoh::time::Timestamp::new(
898 zenoh::time::NTP64::from(Duration::from_secs(secs)),
899 zenoh::time::TimestampId::rand(),
900 )
901 }
902
903 fn value(key: &str, timestamp: Option<zenoh::time::Timestamp>) -> FetchedValue {
904 FetchedValue {
905 key: key.to_string(),
906 payload: zenoh::bytes::ZBytes::from(vec![0u8]),
907 encoding: "application/json".to_string(),
908 timestamp,
909 attachment: None,
910 source: ValueSource::Storage,
911 }
912 }
913
914 #[test]
915 fn the_latest_hlc_wins_whatever_order_the_replies_arrived_in() {
916 let pick = |order: [u64; 3]| {
917 pick_latest(order.map(|s| value(&format!("k/{s}"), Some(stamp(s)))))
918 .expect("three candidates")
919 .key
920 };
921 assert_eq!(pick([1, 2, 3]), "k/3");
922 assert_eq!(pick([3, 2, 1]), "k/3", "arrival order is not the rule");
923 assert_eq!(pick([2, 3, 1]), "k/3");
924 }
925
926 /// RFC 04 §1.2: a storage that does not stamp cannot outrank one that
927 /// does — in either arrival order. That asymmetry is the whole reason
928 /// this is not a `max_by_key` on the timestamp.
929 #[test]
930 fn a_stamped_answer_beats_an_unstamped_one_both_ways_round() {
931 let stamped = || value("stamped", Some(stamp(7)));
932 let bare = || value("bare", None);
933 assert_eq!(pick_latest([bare(), stamped()]).unwrap().key, "stamped");
934 assert_eq!(pick_latest([stamped(), bare()]).unwrap().key, "stamped");
935 }
936
937 #[test]
938 fn nothing_answered_is_nothing_picked_and_a_tie_keeps_the_first() {
939 assert!(
940 pick_latest(Vec::new()).is_none(),
941 "silence is not a value (RFC 05 §3.1)"
942 );
943 let ts = stamp(4);
944 assert_eq!(
945 pick_latest([value("first", Some(ts)), value("second", Some(ts))])
946 .unwrap()
947 .key,
948 "first",
949 "equal stamps keep arrival order — arbitrary, but stated"
950 );
951 assert_eq!(
952 pick_latest([value("first", None), value("second", None)])
953 .unwrap()
954 .key,
955 "first",
956 "two unstamped answers cannot be ordered; the first stands"
957 );
958 }
959}