1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
//! Command execution: the half of [`Shard`] that turns parsed commands into
//! shard-local work and reduces the (possibly multi-shard) results.
//!
//! [`crate::shard`] owns the reactor (sockets, the inbound queue, flushing);
//! this module owns the *semantics* — transaction state, routing a command to
//! the shard(s) that own its keys, executing one op against the local store,
//! and folding sub-results into each connection's seq-ordered ring.
use crate::message::{Agg, DispatchMeta, Inbound, Op, Part, PendingSlot, SmallReply};
use crate::reduce::{drain_front, materialize};
use crate::shard::Shard;
use crate::{Commands, ResolvedCmd, Route, TxnKind};
use kevy_resp::{Argv, ArgvView, RespVersion, encode_array_len};
impl<C: Commands> Shard<C> {
/// Apply transaction state (queue inside MULTI), else dispatch the command.
pub(crate) fn handle_command<A: ArgvView + ?Sized>(&mut self, conn_id: u64, args: &A) {
// v2.0.16: CLIENT SETNAME / CLIENT GETNAME intercept (closes
// v1.52.x finding). These need per-conn state which the
// stateless `cmd_client` dispatch can't access; handle in-line
// here where we already own `&mut Conn` via `self.conns`. All
// other CLIENT subcommands fall through to the standard
// dispatch unchanged.
if Self::try_intercept_client(self, conn_id, args) {
return;
}
// One verb-resolution per cmd (was 4: txn_kind + route + is_quit +
// is_write each scanned the verb separately). KevyCommands overrides
// resolve() with a single match; non-overriding impls still pay 4×.
let resolved = self.commands.resolve(args);
// One conns probe serves the whole pre-dispatch phase — the MULTI
// check, the per-cmd proto capture, and (for the dispatching hot
// arms) the seq assignment. These were three separate map probes
// per command (in_multi here + next_seq_for + start_single's proto
// read).
let Some(c) = self.conns.get_mut(&conn_id) else { return };
let in_multi = c.multi.is_some();
let proto = c.proto;
let cluster_conn = c.cluster;
if !in_multi && matches!(resolved.txn_kind, TxnKind::Other | TxnKind::Watch) {
let seq = c.next_seq;
c.next_seq += 1;
self.start_command(conn_id, seq, proto, args, resolved, cluster_conn);
return;
}
// Transaction-state arms (rare next to the dispatch path above);
// each re-probes internally / via immediate_reply.
match (in_multi, &resolved.txn_kind) {
(false, TxnKind::Multi) => {
if let Some(c) = self.conns.get_mut(&conn_id) {
c.multi = Some(Vec::new());
}
self.immediate_reply(conn_id, b"+OK\r\n".to_vec());
}
(false, TxnKind::Exec) => {
self.immediate_reply(conn_id, b"-ERR EXEC without MULTI\r\n".to_vec());
}
(false, TxnKind::Discard) => {
self.immediate_reply(conn_id, b"-ERR DISCARD without MULTI\r\n".to_vec());
}
(true, TxnKind::Multi) => {
self.immediate_reply(conn_id, b"-ERR MULTI calls can not be nested\r\n".to_vec());
}
(true, TxnKind::Discard) => {
// DISCARD drops the queued cmds AND any `WATCH`-ed keys
// (Redis semantics — see https://redis.io/commands/discard).
if let Some(c) = self.conns.get_mut(&conn_id) {
c.multi = None;
c.watched.clear();
}
self.immediate_reply(conn_id, b"+OK\r\n".to_vec());
}
(true, TxnKind::Exec) => self.exec_transaction(conn_id),
(true, TxnKind::Watch) => self.immediate_reply(
conn_id,
b"-ERR WATCH inside MULTI is not allowed\r\n".to_vec(),
),
(true, TxnKind::Other) => {
if let Some(q) = self.conns.get_mut(&conn_id).and_then(|c| c.multi.as_mut()) {
q.push(args.to_argv());
}
self.immediate_reply(conn_id, b"+QUEUED\r\n".to_vec());
}
// (false, Other | Watch) dispatched on the early path above.
(false, TxnKind::Other | TxnKind::Watch) => {}
}
}
/// Push a slot that resolves immediately to `bytes` (preserves seq order).
pub(crate) fn immediate_reply(&mut self, conn_id: u64, bytes: Vec<u8>) {
let seq = match self.conns.get_mut(&conn_id) {
Some(c) => {
let s = c.next_seq;
c.next_seq += 1;
s
}
None => return,
};
if let Some(c) = self.conns.get_mut(&conn_id) {
let proto = c.proto;
c.pending.push_back(PendingSlot {
remaining: 1,
agg: Agg::First(None),
done: None,
proto,
});
}
self.fold(conn_id, seq, Part::Reply(SmallReply::from_vec(bytes)));
}
/// `EXEC` — emit a `*N` array header, then run the queued commands in order.
/// The seq-ordered ring concatenates their replies into one valid array.
/// If the conn has any `WATCH`-ed keys, delegate to the pre-check fan-out
/// path in [`crate::exec_watch`] (aborts if any watched key is dirty).
fn exec_transaction(&mut self, conn_id: u64) {
let (queued, watched) = match self.conns.get_mut(&conn_id) {
Some(c) => (
c.multi.take().unwrap_or_default(),
std::mem::take(&mut c.watched),
),
None => return,
};
if !watched.is_empty() {
self.exec_transaction_watched(conn_id, queued, watched);
return;
}
let mut header = Vec::new();
encode_array_len(&mut header, queued.len() as i64);
self.immediate_reply(conn_id, header);
for cmd in &queued {
let resolved = self.commands.resolve(cmd);
// EXEC's queued cmds inherit the conn's proto at execution
// time (same per-cmd capture as the live dispatch path).
let Some(c) = self.conns.get_mut(&conn_id) else { return };
let seq = c.next_seq;
c.next_seq += 1;
let proto = c.proto;
// cluster_conn = false: queued transactions execute with full
// cross-shard fan-out even on a cluster conn (superset
// behaviour — the redirect already happened, or never will).
self.start_command(conn_id, seq, proto, cmd, resolved, false);
}
}
/// Hand off to the per-shape starter (pub/sub / single-target /
/// multi-target). Each starter owns the rest of the command's life
/// cycle: pending-slot bookkeeping, local exec, and cross-shard
/// forwarding. `seq` and `proto` arrive from the caller's single
/// pre-dispatch conns probe (see [`Self::handle_command`]).
fn start_command<A: ArgvView + ?Sized>(
&mut self,
conn_id: u64,
seq: u64,
proto: RespVersion,
args: &A,
resolved: ResolvedCmd,
cluster_conn: bool,
) {
// One client command at the dispatch boundary (before fan-out, so a
// multi-key command counts once) — INFO's total_commands_processed.
self.commands.on_command();
let ResolvedCmd {
route,
is_quit,
is_write,
block_hint,
wake_idx,
..
} = resolved;
// v3.14 A0: role-gated write rejection (read-only replica).
// `seq` is already assigned by handle_command — resolve it
// directly (immediate_reply would double-assign and wedge the
// emit order).
if is_write && let Some(err) = self.commands.write_denied() {
self.push_pending_slot(conn_id, 1, Agg::First(None), false);
self.fold(conn_id, seq, Part::Reply(SmallReply::from_vec(err)));
return;
}
if !is_write && let Some(err) = self.commands.read_denied() {
self.push_pending_slot(conn_id, 1, Agg::First(None), false);
self.fold(conn_id, seq, Part::Reply(SmallReply::from_vec(err)));
return;
}
match route {
Route::Subscribe => self.do_subscribe(conn_id, seq, args, true),
Route::Unsubscribe => self.do_subscribe(conn_id, seq, args, false),
Route::Psubscribe => self.do_psubscribe(conn_id, seq, args),
Route::Punsubscribe => self.do_punsubscribe(conn_id, seq, args),
Route::Publish => self.do_publish(conn_id, seq, args),
Route::Watch => self.do_watch(conn_id, seq, args),
Route::Unwatch => self.do_unwatch(conn_id, seq),
Route::Hello => self.do_hello(conn_id, seq, args),
Route::Rename { nx } => self.start_rename(conn_id, seq, args, nx),
// v2.3 FEED.* — parse + shard-index dispatch live in
// [`crate::exec_feed`] (500-LOC house rule).
r @ (Route::FeedShards | Route::FeedTail | Route::FeedRead) => {
self.start_feed_route(conn_id, seq, args, &r, is_quit);
}
Route::Slowlog(sub) => self.start_slowlog(conn_id, seq, sub),
// v3.16 WAIT / REPL.WAIT — deferred all-shard barriers; own
// starters (not dispatch_targets) so a parked waiter never
// rides `xshard_inflight` (see [`crate::exec_replwait`]).
Route::ReplWait { numreplicas, timeout_ms } => {
self.start_repl_wait(conn_id, seq, numreplicas, timeout_ms);
}
Route::ReplBarrier { offsets, timeout_ms, miss } => {
self.start_repl_barrier(conn_id, seq, offsets, timeout_ms, miss);
}
Route::Local => {
let meta = DispatchMeta { is_write, wake_idx, key_idx: None };
self.start_single(conn_id, seq, proto, args, self.id, is_quit, block_hint, meta);
}
Route::Single(idx) => {
let shard = self.shard_of(&args[idx]);
// Cluster conns own their shard's slots only: a wrong-shard
// key redirects (`-MOVED`) instead of forwarding, keeping a
// cluster client's topology honest. `cluster_conn` is only
// ever true in cluster mode, so the compat / cluster-off
// path pays one always-false branch. (EXEC replay passes
// false — queued transactions keep full fan-out semantics.)
if cluster_conn
&& shard != self.id
&& let Some(topo) = &self.cluster
{
let slot = kevy_hash::key_hash_slot(&args[idx]);
let bytes = topo.moved(slot, shard);
self.push_pending_slot(conn_id, 1, Agg::First(None), is_quit);
self.fold(conn_id, seq, Part::Reply(SmallReply::from_vec(bytes)));
return;
}
// Keyed routes put the key at argv[1] (or argv[2] for
// XGROUP/XINFO) — well inside u8.
let meta = DispatchMeta { is_write, wake_idx, key_idx: Some(idx as u8) };
self.start_single(conn_id, seq, proto, args, shard, is_quit, block_hint, meta);
}
// v1.56: cluster conns get `-CROSSSLOT` on cross-slot multi-key
// (MGET/MSET/SINTER/SUNION/SDIFF); else fan-out as before.
other => self.start_multi_or_crossslot(
conn_id, seq, args, other, is_quit, cluster_conn,
),
}
}
// `start_single` + `try_inline_local` (and their helpers `park_blocked`
// / `post_write_housekeeping`) live in [`crate::exec_dispatch`] —
// same `impl<C: Commands> Shard<C>`, split out so this file stays
// under the 500-LOC house rule. v1.56 CROSSSLOT helpers live in
// [`crate::exec_crossslot`] for the same reason.
/// Multi-target / aggregating command (DEL, MGET, DBSIZE, fan-outs, …).
/// Builds the per-shard target list, registers a pending slot for the
/// aggregator, then dispatches each target (locally exec or cross-core
/// send).
pub(crate) fn start_multi<A: ArgvView + ?Sized>(
&mut self,
conn_id: u64,
seq: u64,
args: &A,
route: Route,
is_quit: bool,
) {
let (targets, agg) = self.build_multi_targets(args, route);
let remaining = targets.len().max(1) as u32;
self.push_pending_slot(conn_id, remaining, agg, is_quit);
// An empty key set (shouldn't happen given routing) still resolves.
if targets.is_empty() {
self.fold(conn_id, seq, Part::Int(0));
return;
}
self.dispatch_targets(conn_id, seq, targets);
}
/// Register a `PendingSlot` for `conn_id` waiting on `remaining` parts
/// to fold via `agg`. Pushed in seq order, so the slot's index is
/// `seq - next_emit`. Captures the conn's current `proto` so a
/// later `materialize` (run when the last sub-reply lands) shapes
/// the bytes per the proto that was in effect at dispatch time.
pub(crate) fn push_pending_slot(&mut self, conn_id: u64, remaining: u32, agg: Agg, is_quit: bool) {
if let Some(c) = self.conns.get_mut(&conn_id) {
let proto = c.proto;
c.pending.push_back(PendingSlot {
remaining,
agg,
done: None,
proto,
});
if is_quit {
c.closing = true;
}
}
}
/// Fan a built target list out: locally exec on this shard, or send the
/// unbatched `Inbound::Request` to the owning peer. Single-key forwards
/// never come through here — `start_single` pushes them straight onto
/// `request_batch` (the hot batched lane).
pub(crate) fn dispatch_targets(&mut self, conn_id: u64, seq: u64, targets: Vec<(usize, Op)>) {
for (shard, op) in targets {
if shard == self.id {
let part = self.exec_op(op);
self.fold(conn_id, seq, part);
} else {
// Multi-key ops (Del/MSet/Gather/…) use the unbatched path.
self.xshard_inflight += 1;
self.send_to(
shard,
Inbound::Request {
origin: self.id,
conn: conn_id,
seq,
op,
},
);
}
}
}
/// Flush each shard's accumulated single-key dispatch batch as one
/// cross-core `RequestBatch`. Call once per reactor loop. The bitmap
/// short-circuit (D3 2026-06-20) early-returns when no shard has
/// pending requests. E17 tried splitting the slow body into a
/// `#[inline(never)]` helper and reverted — body is small enough
/// that LLVM inlines it cleanly; forcing the outline added a fn
/// call on the cross-shard hot path with no upside.
#[inline]
pub(crate) fn flush_requests(&mut self) {
if self.request_batch_nonempty == 0 {
return;
}
let mut mask = self.request_batch_nonempty;
self.request_batch_nonempty = 0;
while mask != 0 {
let s = mask.trailing_zeros() as usize;
mask &= mask - 1;
if s == self.id || self.request_batch[s].is_empty() {
continue;
}
let reqs = std::mem::take(&mut self.request_batch[s]);
self.xshard_inflight += reqs.len() as u64;
self.send_to(s, Inbound::RequestBatch { origin: self.id, reqs });
}
}
// `build_multi_targets` / `group_keys` / `build_gather` / `fanout_keys` /
// `build_mset_targets` live in [`crate::exec_build`] so this file stays
// under the 500-LOC house rule; still on the same `impl Shard`.
//
// `exec_op` (the cross-shard request dispatcher) lives in
// [`crate::exec_op`]; do_subscribe / do_publish / deliver_publish /
// flush_publish live in [`crate::exec_pubsub`]. All still on the same
// `impl Shard`, but split so this file stays under 500 LOC.
/// Append a mutating command to this shard's AOF, if enabled (best-effort).
pub(crate) fn log<A: ArgvView + ?Sized>(&mut self, args: &A) {
if let Some(aof) = &mut self.aof
&& let Err(e) = aof.append(args)
{
eprintln!("kevy: shard {} aof append failed: {e}", self.id);
}
}
/// Like [`Self::log`] but TTL-persistence-safe. After logging `args`, if
/// it is a *relative*-TTL write (`EXPIRE`/`PEXPIRE`/`SETEX`/`PSETEX`/
/// `SET … EX|PX`) it appends an absolute `PEXPIREAT key <unix_ms>` derived
/// from the key's post-exec deadline. AOF replay re-anchors a relative TTL
/// to restart-time — resetting every key to a fresh full TTL
/// (INC-2026-06-09) — so the absolute follow-up overwrites that with the
/// original wall-clock deadline. Already-absolute writes (`EXPIREAT`/
/// `PEXPIREAT`) replay correctly and need no follow-up.
pub(crate) fn log_write<A: ArgvView + ?Sized>(&mut self, args: &A) {
self.log(args);
// v2.4: hash field-TTL relative forms get the same absolute
// follow-up discipline — `HPEXPIREAT key <abs> FIELDS …`
// re-anchors the replay-time deadline to the original wall
// clock. HPEXPIREAT itself is already absolute.
if args
.get(0)
.is_some_and(|v| v.eq_ignore_ascii_case(b"HEXPIRE") || v.eq_ignore_ascii_case(b"HPEXPIRE"))
{
self.log_hash_ttl_followup(args);
return;
}
if !relative_ttl_write(args) {
return;
}
let Some(key) = args.get(1) else { return };
let pttl = self.store.pttl(key);
if pttl < 0 {
return; // command left no live TTL (key gone / TTL cleared)
}
let abs = kevy_store::now_unix_ms().saturating_add(pttl as u64);
let key = key.to_vec();
let mut c = Argv::with_capacity(3, 0);
c.push(b"PEXPIREAT");
c.push(&key);
c.push(abs.to_string().as_bytes());
self.log(&c);
}
/// v2.4 log_write helper: rewrite a relative `HEXPIRE`/`HPEXPIRE`
/// frame's deadline as absolute unix-ms and append the canonical
/// `HPEXPIREAT` follow-up (fields tail copied verbatim).
fn log_hash_ttl_followup<A: ArgvView + ?Sized>(&mut self, args: &A) {
if args.len() < 6 {
return;
}
let Some(raw) = std::str::from_utf8(&args[2]).ok().and_then(|s| s.parse::<i64>().ok())
else {
return;
};
let ms = if args[0].eq_ignore_ascii_case(b"HEXPIRE") {
raw.saturating_mul(1000)
} else {
raw
};
let abs = kevy_store::now_unix_ms().saturating_add_signed(ms);
let mut c = Argv::with_capacity(args.len(), 0);
c.push(b"HPEXPIREAT");
c.push(&args[1]);
c.push(abs.to_string().as_bytes());
for i in 3..args.len() {
c.push(&args[i]);
}
self.log(&c);
}
/// Fold a sub-result into its slot; emit completed replies in seq order.
/// The `WatchCollect` / `ExecPrep` accumulators don't materialise to RESP
/// bytes — they hand off to [`crate::exec_watch`] for the conn-state
/// mutation + downstream dispatch they require.
pub(crate) fn fold(&mut self, conn_id: u64, seq: u64, part: Part) {
let watch_agg: Option<Agg> = {
let Some(conn) = self.conns.get_mut(&conn_id) else {
return;
};
if seq < conn.next_emit {
return; // already emitted (defensive — shouldn't happen)
}
let idx = (seq - conn.next_emit) as usize;
let Some(slot) = conn.pending.get_mut(idx) else {
return;
};
// (Agg::AllOk, Part::Ok) `{}` body matches the catch-all `_ => {}`
// body but documents the *expected* aggregator/part pairing — the
// wildcard arm is the fallback for impossible combinations after
// the dispatcher arms. match_same_arms would collapse the two and
// hide the contract; keep them separate.
#[allow(clippy::match_same_arms)]
match (&mut slot.agg, part) {
(Agg::First(dst), Part::Reply(b)) => *dst = Some(b),
(Agg::SumInt(acc), Part::Int(n)) => *acc += n,
// v3.16 D1 WAIT: reply = MIN over per-shard acked counts.
(Agg::MinInt(acc), Part::Int(n)) => *acc = (*acc).min(n),
// v3.16 D2 REPL.WAIT: every shard must report 1 (met).
(Agg::ReplBarrier { ok, .. }, Part::Int(n)) => *ok &= n > 0,
// v3.16 D2 REPL.TOKEN: pairs drop in by shard id.
(
Agg::ReplTokens { slots },
Part::ReplToken { shard, generation, next_offset },
) => {
if let Some(s) = slots.get_mut(shard as usize) {
*s = Some((generation, next_offset));
}
}
(Agg::AllOk, Part::Ok) => {}
(Agg::ExtensionGather { chunks, .. }, Part::ExtensionChunk(c)) => {
chunks.push(c);
}
(Agg::Gather { got, .. }, Part::Gathered(items))
| (Agg::ZStoreGather { got, .. }, Part::Gathered(items)) => {
for (k, g) in items {
got.insert(k, g);
}
}
(Agg::Keys { acc, .. }, Part::Keys(ks)) => acc.extend(ks),
(
Agg::PrefixStats { keys, expires },
Part::PrefixStats { keys: k, expires: e },
) => {
*keys += k;
*expires += e;
}
(Agg::SlowlogGet { entries, .. }, Part::SlowlogEntries(es)) => {
entries.extend(es);
}
(Agg::WatchCollect { pairs }, Part::WatchVersions(items)) => {
pairs.extend(items);
}
// Cross-shard XREAD gather: drop each stream's element into
// its request-order slot.
(Agg::XReadGather { slots }, Part::XReadElement { index, element }) => {
if let Some(slot) = slots.get_mut(index as usize) {
*slot = element;
}
}
(Agg::ExecPrep { dirty, .. }, Part::Int(n)) => *dirty |= n != 0,
// Cross-shard RENAME orchestrator: buffer the step-1
// result in the agg so finalize can ship step 2.
(
Agg::RenameOrchestrator { taken, .. },
Part::RenameTaken { value, ttl_ms },
) => *taken = Some((value, ttl_ms)),
// Step 2's put result: `refused = None` → stored; `Some`
// → NX-blocked, and the handed-back value lands in `taken`
// so finalize can restore src before the `:0` reply.
(
Agg::RenameOrchestrator { put_stored, taken, .. },
Part::RenamePutDone { refused },
) => {
*put_stored = Some(refused.is_none());
if refused.is_some() {
*taken = refused;
}
}
// The terminal step-1 miss (RenameNoSuchSrc) leaves
// `taken == None`; finalize reads that as "missing src".
_ => {}
}
slot.remaining -= 1;
if slot.remaining == 0 {
let proto = slot.proto;
let agg = std::mem::replace(&mut slot.agg, Agg::AllOk);
if matches!(
agg,
Agg::WatchCollect { .. }
| Agg::ExecPrep { .. }
| Agg::RenameOrchestrator { .. }
| Agg::ZStoreGather { .. }
| Agg::ExtensionGather { .. }
) {
Some(agg)
} else {
slot.done = Some(materialize(agg, proto));
drain_front(conn);
None
}
} else {
None
}
};
if let Some(agg) = watch_agg {
match agg {
Agg::WatchCollect { .. } | Agg::ExecPrep { .. } => {
self.finalize_watch_agg(conn_id, seq, agg);
}
Agg::RenameOrchestrator { .. } => self.finalize_rename_agg(conn_id, seq, agg),
Agg::ZStoreGather { .. } => self.finalize_zstore_agg(conn_id, seq, agg),
Agg::ExtensionGather { argv, chunks } => {
let proto = self
.conns
.get(&conn_id)
.map_or(kevy_resp::RespVersion::V2, |c| c.proto);
let reply = self.commands.extension_reduce_v3(&argv, chunks, proto);
// v2.6: a reply starting with 0x00 is a CONTINUATION —
// the remainder encodes a second fan-out argv
// (length-prefixed items). RESP replies never start
// with NUL, so the convention is unambiguous. Phase
// state rides inside the continuation argv itself
// (stateless two-phase, no new agg variant).
if reply.first() == Some(&0) {
if let Some(argv2) = decode_continuation(&reply[1..]) {
self.start_extension_phase(conn_id, seq, argv2);
} else {
self.fill_extension_slot(
conn_id,
seq,
b"-ERR internal: bad extension continuation\r\n".to_vec(),
);
}
} else {
self.fill_extension_slot(conn_id, seq, reply);
}
}
// The match above is exhaustive over what fold ever puts
// into `watch_agg` (only the orchestrator aggs). Anything
// else is a bug; ignore so a stray slot doesn't crash
// the reactor.
_ => {}
}
}
}
pub(crate) fn protocol_error(&mut self, conn_id: u64) {
let seq = match self.conns.get_mut(&conn_id) {
Some(c) => {
let s = c.next_seq;
c.next_seq += 1;
c.closing = true;
let proto = c.proto;
c.pending.push_back(PendingSlot {
remaining: 1,
agg: Agg::First(None),
done: None,
proto,
});
s
}
None => return,
};
self.fold(
conn_id,
seq,
Part::Reply(SmallReply::from_slice(b"-ERR Protocol error\r\n")),
);
}
}
/// Does `args` set a TTL via a *relative* duration (vs absolute `*AT`)? Such
/// writes need an absolute `PEXPIREAT` follow-up in the AOF — see
/// [`Shard::log_write`]. `SET … EXAT|PXAT` aren't parsed by the server's SET,
/// so only `EX`/`PX` count here.
fn relative_ttl_write<A: ArgvView + ?Sized>(args: &A) -> bool {
if args.len() < 3 {
return false;
}
let verb = &args[0];
if verb.eq_ignore_ascii_case(b"EXPIRE")
|| verb.eq_ignore_ascii_case(b"PEXPIRE")
|| verb.eq_ignore_ascii_case(b"SETEX")
|| verb.eq_ignore_ascii_case(b"PSETEX")
{
return true;
}
if verb.eq_ignore_ascii_case(b"SET") {
return (3..args.len())
.any(|i| args[i].eq_ignore_ascii_case(b"EX") || args[i].eq_ignore_ascii_case(b"PX"));
}
false
}
/// Decode a continuation payload: `[n: u32 LE][(len: u32 LE, bytes)*]`.
fn decode_continuation(b: &[u8]) -> Option<Vec<Vec<u8>>> {
let mut pos = 0usize;
let n = u32::from_le_bytes(b.get(pos..pos + 4)?.try_into().ok()?) as usize;
pos += 4;
let mut out = Vec::with_capacity(n);
for _ in 0..n {
let len = u32::from_le_bytes(b.get(pos..pos + 4)?.try_into().ok()?) as usize;
pos += 4;
out.push(b.get(pos..pos + len)?.to_vec());
pos += len;
}
Some(out)
}