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
//! Consumer runtime: poll → dedupe → dispatch → ack / retry / dead-letter.
use std::collections::HashMap;
use std::sync::Arc;
use std::time::Duration;
use futures::future::BoxFuture;
use crate::core::engine::FrozenDiContainer;
use crate::messaging::{
EventContext, EventError, EventHandlerDescriptor, InboundMessage, MessageTransport,
};
type Handler = fn(EventContext) -> BoxFuture<'static, Result<(), EventError>>;
/// Drives one transport against the link-time handler registry.
///
/// Spawn from `ArclyPlugin::on_start`. The dispatch map is frozen at spawn —
/// the polling loop performs immutable `HashMap` reads only.
pub struct ConsumerRuntime {
pub transport: Arc<dyn MessageTransport>,
pub container: &'static FrozenDiContainer,
pub poll: Duration,
pub batch: usize,
/// Failed deliveries per message before dead-lettering.
pub max_retries: u32,
/// TTL for the consume-side dedupe claim (when an `IdempotencyStore`
/// is available in the DI container).
pub dedupe_ttl_secs: u64,
/// Messages processed concurrently per polled batch. `0`/`1` keeps the
/// historical strictly-ordered sequential behaviour; higher values trade
/// cross-message ordering within a batch for throughput (per-message
/// dedupe and retries are unaffected — `attempts` is a sharded map).
pub concurrency: usize,
}
impl ConsumerRuntime {
pub fn spawn(self) {
// Freeze the dispatch map from the inventory registry — once, here.
let dispatch: HashMap<&'static str, Handler> =
inventory::iter::<&'static EventHandlerDescriptor>
.into_iter()
.map(|d| (d.topic, d.handler))
.collect();
for d in inventory::iter::<&'static EventHandlerDescriptor> {
tracing::info!(
topic = d.topic,
consumer = d.consumer,
"event handler registered"
);
}
let runtime = Arc::new(self);
tokio::spawn(async move {
// Per-key failure counts — sharded so concurrent processing
// never contends on a single lock.
let attempts: Arc<dashmap::DashMap<String, u32>> = Arc::new(dashmap::DashMap::new());
let mut tick = tokio::time::interval(runtime.poll);
let limit = runtime.concurrency.max(1);
loop {
tick.tick().await;
// Drain-aware (like the outbox relay): once shutdown begins,
// stop claiming work — unacked messages redeliver elsewhere.
if crate::observability::health::is_draining() {
tracing::info!("consumer runtime: drain flag set — stopping");
return;
}
let batch = match runtime.transport.poll(runtime.batch).await {
Ok(b) => b,
Err(e) => {
tracing::warn!(error = %e, "transport poll failed — retrying next tick");
continue;
}
};
use futures::StreamExt;
futures::stream::iter(batch)
.for_each_concurrent(limit, |msg| {
let runtime = Arc::clone(&runtime);
let attempts = Arc::clone(&attempts);
let dispatch = &dispatch;
async move {
runtime.process(dispatch, &attempts, msg).await;
}
})
.await;
}
});
}
async fn process(
&self,
dispatch: &HashMap<&'static str, Handler>,
attempts: &dashmap::DashMap<String, u32>,
msg: InboundMessage,
) {
let Some(handler) = dispatch.get(msg.topic.as_str()) else {
// No subscriber for this topic — ack so it doesn't loop forever.
tracing::debug!(topic = %msg.topic, "no handler — acking unrouted message");
let _ = self.transport.ack(&msg).await;
return;
};
// Consume-side dedupe (at-least-once → effectively-once per TTL).
// Same contract as the HTTP idempotency layer: claim before running,
// `complete` only AFTER success, `release` on failure — so retries of
// a failed delivery pass through instead of being replay-swallowed.
let store = self
.container
.try_get::<Box<dyn crate::web::idempotency::IdempotencyStore>>();
let dedupe_key = format!("consume:{}:{}", msg.topic, msg.idempotency_key);
if let Some(store) = store {
match store.claim(&dedupe_key, self.dedupe_ttl_secs).await {
crate::web::idempotency::IdempotencyDecision::Fresh => {}
crate::web::idempotency::IdempotencyDecision::Unavailable => {}
crate::web::idempotency::IdempotencyDecision::Replay { .. } => {
metrics::counter!("events_deduped_total").increment(1);
let _ = self.transport.ack(&msg).await;
return;
}
crate::web::idempotency::IdempotencyDecision::InFlight => {
// Our own released-then-retried claim or a concurrent
// consumer; requeue and let the next poll settle it.
let _ = self.transport.nack(&msg).await;
return;
}
}
}
// One shared extraction (pipeline::Provenance): the producer's trace
// continues (fresh root when none was carried) and the envelope's
// tenant id is validated against the SAME registry as HTTP traffic.
let provenance = crate::pipeline::Provenance::from_message(&msg, self.container);
// A suspended (or unknown) tenant's queued events must stop being
// processed, exactly like its HTTP requests — park them out of band
// rather than retry-looping or silently processing. Only enforced
// when a TenantRegistry is actually configured.
if msg.tenant.is_some()
&& provenance.tenant.is_none()
&& self
.container
.try_get::<crate::web::tenant::TenantRegistry>()
.is_some()
{
metrics::counter!("events_tenant_rejected_total").increment(1);
tracing::warn!(
topic = %msg.topic,
tenant = msg.tenant.as_deref().unwrap_or(""),
"event tenant suspended or unknown — dead-lettering"
);
if let Some(store) = store {
store.release(&dedupe_key).await;
}
let _ = self
.transport
.dead_letter(&msg, "tenant suspended or unknown")
.await;
return;
}
let ctx = EventContext {
message: msg.clone(),
container: self.container,
trace: provenance.trace,
tenant: provenance.tenant,
};
match handler(ctx).await {
Ok(()) => {
attempts.remove(&msg.idempotency_key);
if let Some(store) = store {
store
.complete(&dedupe_key, 200, b"", self.dedupe_ttl_secs)
.await;
}
metrics::counter!("events_consumed_total", "topic" => msg.topic.clone())
.increment(1);
if let Err(e) = self.transport.ack(&msg).await {
tracing::warn!(error = %e, "ack failed — message may redeliver");
}
}
Err(error) => {
if let Some(store) = store {
store.release(&dedupe_key).await; // allow the retry through
}
// Typed fate: a permanent failure skips the retry budget and
// parks immediately; transient failures take the bounded
// retry path below.
let (reason, poison) = match error {
EventError::DeadLetter(m) => {
metrics::counter!("events_poisoned_total").increment(1);
(m, true)
}
EventError::Retry(m) => (m, false),
};
// Bump-and-read, dropping the shard guard BEFORE any await
// below (dead_letter / nack are async).
let n = {
let mut entry = attempts.entry(msg.idempotency_key.clone()).or_insert(0);
*entry += 1;
*entry
};
if poison || n > self.max_retries {
attempts.remove(&msg.idempotency_key);
metrics::counter!("events_dead_lettered_total").increment(1);
tracing::error!(topic = %msg.topic, key = %msg.idempotency_key,
attempts = n, reason = %reason,
"poison message → dead letter");
// PII never parks raw in the DLQ: dead-lettered payloads
// are masked first (the queue copy already served its
// delivery purpose; the DLQ copy is for forensics).
let mut parked = msg.clone();
if let Some(masker) = self.container.try_get::<crate::compliance::Masker>() {
masker.apply(&mut parked.payload);
}
let _ = self.transport.dead_letter(&parked, &reason).await;
} else {
metrics::counter!("events_retried_total").increment(1);
tracing::warn!(topic = %msg.topic, attempt = n, reason = %reason,
"event handler failed — nack for retry");
let _ = self.transport.nack(&msg).await;
}
}
}
}
}