service_v2.rs 24.2 KB
Newer Older
1
// SPDX-FileCopyrightText: Copyright (c) 2024-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
2
// SPDX-License-Identifier: Apache-2.0
3

4
use std::collections::HashMap;
5
use std::env::var;
Graham King's avatar
Graham King committed
6
use std::path::PathBuf;
7
use std::sync::Arc;
8
9
use std::sync::atomic::AtomicBool;
use std::sync::atomic::Ordering;
10
use std::time::Duration;
11

12
13
14
use axum::body::Body;
use axum::http::Response;

15
use super::Metrics;
16
use super::RouteDoc;
17
use super::metrics;
18
use super::metrics::register_worker_timing_metrics;
19
use crate::discovery::ModelManager;
20
use crate::endpoint_type::EndpointType;
21
22
23
use crate::kv_router::metrics::{
    RouterRequestMetrics, RoutingOverheadMetrics, register_worker_load_metrics,
};
24
use crate::request_template::RequestTemplate;
25
use anyhow::Result;
Graham King's avatar
Graham King committed
26
use axum_server::tls_rustls::RustlsConfig;
27
use derive_builder::Builder;
28
use dynamo_runtime::config::env_is_truthy;
29
use dynamo_runtime::config::environment_names::llm as env_llm;
30
use dynamo_runtime::discovery::Discovery;
31
use dynamo_runtime::logging::make_request_span;
Graham King's avatar
Graham King committed
32
use std::net::SocketAddr;
33
use tokio::task::JoinHandle;
34
use tokio_util::sync::CancellationToken;
35
use tower_http::trace::TraceLayer;
36

37
38
39
40
/// HTTP service shared state
pub struct State {
    metrics: Arc<Metrics>,
    manager: Arc<ModelManager>,
41
    discovery_client: Arc<dyn Discovery>,
42
    flags: StateFlags,
43
    cancel_token: CancellationToken,
44
45
46
47
48
49
50
}

#[derive(Default, Debug)]
struct StateFlags {
    chat_endpoints_enabled: AtomicBool,
    cmpl_endpoints_enabled: AtomicBool,
    embeddings_endpoints_enabled: AtomicBool,
51
    images_endpoints_enabled: AtomicBool,
52
    videos_endpoints_enabled: AtomicBool,
53
    responses_endpoints_enabled: AtomicBool,
54
    anthropic_endpoints_enabled: AtomicBool,
55
56
57
58
59
60
61
62
}

impl StateFlags {
    pub fn get(&self, endpoint_type: &EndpointType) -> bool {
        match endpoint_type {
            EndpointType::Chat => self.chat_endpoints_enabled.load(Ordering::Relaxed),
            EndpointType::Completion => self.cmpl_endpoints_enabled.load(Ordering::Relaxed),
            EndpointType::Embedding => self.embeddings_endpoints_enabled.load(Ordering::Relaxed),
63
            EndpointType::Images => self.images_endpoints_enabled.load(Ordering::Relaxed),
64
            EndpointType::Videos => self.videos_endpoints_enabled.load(Ordering::Relaxed),
65
66
            // TODO: add audios_endpoints_enabled flag
            EndpointType::Audios => false,
67
            EndpointType::Responses => self.responses_endpoints_enabled.load(Ordering::Relaxed),
68
69
70
            EndpointType::AnthropicMessages => {
                self.anthropic_endpoints_enabled.load(Ordering::Relaxed)
            }
71
72
73
74
75
76
77
78
79
80
81
82
83
84
        }
    }

    pub fn set(&self, endpoint_type: &EndpointType, enabled: bool) {
        match endpoint_type {
            EndpointType::Chat => self
                .chat_endpoints_enabled
                .store(enabled, Ordering::Relaxed),
            EndpointType::Completion => self
                .cmpl_endpoints_enabled
                .store(enabled, Ordering::Relaxed),
            EndpointType::Embedding => self
                .embeddings_endpoints_enabled
                .store(enabled, Ordering::Relaxed),
85
86
87
            EndpointType::Images => self
                .images_endpoints_enabled
                .store(enabled, Ordering::Relaxed),
88
89
90
            EndpointType::Videos => self
                .videos_endpoints_enabled
                .store(enabled, Ordering::Relaxed),
91
92
            // TODO: add audios_endpoints_enabled flag
            EndpointType::Audios => {}
93
94
95
            EndpointType::Responses => self
                .responses_endpoints_enabled
                .store(enabled, Ordering::Relaxed),
96
97
98
            EndpointType::AnthropicMessages => self
                .anthropic_endpoints_enabled
                .store(enabled, Ordering::Relaxed),
99
100
        }
    }
101
102
103
}

impl State {
104
105
    pub fn new(
        manager: Arc<ModelManager>,
106
        discovery_client: Arc<dyn Discovery>,
107
108
        cancel_token: CancellationToken,
    ) -> Self {
109
110
111
        Self {
            manager,
            metrics: Arc::new(Metrics::default()),
112
            discovery_client,
113
114
115
116
            flags: StateFlags {
                chat_endpoints_enabled: AtomicBool::new(false),
                cmpl_endpoints_enabled: AtomicBool::new(false),
                embeddings_endpoints_enabled: AtomicBool::new(false),
117
                images_endpoints_enabled: AtomicBool::new(false),
118
                videos_endpoints_enabled: AtomicBool::new(false),
119
                responses_endpoints_enabled: AtomicBool::new(false),
120
                anthropic_endpoints_enabled: AtomicBool::new(false),
121
            },
122
            cancel_token,
123
124
125
        }
    }

126
127
128
129
130
131
132
133
134
135
136
137
138
    /// Get the Prometheus [`Metrics`] object which tracks request counts and inflight requests
    pub fn metrics_clone(&self) -> Arc<Metrics> {
        self.metrics.clone()
    }

    pub fn manager(&self) -> &ModelManager {
        Arc::as_ref(&self.manager)
    }

    pub fn manager_clone(&self) -> Arc<ModelManager> {
        self.manager.clone()
    }

139
140
141
142
    pub fn discovery(&self) -> Arc<dyn Discovery> {
        self.discovery_client.clone()
    }

143
144
145
146
147
148
149
150
151
152
    /// Check if the service is shutting down
    pub fn is_cancelled(&self) -> bool {
        self.cancel_token.is_cancelled()
    }

    /// Get the cancellation token
    pub fn cancel_token(&self) -> &CancellationToken {
        &self.cancel_token
    }

153
154
155
156
157
158
    // TODO
    pub fn sse_keep_alive(&self) -> Option<Duration> {
        None
    }
}

159
160
#[derive(Clone)]
pub struct HttpService {
161
162
163
    // The state we share with every request handler
    state: Arc<State>,

164
165
    router: axum::Router,
    port: u16,
166
    host: String,
Graham King's avatar
Graham King committed
167
168
169
    enable_tls: bool,
    tls_cert_path: Option<PathBuf>,
    tls_key_path: Option<PathBuf>,
170
    route_docs: Vec<RouteDoc>,
171
172
173
}

#[derive(Clone, Builder)]
174
#[builder(pattern = "owned", build_fn(private, name = "build_internal"))]
175
176
177
178
pub struct HttpServiceConfig {
    #[builder(default = "8787")]
    port: u16,

179
180
181
    #[builder(setter(into), default = "String::from(\"0.0.0.0\")")]
    host: String,

Graham King's avatar
Graham King committed
182
183
184
185
186
187
188
189
190
    #[builder(default = "false")]
    enable_tls: bool,

    #[builder(default = "None")]
    tls_cert_path: Option<PathBuf>,

    #[builder(default = "None")]
    tls_key_path: Option<PathBuf>,

191
192
    // #[builder(default)]
    // custom: Vec<axum::Router>
193
    #[builder(default = "false")]
194
195
    enable_chat_endpoints: bool,

196
    #[builder(default = "false")]
197
    enable_cmpl_endpoints: bool,
198

199
    #[builder(default = "true")]
200
201
    enable_embeddings_endpoints: bool,

202
203
204
    #[builder(default = "true")]
    enable_responses_endpoints: bool,

205
206
207
    #[builder(default = "false")]
    enable_anthropic_endpoints: bool,

208
209
    #[builder(default = "None")]
    request_template: Option<RequestTemplate>,
210

211
212
    #[builder(default = "None")]
    discovery: Option<Arc<dyn Discovery>>,
213
214
215

    #[builder(default = "None")]
    cancel_token: Option<CancellationToken>,
216
217
218
219
220
221
222
223
224
225

    /// When set, the `/metrics` endpoint will also expose metrics from the
    /// DRT's registry tree (anything created via `metrics().create*()`).
    #[builder(default = "None")]
    drt_metrics: Option<dynamo_runtime::metrics::MetricsRegistry>,

    /// When set (e.g. DRT discovery), router metrics (dynamo_router_* with router_id label)
    /// are registered using discovery.instance_id() and exposed on /metrics.
    #[builder(default = "None")]
    drt_discovery: Option<Arc<dyn Discovery>>,
226
227
228
229
230
231
232
}

impl HttpService {
    pub fn builder() -> HttpServiceConfigBuilder {
        HttpServiceConfigBuilder::default()
    }

233
234
235
236
237
238
239
240
    pub fn state_clone(&self) -> Arc<State> {
        self.state.clone()
    }

    pub fn state(&self) -> &State {
        Arc::as_ref(&self.state)
    }

241
    pub fn model_manager(&self) -> &ModelManager {
242
        self.state().manager()
243
244
    }

245
246
247
248
249
250
    pub async fn spawn(&self, cancel_token: CancellationToken) -> JoinHandle<Result<()>> {
        let this = self.clone();
        tokio::spawn(async move { this.run(cancel_token).await })
    }

    pub async fn run(&self, cancel_token: CancellationToken) -> Result<()> {
251
        let address = format!("{}:{}", self.host, self.port);
Graham King's avatar
Graham King committed
252
253
        let protocol = if self.enable_tls { "HTTPS" } else { "HTTP" };
        tracing::info!(protocol, address, "Starting HTTP(S) service");
254
255
256
257

        let router = self.router.clone();
        let observer = cancel_token.child_token();

258
259
        let state_cancel = self.state.cancel_token().clone();

Graham King's avatar
Graham King committed
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
        let addr: SocketAddr = address
            .parse()
            .map_err(|e| anyhow::anyhow!("Invalid address '{}': {}", address, e))?;

        if self.enable_tls {
            let cert_path = self
                .tls_cert_path
                .as_ref()
                .ok_or_else(|| anyhow::anyhow!("TLS certificate path not provided"))?;
            let key_path = self
                .tls_key_path
                .as_ref()
                .ok_or_else(|| anyhow::anyhow!("TLS private key path not provided"))?;

            // aws_lc_rs is the default but other crates pull in `ring` also,
            // so rustls doesn't know which one to use. Tell it.
            if let Err(e) = rustls::crypto::aws_lc_rs::default_provider().install_default() {
                tracing::debug!("TLS crypto provider already installed: {e:?}");
            }

            let config = RustlsConfig::from_pem_file(cert_path, key_path)
                .await
                .map_err(|e| anyhow::anyhow!("Failed to create TLS config: {}", e))?;

            let handle = axum_server::Handle::new();
            let server = axum_server::bind_rustls(addr, config)
                .handle(handle.clone())
                .serve(router.into_make_service());

            tokio::select! {
                result = server => {
                    result.map_err(|e| anyhow::anyhow!("HTTPS server error: {}", e))?;
                }
                _ = observer.cancelled() => {
294
                    state_cancel.cancel();
Graham King's avatar
Graham King committed
295
                    tracing::info!("HTTPS server shutdown requested");
296
297
298
                    // accepting requests for 5 more seconds, to allow incorrectly routed requests to arrive
                    handle.graceful_shutdown(Some(Duration::from_secs(get_graceful_shutdown_timeout() as u64)));
                    // no longer accepting requests, draining all existing connections
Graham King's avatar
Graham King committed
299
300
301
                }
            }
        } else {
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
            let listener = tokio::net::TcpListener::bind(addr).await.map_err(|e| {
                tracing::error!(
                    protocol = %protocol,
                    address = %address,
                    error = %e,
                    "Failed to bind server to address"
                );
                match e.kind() {
                    std::io::ErrorKind::AddrInUse => anyhow::anyhow!(
                        "Failed to start {} server: port {} already in use. Use --http-port to specify a different port.",
                        protocol,
                        self.port
                    ),
                    _ => anyhow::anyhow!(
                        "Failed to start {} server on {}: {}",
                        protocol,
                        address,
                        e
                    ),
                }
            })?;
Graham King's avatar
Graham King committed
323
324

            axum::serve(listener, router)
325
326
327
328
329
330
331
332
333
                .with_graceful_shutdown(async move {
                    observer.cancelled_owned().await;
                    state_cancel.cancel();
                    tracing::info!("HTTP server shutdown requested");
                    // accepting requests for 5 more seconds, to allow incorrectly routed requests to arrive
                    tokio::time::sleep(Duration::from_secs(get_graceful_shutdown_timeout() as u64))
                        .await;
                    // no longer accepting requests, draining all existing connections
                })
Graham King's avatar
Graham King committed
334
335
336
                .await
                .inspect_err(|_| cancel_token.cancel())?;
        }
337
338

        Ok(())
339
    }
340
341
342
343
344

    /// Documentation of exposed HTTP endpoints
    pub fn route_docs(&self) -> &[RouteDoc] {
        &self.route_docs
    }
345

346
    pub fn enable_model_endpoint(&self, endpoint_type: EndpointType, enable: bool) {
347
348
349
350
351
352
353
        self.state.flags.set(&endpoint_type, enable);
        tracing::info!(
            "{} endpoints {}",
            endpoint_type.as_str(),
            if enable { "enabled" } else { "disabled" }
        );
    }
354
355
}

356
357
358
359
360
361
362
fn get_graceful_shutdown_timeout() -> usize {
    std::env::var(env_llm::DYN_HTTP_GRACEFUL_SHUTDOWN_TIMEOUT_SECS)
        .ok()
        .and_then(|s| s.parse::<usize>().ok())
        .unwrap_or(5)
}

363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
/// Environment variable to set the metrics endpoint path (default: `/metrics`)
static HTTP_SVC_METRICS_PATH_ENV: &str = "DYN_HTTP_SVC_METRICS_PATH";
/// Environment variable to set the models endpoint path (default: `/v1/models`)
static HTTP_SVC_MODELS_PATH_ENV: &str = "DYN_HTTP_SVC_MODELS_PATH";
/// Environment variable to set the health endpoint path (default: `/health`)
static HTTP_SVC_HEALTH_PATH_ENV: &str = "DYN_HTTP_SVC_HEALTH_PATH";
/// Environment variable to set the live endpoint path (default: `/live`)
static HTTP_SVC_LIVE_PATH_ENV: &str = "DYN_HTTP_SVC_LIVE_PATH";
/// Environment variable to set the chat completions endpoint path (default: `/v1/chat/completions`)
static HTTP_SVC_CHAT_PATH_ENV: &str = "DYN_HTTP_SVC_CHAT_PATH";
/// Environment variable to set the completions endpoint path (default: `/v1/completions`)
static HTTP_SVC_CMP_PATH_ENV: &str = "DYN_HTTP_SVC_CMP_PATH";
/// Environment variable to set the embeddings endpoint path (default: `/v1/embeddings`)
static HTTP_SVC_EMB_PATH_ENV: &str = "DYN_HTTP_SVC_EMB_PATH";
/// Environment variable to set the responses endpoint path (default: `/v1/responses`)
static HTTP_SVC_RESPONSES_PATH_ENV: &str = "DYN_HTTP_SVC_RESPONSES_PATH";
379
380
/// Environment variable to set the anthropic messages endpoint path (default: `/v1/messages`)
static HTTP_SVC_ANTHROPIC_PATH_ENV: &str = "DYN_HTTP_SVC_ANTHROPIC_PATH";
381

382
383
impl HttpServiceConfigBuilder {
    pub fn build(self) -> Result<HttpService, anyhow::Error> {
384
        let config: HttpServiceConfig = self.build_internal()?;
385

386
        let model_manager = Arc::new(ModelManager::new());
387
        let cancel_token = config.cancel_token.unwrap_or_default();
388
389
390
391
392
393
        // Use the provided discovery client, or fall back to a no-op memory-backed one
        // (for in-process modes that don't need discovery)
        let discovery_client = config.discovery.unwrap_or_else(|| {
            use dynamo_runtime::discovery::KVStoreDiscovery;
            Arc::new(KVStoreDiscovery::new(
                dynamo_runtime::storage::kv::Manager::memory(),
394
                cancel_token.child_token(),
395
396
            )) as Arc<dyn Discovery>
        });
397
        let state = Arc::new(State::new(model_manager, discovery_client, cancel_token));
398
399
400
401
402
403
404
405
406
407
408
409
        state
            .flags
            .set(&EndpointType::Chat, config.enable_chat_endpoints);
        state
            .flags
            .set(&EndpointType::Completion, config.enable_cmpl_endpoints);
        state
            .flags
            .set(&EndpointType::Embedding, config.enable_embeddings_endpoints);
        state
            .flags
            .set(&EndpointType::Responses, config.enable_responses_endpoints);
410
411
412
413
        state.flags.set(
            &EndpointType::AnthropicMessages,
            config.enable_anthropic_endpoints,
        );
414

415
416
        // enable prometheus metrics
        let registry = metrics::Registry::new();
417
        state.metrics_clone().register(&registry)?;
418

419
420
421
422
423
424
425
426
427
428
429
430
        // Register worker load metrics (active_decode_blocks, active_prefill_tokens per worker)
        // These are updated by KvWorkerMonitor when receiving ActiveLoad events
        if let Err(e) = register_worker_load_metrics(&registry) {
            tracing::warn!("Failed to register worker load metrics: {}", e);
        }

        // Register worker timing metrics (last_ttft, last_itl per worker)
        // These are updated by ResponseMetricCollector when observing TTFT/ITL
        if let Err(e) = register_worker_timing_metrics(&registry) {
            tracing::warn!("Failed to register worker timing metrics: {}", e);
        }

431
432
433
434
435
436
437
438
        if let Some(ref discovery) = config.drt_discovery {
            let instance_id = discovery.instance_id();
            if let Err(e) = RouterRequestMetrics::register(&registry, instance_id) {
                tracing::warn!("Failed to register router request metrics: {}", e);
            }
            if let Err(e) = RoutingOverheadMetrics::register(&registry, instance_id) {
                tracing::warn!("Failed to register routing overhead metrics: {}", e);
            }
439
440
        }

441
        let mut router = axum::Router::new();
442

443
444
445
        let mut all_docs = Vec::new();

        let mut routes = vec![
446
447
448
449
450
            metrics::router(
                registry,
                var(HTTP_SVC_METRICS_PATH_ENV).ok(),
                config.drt_metrics,
            ),
451
452
453
            super::openai::list_models_router(state.clone(), var(HTTP_SVC_MODELS_PATH_ENV).ok()),
            super::health::health_check_router(state.clone(), var(HTTP_SVC_HEALTH_PATH_ENV).ok()),
            super::health::live_check_router(state.clone(), var(HTTP_SVC_LIVE_PATH_ENV).ok()),
454
            super::busy_threshold::busy_threshold_router(state.clone(), None),
455
456
        ];

457
458
459
460
        let endpoint_routes =
            HttpServiceConfigBuilder::get_endpoints_router(state.clone(), &config.request_template);
        routes.extend(endpoint_routes);
        for (route_docs, route) in routes {
461
462
463
464
            router = router.merge(route);
            all_docs.extend(route_docs);
        }

465
466
467
468
469
470
471
        // Add OpenAPI documentation routes (must be after all other routes so it can document them)
        // Note: The path parameter is currently unused as SwaggerUi requires static paths
        let (openapi_docs, openapi_route) =
            super::openapi_docs::openapi_router(all_docs.clone(), None);
        router = router.merge(openapi_route);
        all_docs.extend(openapi_docs);

472
        // Add span for tracing
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
        // Add on_response callback for logging response status code
        router = router.layer(
            TraceLayer::new_for_http()
                .make_span_with(make_request_span)
                .on_response(
                    |response: &Response<Body>, latency: Duration, _span: &tracing::Span| {
                        let status = response.status();
                        let latency_ms = latency.as_millis();

                        if status.is_server_error() {
                            tracing::error!(
                                status = %status.as_u16(),
                                latency_ms = %latency_ms,
                                "request completed with server error"
                            );
                        } else if status.is_client_error() {
                            tracing::warn!(
                                status = %status.as_u16(),
                                latency_ms = %latency_ms,
                                "request completed with client request error"
                            );
                        } else {
                            tracing::debug!(
                                status = %status.as_u16(),
                                latency_ms = %latency_ms,
                                "request completed"
                            );
                        }
                    },
                ),
        );
504

505
        Ok(HttpService {
506
            state,
507
508
            router,
            port: config.port,
509
            host: config.host,
Graham King's avatar
Graham King committed
510
511
512
            enable_tls: config.enable_tls,
            tls_cert_path: config.tls_cert_path,
            tls_key_path: config.tls_key_path,
513
            route_docs: all_docs,
514
515
        })
    }
516
517
518
519
520

    pub fn with_request_template(mut self, request_template: Option<RequestTemplate>) -> Self {
        self.request_template = Some(request_template);
        self
    }
521

522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
    fn get_endpoints_router(
        state: Arc<State>,
        request_template: &Option<RequestTemplate>,
    ) -> Vec<(Vec<RouteDoc>, axum::Router)> {
        let mut routes = Vec::new();
        // Add chat completions route with conditional middleware
        let (chat_docs, chat_route) = super::openai::chat_completions_router(
            state.clone(),
            request_template.clone(),
            var(HTTP_SVC_CHAT_PATH_ENV).ok(),
        );
        let (cmpl_docs, cmpl_route) =
            super::openai::completions_router(state.clone(), var(HTTP_SVC_CMP_PATH_ENV).ok());
        let (embed_docs, embed_route) =
            super::openai::embeddings_router(state.clone(), var(HTTP_SVC_EMB_PATH_ENV).ok());
537
        let (images_docs, images_route) = super::openai::images_router(state.clone(), None);
538
        let (videos_docs, videos_route) = super::openai::videos_router(state.clone(), None);
539
540
541
542
543
544
545
546
547
        let (responses_docs, responses_route) = super::openai::responses_router(
            state.clone(),
            request_template.clone(),
            var(HTTP_SVC_RESPONSES_PATH_ENV).ok(),
        );
        let mut endpoint_routes = HashMap::new();
        endpoint_routes.insert(EndpointType::Chat, (chat_docs, chat_route));
        endpoint_routes.insert(EndpointType::Completion, (cmpl_docs, cmpl_route));
        endpoint_routes.insert(EndpointType::Embedding, (embed_docs, embed_route));
548
        endpoint_routes.insert(EndpointType::Images, (images_docs, images_route));
549
        endpoint_routes.insert(EndpointType::Videos, (videos_docs, videos_route));
550
551
        endpoint_routes.insert(EndpointType::Responses, (responses_docs, responses_route));

552
553
554
555
556
557
558
559
560
561
562
563
564
        if env_is_truthy(env_llm::DYN_ENABLE_ANTHROPIC_API) {
            tracing::warn!("Anthropic Messages API (/v1/messages) is experimental.");
            let (anthropic_docs, anthropic_route) = super::anthropic::anthropic_messages_router(
                state.clone(),
                request_template.clone(),
                var(HTTP_SVC_ANTHROPIC_PATH_ENV).ok(),
            );
            endpoint_routes.insert(
                EndpointType::AnthropicMessages,
                (anthropic_docs, anthropic_route),
            );
        }

565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
        for endpoint_type in EndpointType::all() {
            let state_route = state.clone();
            if !endpoint_routes.contains_key(&endpoint_type) {
                tracing::debug!("{} endpoints are disabled", endpoint_type.as_str());
                continue;
            }
            let (docs, route) = endpoint_routes.get(&endpoint_type).cloned().unwrap();
            let route = route.route_layer(axum::middleware::from_fn(
                move |req: axum::http::Request<axum::body::Body>, next: axum::middleware::Next| {
                    let state: Arc<State> = state_route.clone();
                    async move {
                        // Check if the endpoint is enabled
                        let enabled = state.flags.get(&endpoint_type);
                        if enabled {
                            Ok(next.run(req).await)
                        } else {
                            tracing::debug!("{} endpoints are disabled", endpoint_type.as_str());
582
                            Err(axum::http::StatusCode::NOT_FOUND)
583
584
585
586
587
588
589
590
                        }
                    }
                },
            ));
            routes.push((docs, route));
        }
        routes
    }
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

#[cfg(test)]
mod tests {
    use super::*;
    use serial_test::serial;
    use std::sync::Arc;
    use tokio_util::sync::CancellationToken;

    #[tokio::test]
    #[serial]
    async fn test_liveness_endpoint_reflects_cancellation() {
        // 1. Setup service & token
        let cancel_token = Arc::new(CancellationToken::new());
        let service = HttpService::builder().build().unwrap();
        let port = service.port;

        // 2. Spawn service with shared token
        let service_token = cancel_token.clone();
        let handle = tokio::spawn(async move {
            service.run((*service_token).clone()).await.unwrap();
        });

        tokio::time::sleep(std::time::Duration::from_millis(1)).await;

        // 3. Cancel the token
        cancel_token.cancel();

        // 4. Wait a tiny bit for propagation
        tokio::time::sleep(std::time::Duration::from_millis(20)).await;

        // 5. Hit the endpoint
        let client = reqwest::Client::new();
        let resp = client
            .get(format!("http://localhost:{}/live", port))
            .send()
            .await
            .expect("Request failed");

        // 6. ASSERTION: Should be 503 Service Unavailable
        assert_eq!(resp.status(), reqwest::StatusCode::SERVICE_UNAVAILABLE);

        // Clean up
        handle.abort();
    }
}