router.rs 58.6 KB
Newer Older
1
use crate::config::types::RetryConfig;
2
use crate::core::{
3
4
    is_retryable_status, BasicWorker, CircuitBreakerConfig, HealthConfig, RetryExecutor, Worker,
    WorkerRegistry, WorkerType,
5
};
6
use crate::metrics::RouterMetrics;
7
use crate::policies::{LoadBalancingPolicy, PolicyRegistry};
8
use crate::protocols::spec::{
9
10
    ChatCompletionRequest, CompletionRequest, GenerateRequest, GenerationRequest, RerankRequest,
    RerankResponse, RerankResult, ResponsesRequest,
11
};
12
use crate::routers::header_utils;
13
use crate::routers::{RouterTrait, WorkerManagement};
14
use axum::body::to_bytes;
15
16
17
use axum::{
    body::Body,
    extract::Request,
18
    http::{header::CONTENT_LENGTH, header::CONTENT_TYPE, HeaderMap, HeaderValue, StatusCode},
19
20
21
22
    response::{IntoResponse, Response},
    Json,
};
use futures_util::StreamExt;
23
use reqwest::Client;
24
use std::collections::HashMap;
25
use std::sync::Arc;
26
use std::time::{Duration, Instant};
27
use tokio_stream::wrappers::UnboundedReceiverStream;
28
29
30
31
32
use tracing::{debug, error, info, warn};

/// Regular router that uses injected load balancing policies
#[derive(Debug)]
pub struct Router {
33
34
    worker_registry: Arc<WorkerRegistry>,
    policy_registry: Arc<PolicyRegistry>,
35
    client: Client,
36
37
    worker_startup_timeout_secs: u64,
    worker_startup_check_interval_secs: u64,
38
39
    dp_aware: bool,
    api_key: Option<String>,
40
    retry_config: RetryConfig,
41
    circuit_breaker_config: CircuitBreakerConfig,
42
43
44
45
46
    _worker_loads: Arc<tokio::sync::watch::Receiver<HashMap<String, isize>>>,
    _load_monitor_handle: Option<Arc<tokio::task::JoinHandle<()>>>,
}

impl Router {
47
    /// Create a new router with injected policy and client
48
    #[allow(clippy::too_many_arguments)]
49
    pub async fn new(
50
        worker_urls: Vec<String>,
51
        ctx: &Arc<crate::server::AppContext>,
52
53
    ) -> Result<Self, String> {
        // Update active workers gauge
54
        RouterMetrics::set_active_workers(worker_urls.len());
55
56
57

        // Wait for workers to be healthy (skip if empty - for service discovery mode)
        if !worker_urls.is_empty() {
58
59
            Self::wait_for_healthy_workers(
                &worker_urls,
60
61
                ctx.router_config.worker_startup_timeout_secs,
                ctx.router_config.worker_startup_check_interval_secs,
62
63
            )
            .await?;
64
65
        }

66
        let worker_urls = if ctx.router_config.dp_aware {
67
            // worker address now in the format of "http://host:port@dp_rank"
68
            Self::get_dp_aware_workers(&worker_urls, &ctx.router_config.api_key)
69
70
71
72
73
                .map_err(|e| format!("Failed to get dp-aware workers: {}", e))?
        } else {
            worker_urls
        };

74
        // Convert config CircuitBreakerConfig to core CircuitBreakerConfig
75
        let circuit_breaker_config = ctx.router_config.effective_circuit_breaker_config();
76
77
78
        let core_cb_config = CircuitBreakerConfig {
            failure_threshold: circuit_breaker_config.failure_threshold,
            success_threshold: circuit_breaker_config.success_threshold,
79
80
            timeout_duration: Duration::from_secs(circuit_breaker_config.timeout_duration_secs),
            window_duration: Duration::from_secs(circuit_breaker_config.window_duration_secs),
81
82
        };

83
84
85
86
87
88
89
90
91
92
93
94
95
96
        // Register workers in the registry
        // In IGW mode, we need to fetch model info from workers
        for url in &worker_urls {
            // TODO: In IGW mode, fetch model_id from worker's /get_model_info endpoint
            // For now, create worker without model_id
            let worker = BasicWorker::new(url.clone(), WorkerType::Regular)
                .with_circuit_breaker_config(core_cb_config.clone())
                .with_health_config(HealthConfig {
                    timeout_secs: ctx.router_config.health_check.timeout_secs,
                    check_interval_secs: ctx.router_config.health_check.check_interval_secs,
                    endpoint: ctx.router_config.health_check.endpoint.clone(),
                    failure_threshold: ctx.router_config.health_check.failure_threshold,
                    success_threshold: ctx.router_config.health_check.success_threshold,
                });
97

98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
            let worker_arc = Arc::new(worker);
            ctx.worker_registry.register(worker_arc.clone());

            // Notify PolicyRegistry about the new worker
            let model_id = worker_arc.model_id();
            let policy = ctx.policy_registry.on_worker_added(model_id, None);

            // If this is a cache-aware policy and it's the first worker for this model,
            // initialize it with the worker
            if policy.name() == "cache_aware" {
                if let Some(cache_aware) = policy
                    .as_any()
                    .downcast_ref::<crate::policies::CacheAwarePolicy>()
                {
                    let worker_dyn: Arc<dyn Worker> = worker_arc.clone();
                    cache_aware.init_workers(std::slice::from_ref(&worker_dyn));
                }
            }
116
117
118
119
120
121
        }

        // Setup load monitoring for PowerOfTwo policy
        let (tx, rx) = tokio::sync::watch::channel(HashMap::new());
        let worker_loads = Arc::new(rx);

122
123
124
        // Check if default policy is power_of_two for load monitoring
        let default_policy = ctx.policy_registry.get_default_policy();
        let load_monitor_handle = if default_policy.name() == "power_of_two" {
125
            let monitor_urls = worker_urls.clone();
126
            let monitor_interval = ctx.router_config.worker_startup_check_interval_secs;
127
            let policy_clone = default_policy.clone();
128
            let client_clone = ctx.client.clone();
129
130

            Some(Arc::new(tokio::spawn(async move {
131
132
133
134
135
136
137
138
                Self::monitor_worker_loads(
                    monitor_urls,
                    tx,
                    monitor_interval,
                    policy_clone,
                    client_clone,
                )
                .await;
139
140
141
142
143
144
            })))
        } else {
            None
        };

        Ok(Router {
145
146
            worker_registry: ctx.worker_registry.clone(),
            policy_registry: ctx.policy_registry.clone(),
147
148
149
150
151
152
153
154
            client: ctx.client.clone(),
            worker_startup_timeout_secs: ctx.router_config.worker_startup_timeout_secs,
            worker_startup_check_interval_secs: ctx
                .router_config
                .worker_startup_check_interval_secs,
            dp_aware: ctx.router_config.dp_aware,
            api_key: ctx.router_config.api_key.clone(),
            retry_config: ctx.router_config.effective_retry_config(),
155
            circuit_breaker_config: core_cb_config,
156
157
158
159
160
161
162
            _worker_loads: worker_loads,
            _load_monitor_handle: load_monitor_handle,
        })
    }

    /// Get the current list of worker URLs
    pub fn get_worker_urls(&self) -> Vec<String> {
163
164
165
166
167
168
169
170
171
172
        self.worker_registry.get_all_urls()
    }

    /// Get worker URLs for a specific model
    pub fn get_worker_urls_for_model(&self, model_id: Option<&str>) -> Vec<String> {
        let workers = match model_id {
            Some(model) => self.worker_registry.get_by_model_fast(model),
            None => self.worker_registry.get_all(),
        };
        workers.iter().map(|w| w.url().to_string()).collect()
173
174
    }

175
    pub async fn wait_for_healthy_workers(
176
        worker_urls: &[String],
177
178
        worker_startup_timeout_secs: u64,
        worker_startup_check_interval_secs: u64,
179
    ) -> Result<(), String> {
180
181
182
183
184
185
        if worker_urls.is_empty() {
            return Err(
                "Timeout waiting for workers to become healthy: no workers provided".to_string(),
            );
        }

186
        // Perform health check asynchronously
187
188
189
190
191
192
        Self::wait_for_healthy_workers_async(
            worker_urls,
            worker_startup_timeout_secs,
            worker_startup_check_interval_secs,
        )
        .await
193
194
195
196
    }

    async fn wait_for_healthy_workers_async(
        worker_urls: &[String],
197
198
        worker_startup_timeout_secs: u64,
        worker_startup_check_interval_secs: u64,
199
200
201
202
    ) -> Result<(), String> {
        info!(
            "Waiting for {} workers to become healthy (timeout: {}s)",
            worker_urls.len(),
203
            worker_startup_timeout_secs
204
205
        );

206
        let start_time = std::time::Instant::now();
207
208
        let client = reqwest::Client::builder()
            .timeout(Duration::from_secs(2))
209
210
211
212
            .build()
            .map_err(|e| format!("Failed to create HTTP client: {}", e))?;

        loop {
213
            if start_time.elapsed() > Duration::from_secs(worker_startup_timeout_secs) {
214
215
                error!(
                    "Timeout {}s waiting for workers {:?} to become healthy. Please set --router-worker-startup-timeout-secs (sglang_router.launch_server) or --worker-startup-timeout-secs (sglang_worker.router) to a larger value",
216
                    worker_startup_timeout_secs, worker_urls
217
218
219
                );
                return Err(format!(
                    "Timeout {}s waiting for workers {:?} to become healthy. Please set --router-worker-startup-timeout-secs (sglang_router.launch_server) or --worker-startup-timeout-secs (sglang_worker.router) to a larger value",
220
                    worker_startup_timeout_secs, worker_urls
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
            // Perform all health checks concurrently
            let mut health_checks = Vec::new();
            for url in worker_urls {
                let client_clone = client.clone();
                let url_clone = url.clone();

                let check_health = tokio::spawn(async move {
                    let health_url = format!("{}/health", url_clone);
                    match client_clone.get(&health_url).send().await {
                        Ok(res) => {
                            if res.status().is_success() {
                                None
                            } else {
                                Some((url_clone, format!("status: {}", res.status())))
                            }
                        }
                        Err(_) => Some((url_clone, "not ready".to_string())),
                    }
                });

                health_checks.push(check_health);
            }

            // Wait for all health checks to complete
            let results = futures::future::join_all(health_checks).await;

250
251
252
            let mut all_healthy = true;
            let mut unhealthy_workers = Vec::new();

253
254
255
256
            for result in results {
                match result {
                    Ok(None) => {
                        // Worker is healthy
257
                    }
258
                    Ok(Some((url, reason))) => {
259
                        all_healthy = false;
260
261
262
263
264
265
                        unhealthy_workers.push((url, reason));
                    }
                    Err(e) => {
                        all_healthy = false;
                        unhealthy_workers
                            .push(("unknown".to_string(), format!("task error: {}", e)));
266
267
268
269
270
                    }
                }
            }

            if all_healthy {
271
                info!("All {} workers are healthy", worker_urls.len());
272
273
                return Ok(());
            } else {
274
                debug!(
275
                    "Waiting for {} workers to become healthy ({} unhealthy: {:?})",
276
                    worker_urls.len(),
277
278
                    unhealthy_workers.len(),
                    unhealthy_workers
279
                );
280
                tokio::time::sleep(Duration::from_secs(worker_startup_check_interval_secs)).await;
281
282
283
284
            }
        }
    }

285
286
    fn get_worker_dp_size(worker_url: &str, api_key: &Option<String>) -> Result<usize, String> {
        let sync_client = reqwest::blocking::Client::new();
287
        let mut req_builder = sync_client.get(format!("{}/get_server_info", worker_url));
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
        if let Some(key) = api_key {
            req_builder = req_builder.bearer_auth(key);
        }

        match req_builder.send() {
            Ok(res) => {
                if res.status().is_success() {
                    let server_info = res
                        .text()
                        .map_err(|e| format!("failed to read text from response: {}", e))?;

                    let server_info: serde_json::Value = serde_json::from_str(&server_info)
                        .map_err(|e| format!("failed to decode JSON: {}", e))?;

                    let dp_size = server_info
                        .get("dp_size")
                        .and_then(|v| v.as_u64())
                        .ok_or_else(|| String::from("dp_size not found or not an u64"))?;

                    Ok(if dp_size > usize::MAX as u64 {
                        return Err(format!("dp_size is too large: {}", dp_size));
                    } else {
                        dp_size as usize
                    })
                } else {
                    Err(format!("unexpected status code: {}", res.status()))
                }
            }
            Err(e) => Err(format!("error response: {}", e)),
        }
    }

    // Given a list of workers, return a list of workers with dp_rank as suffix
    fn get_dp_aware_workers(
        worker_urls: &[String],
        api_key: &Option<String>,
    ) -> Result<Vec<String>, String> {
        let mut dp_aware_workers: Vec<String> = Vec::new();

        for url in worker_urls {
            match Self::get_worker_dp_size(url, api_key) {
                Ok(dp_size) => {
                    for i in 0..dp_size {
                        dp_aware_workers.push(format!("{}@{}", url, i));
                    }
                }
                Err(e) => return Err(format!("Failed to get DP size for {}: {}", url, e)),
            }
        }

        Ok(dp_aware_workers)
    }

341
    fn select_first_worker(&self) -> Result<String, String> {
342
343
        let workers = self.worker_registry.get_all();
        if workers.is_empty() {
344
345
            Err("No workers are available".to_string())
        } else {
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
            Ok(workers[0].url().to_string())
        }
    }

    #[allow(dead_code)]
    fn select_first_worker_for_model(&self, model_id: Option<&str>) -> Result<String, String> {
        let workers = match model_id {
            Some(model) => self.worker_registry.get_by_model_fast(model),
            None => self.worker_registry.get_all(),
        };
        if workers.is_empty() {
            Err(format!(
                "No workers are available for model: {:?}",
                model_id
            ))
        } else {
            Ok(workers[0].url().to_string())
363
364
365
        }
    }

366
    pub async fn send_health_check(&self, worker_url: &str) -> Response {
367
        let health_url = if self.dp_aware {
368
            // Need to extract the URL from "http://host:port@dp_rank"
369
370
            match Self::extract_dp_rank(worker_url) {
                Ok((worker_url_prefix, _dp_rank)) => worker_url_prefix,
371
                Err(e) => {
372
373
374
375
376
377
                    error!("Failed to extract dp_rank for health check: {}", e);
                    return (
                        StatusCode::INTERNAL_SERVER_ERROR,
                        format!("Failed to extract dp_rank: {}", e),
                    )
                        .into_response();
378
                }
379
            }
380
381
382
383
        } else {
            worker_url
        };

384
        let request_builder = self.client.get(format!("{}/health", health_url));
385
386
387

        let response = match request_builder.send().await {
            Ok(res) => {
388
389
                let status = StatusCode::from_u16(res.status().as_u16())
                    .unwrap_or(StatusCode::INTERNAL_SERVER_ERROR);
390
391

                match res.bytes().await {
392
                    Ok(body) => (status, body).into_response(),
393
394
                    Err(e) => {
                        error!(
395
                            worker_url = %health_url,
396
                            error = %e,
397
                            "Failed to read health response body"
398
                        );
399
400
401
402
403
                        (
                            StatusCode::INTERNAL_SERVER_ERROR,
                            format!("Failed to read response body: {}", e),
                        )
                            .into_response()
404
                    }
405
406
                }
            }
407
408
            Err(e) => {
                error!(
409
                    worker_url = %health_url,
410
                    error = %e,
411
                    "Failed to send health request to worker"
412
                );
413
414
415
416
417
                (
                    StatusCode::INTERNAL_SERVER_ERROR,
                    format!("Failed to send request to worker {}: {}", health_url, e),
                )
                    .into_response()
418
            }
419
420
        };

421
        // Don't record metrics for health checks
422
423
424
        response
    }

425
    // Helper method to proxy GET requests to the first available worker
426
    async fn proxy_get_request(&self, req: Request<Body>, endpoint: &str) -> Response {
427
        let headers = header_utils::copy_request_headers(&req);
428
429
430

        match self.select_first_worker() {
            Ok(worker_url) => {
431
                let mut request_builder = self.client.get(format!("{}/{}", worker_url, endpoint));
432
                for (name, value) in headers {
433
434
                    let name_lc = name.to_lowercase();
                    if name_lc != "content-type" && name_lc != "content-length" {
435
436
437
                        request_builder = request_builder.header(name, value);
                    }
                }
438

439
440
441
442
                match request_builder.send().await {
                    Ok(res) => {
                        let status = StatusCode::from_u16(res.status().as_u16())
                            .unwrap_or(StatusCode::INTERNAL_SERVER_ERROR);
443
444
445
446
447

                        // Preserve headers from backend
                        let response_headers =
                            header_utils::preserve_response_headers(res.headers());

448
                        match res.bytes().await {
449
450
451
452
453
454
                            Ok(body) => {
                                let mut response = Response::new(axum::body::Body::from(body));
                                *response.status_mut() = status;
                                *response.headers_mut() = response_headers;
                                response
                            }
455
456
457
458
459
                            Err(e) => (
                                StatusCode::INTERNAL_SERVER_ERROR,
                                format!("Failed to read response: {}", e),
                            )
                                .into_response(),
460
461
                        }
                    }
462
463
464
465
466
                    Err(e) => (
                        StatusCode::INTERNAL_SERVER_ERROR,
                        format!("Request failed: {}", e),
                    )
                        .into_response(),
467
468
                }
            }
469
            Err(e) => (StatusCode::SERVICE_UNAVAILABLE, e).into_response(),
470
471
472
        }
    }

473
474
475
476
477
478
479
480
481
482
483
484
485
    /// Select worker for a specific model considering circuit breaker state
    fn select_worker_for_model(
        &self,
        model_id: Option<&str>,
        text: Option<&str>,
    ) -> Option<Arc<dyn Worker>> {
        // Get workers for the specified model (O(1) lookup if model_id is provided)
        let workers = match model_id {
            Some(model) => self.worker_registry.get_by_model_fast(model),
            None => self.worker_registry.get_all(),
        };

        let available: Vec<Arc<dyn Worker>> = workers
486
487
            .iter()
            .filter(|w| w.is_available())
488
            .cloned()
489
490
491
492
            .collect();
        if available.is_empty() {
            return None;
        }
493
494
495
496
497
498
499
500
501

        // Get the appropriate policy for this model
        let policy = match model_id {
            Some(model) => self.policy_registry.get_policy_or_default(model),
            None => self.policy_registry.get_default_policy(),
        };

        let idx = policy.select_worker(&available, text)?;
        Some(available[idx].clone())
502
503
    }

504
    pub async fn route_typed_request<T: GenerationRequest + serde::Serialize + Clone>(
505
        &self,
506
        headers: Option<&HeaderMap>,
507
508
        typed_req: &T,
        route: &str,
509
        model_id: Option<&str>,
510
    ) -> Response {
511
        let start = Instant::now();
512
513
514
515
516
517
518
        let is_stream = typed_req.is_stream();
        let text = typed_req.extract_text_for_routing();

        let response = RetryExecutor::execute_response_with_retry(
            &self.retry_config,
            // operation per attempt
            |_: u32| async {
519
                let worker = match self.select_worker_for_model(model_id, Some(&text)) {
520
521
522
523
524
525
526
527
528
529
                    Some(w) => w,
                    None => {
                        RouterMetrics::record_request_error(route, "no_available_workers");
                        return (
                            StatusCode::SERVICE_UNAVAILABLE,
                            "No available workers (all circuits open or unhealthy)",
                        )
                            .into_response();
                    }
                };
530

531
                // Optional load tracking for cache-aware policy
532
533
534
535
536
537
538
                // Get the policy for this model to check if it's cache-aware
                let policy = match model_id {
                    Some(model) => self.policy_registry.get_policy_or_default(model),
                    None => self.policy_registry.get_default_policy(),
                };

                let load_incremented = if policy.name() == "cache_aware" {
539
540
541
                    worker.increment_load();
                    RouterMetrics::set_running_requests(worker.url(), worker.load());
                    true
542
543
544
545
                } else {
                    false
                };

546
547
548
549
550
551
552
                // Keep a clone for potential cleanup on retry
                let worker_for_cleanup = if load_incremented {
                    Some(worker.clone_worker())
                } else {
                    None
                };

553
554
                let response = self
                    .send_typed_request(
555
                        headers,
556
557
                        typed_req,
                        route,
558
                        worker.url(),
559
560
561
562
563
                        is_stream,
                        load_incremented,
                    )
                    .await;

564
                worker.record_outcome(response.status().is_success());
565
566
567
568
569
570
571
572
573
574
575
576
577

                // For retryable failures, we need to decrement load since send_typed_request
                // won't have done it (it only decrements on success or non-retryable failures)
                if is_retryable_status(response.status()) && load_incremented {
                    if let Some(cleanup_worker) = worker_for_cleanup {
                        cleanup_worker.decrement_load();
                        RouterMetrics::set_running_requests(
                            cleanup_worker.url(),
                            cleanup_worker.load(),
                        );
                    }
                }

578
579
580
                response
            },
            // should_retry predicate
581
            |res, _attempt| is_retryable_status(res.status()),
582
583
584
585
586
587
588
            // on_backoff hook
            |delay, attempt| {
                RouterMetrics::record_retry(route);
                RouterMetrics::record_retry_backoff_duration(delay, attempt);
            },
            // on_exhausted hook
            || RouterMetrics::record_retries_exhausted(route),
589
        )
590
591
592
593
594
595
        .await;

        if response.status().is_success() {
            let duration = start.elapsed();
            RouterMetrics::record_request(route);
            RouterMetrics::record_generate_duration(duration);
596
        } else if !is_retryable_status(response.status()) {
597
            RouterMetrics::record_request_error(route, "non_retryable_error");
598
        }
599
600

        response
601
602
    }

603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
    // TODO (rui): Better accommodate to the Worker abstraction
    fn extract_dp_rank(worker_url: &str) -> Result<(&str, usize), String> {
        let parts: Vec<&str> = worker_url.split('@').collect();
        if parts.len() != 2 {
            return Err(format!("invalid worker_url format: {}", worker_url));
        }

        // Parse the second part (dp_rank) into an integer
        match parts[1].parse::<usize>() {
            Ok(dp_rank) => Ok((parts[0], dp_rank)),
            Err(_) => Err(format!(
                "failed to parse dp_rank from worker_url: {}",
                worker_url
            )),
        }
    }

620
621
622
    // Send typed request directly without conversion
    async fn send_typed_request<T: serde::Serialize>(
        &self,
623
        headers: Option<&HeaderMap>,
624
625
626
627
628
        typed_req: &T,
        route: &str,
        worker_url: &str,
        is_stream: bool,
        load_incremented: bool, // Whether load was incremented for this request
629
    ) -> Response {
630
631
632
633
634
        let mut request_builder = if self.dp_aware {
            let (worker_url_prefix, dp_rank) = match Self::extract_dp_rank(worker_url) {
                Ok(tup) => tup,
                Err(e) => {
                    error!("Failed to extract dp_rank: {}", e);
635
636
637
638
639
                    return (
                        StatusCode::INTERNAL_SERVER_ERROR,
                        format!("Failed to extract dp_rank: {}", e),
                    )
                        .into_response();
640
641
642
643
644
645
646
                }
            };

            // Parse the request body
            let mut json_val = match serde_json::to_value(typed_req) {
                Ok(j) => j,
                Err(e) => {
647
648
649
650
651
                    return (
                        StatusCode::BAD_REQUEST,
                        format!("Convert into serde_json::Value failed: {}", e),
                    )
                        .into_response();
652
653
654
655
656
657
658
659
660
661
662
663
664
665
                }
            };

            // Insert the data_parallel_rank field
            if let Some(map) = json_val.as_object_mut() {
                map.insert(
                    String::from("data_parallel_rank"),
                    serde_json::json!(dp_rank),
                );
                debug!(
                    "Modified request body: {}",
                    serde_json::to_string(&json_val).unwrap_or(String::from("ERR"))
                );
            } else {
666
667
668
669
670
                return (
                    StatusCode::BAD_REQUEST,
                    "Failed to insert the data_parallel_rank field into the request body",
                )
                    .into_response();
671
672
            }

673
            self.client
674
675
676
                .post(format!("{}{}", worker_url_prefix, route))
                .json(&json_val)
        } else {
677
            self.client
678
679
680
                .post(format!("{}{}", worker_url, route))
                .json(typed_req) // Use json() directly with typed request
        };
681

682
683
684
685
        // Copy all headers from original request if provided
        if let Some(headers) = headers {
            for (name, value) in headers {
                // Skip Content-Type and Content-Length as .json() sets them
686
                if *name != CONTENT_TYPE && *name != CONTENT_LENGTH {
687
688
                    request_builder = request_builder.header(name, value);
                }
689
690
691
692
693
694
            }
        }

        let res = match request_builder.send().await {
            Ok(res) => res,
            Err(e) => {
695
696
697
698
                error!(
                    "Failed to send typed request worker_url={} route={} error={}",
                    worker_url, route, e
                );
699
700
701

                // Decrement load on error if it was incremented
                if load_incremented {
702
703
704
                    if let Some(worker) = self.worker_registry.get_by_url(worker_url) {
                        worker.decrement_load();
                        RouterMetrics::set_running_requests(worker_url, worker.load());
705
706
707
                    }
                }

708
709
710
711
712
                return (
                    StatusCode::INTERNAL_SERVER_ERROR,
                    format!("Request failed: {}", e),
                )
                    .into_response();
713
714
715
            }
        };

716
717
        let status = StatusCode::from_u16(res.status().as_u16())
            .unwrap_or(StatusCode::INTERNAL_SERVER_ERROR);
718
719

        if !is_stream {
720
            // For non-streaming requests, preserve headers
721
            let response_headers = header_utils::preserve_response_headers(res.headers());
722

723
            let response = match res.bytes().await {
724
725
726
727
728
729
                Ok(body) => {
                    let mut response = Response::new(axum::body::Body::from(body));
                    *response.status_mut() = status;
                    *response.headers_mut() = response_headers;
                    response
                }
730
                Err(e) => {
731
732
                    // IMPORTANT: Decrement load on error before returning
                    if load_incremented {
733
734
735
                        if let Some(worker) = self.worker_registry.get_by_url(worker_url) {
                            worker.decrement_load();
                            RouterMetrics::set_running_requests(worker_url, worker.load());
736
737
738
                        }
                    }

739
                    let error_msg = format!("Failed to get response body: {}", e);
740
                    (StatusCode::INTERNAL_SERVER_ERROR, error_msg).into_response()
741
742
743
744
                }
            };

            // Decrement load counter for non-streaming requests if it was incremented
745
            if load_incremented {
746
747
748
                if let Some(worker) = self.worker_registry.get_by_url(worker_url) {
                    worker.decrement_load();
                    RouterMetrics::set_running_requests(worker_url, worker.load());
749
750
751
752
753
754
                }
            }

            response
        } else if load_incremented {
            // For streaming with load tracking, we need to manually decrement when done
755
            let registry = Arc::clone(&self.worker_registry);
756
757
            let worker_url = worker_url.to_string();

758
759
760
761
762
            // Preserve headers for streaming response
            let mut response_headers = header_utils::preserve_response_headers(res.headers());
            // Ensure we set the correct content-type for SSE
            response_headers.insert(CONTENT_TYPE, HeaderValue::from_static("text/event-stream"));

763
764
765
766
767
768
            let stream = res.bytes_stream();
            let (tx, rx) = tokio::sync::mpsc::unbounded_channel();

            // Spawn task to forward stream and detect completion
            tokio::spawn(async move {
                let mut stream = stream;
769
                let mut decremented = false;
770
771
772
773
774
775
776
777
778
                while let Some(chunk) = stream.next().await {
                    match chunk {
                        Ok(bytes) => {
                            // Check for stream end marker
                            if bytes
                                .as_ref()
                                .windows(12)
                                .any(|window| window == b"data: [DONE]")
                            {
779
780
781
782
                                if let Some(worker) = registry.get_by_url(&worker_url) {
                                    worker.decrement_load();
                                    RouterMetrics::set_running_requests(&worker_url, worker.load());
                                    decremented = true;
783
784
                                }
                            }
785
786
787
788
789
790
791
792
793
794
                            if tx.send(Ok(bytes)).is_err() {
                                break;
                            }
                        }
                        Err(e) => {
                            let _ = tx.send(Err(format!("Stream error: {}", e)));
                            break;
                        }
                    }
                }
795
                if !decremented {
796
797
798
                    if let Some(worker) = registry.get_by_url(&worker_url) {
                        worker.decrement_load();
                        RouterMetrics::set_running_requests(&worker_url, worker.load());
799
800
                    }
                }
801
802
803
804
805
806
807
            });

            let stream = UnboundedReceiverStream::new(rx);
            let body = Body::from_stream(stream);

            let mut response = Response::new(body);
            *response.status_mut() = status;
808
            *response.headers_mut() = response_headers;
809
            response
810
811
        } else {
            // For requests without load tracking, just stream
812
813
814
815
816
            // Preserve headers for streaming response
            let mut response_headers = header_utils::preserve_response_headers(res.headers());
            // Ensure we set the correct content-type for SSE
            response_headers.insert(CONTENT_TYPE, HeaderValue::from_static("text/event-stream"));

817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
            let stream = res.bytes_stream();
            let (tx, rx) = tokio::sync::mpsc::unbounded_channel();

            // Spawn task to forward stream
            tokio::spawn(async move {
                let mut stream = stream;
                while let Some(chunk) = stream.next().await {
                    match chunk {
                        Ok(bytes) => {
                            if tx.send(Ok(bytes)).is_err() {
                                break;
                            }
                        }
                        Err(e) => {
                            let _ = tx.send(Err(format!("Stream error: {}", e)));
                            break;
                        }
                    }
                }
            });

            let stream = UnboundedReceiverStream::new(rx);
            let body = Body::from_stream(stream);

            let mut response = Response::new(body);
            *response.status_mut() = status;
843
            *response.headers_mut() = response_headers;
844
            response
845
846
847
848
849
850
        }
    }

    pub async fn add_worker(&self, worker_url: &str) -> Result<String, String> {
        let start_time = std::time::Instant::now();
        let client = reqwest::Client::builder()
851
            .timeout(Duration::from_secs(self.worker_startup_timeout_secs))
852
853
854
855
            .build()
            .map_err(|e| format!("Failed to create HTTP client: {}", e))?;

        loop {
856
            if start_time.elapsed() > Duration::from_secs(self.worker_startup_timeout_secs) {
857
858
                error!(
                    "Timeout {}s waiting for worker {} to become healthy. Please set --router-worker-startup-timeout-secs (sglang_router.launch_server) or --worker-startup-timeout-secs (sglang_worker.router) to a larger value",
859
                    self.worker_startup_timeout_secs, worker_url
860
861
862
                );
                return Err(format!(
                    "Timeout {}s waiting for worker {} to become healthy. Please set --router-worker-startup-timeout-secs (sglang_router.launch_server) or --worker-startup-timeout-secs (sglang_worker.router) to a larger value",
863
                    self.worker_startup_timeout_secs, worker_url
864
865
866
                ));
            }

867
            match client.get(format!("{}/health", worker_url)).send().await {
868
869
                Ok(res) => {
                    if res.status().is_success() {
870
871
872
873
874
875
876
877
                        if self.dp_aware {
                            // Need to contact the worker to extract the dp_size,
                            // and add them as multiple workers
                            let url_vec = vec![String::from(worker_url)];
                            let dp_url_vec = Self::get_dp_aware_workers(&url_vec, &self.api_key)
                                .map_err(|e| format!("Failed to get dp-aware workers: {}", e))?;
                            let mut worker_added: bool = false;
                            for dp_url in &dp_url_vec {
878
                                if self.worker_registry.get_by_url(dp_url).is_some() {
879
880
881
882
                                    warn!("Worker {} already exists", dp_url);
                                    continue;
                                }
                                info!("Added worker: {}", dp_url);
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
                                // TODO: In IGW mode, fetch model_id from worker's /get_model_info endpoint
                                let new_worker =
                                    BasicWorker::new(dp_url.to_string(), WorkerType::Regular)
                                        .with_circuit_breaker_config(
                                            self.circuit_breaker_config.clone(),
                                        );

                                let worker_arc = Arc::new(new_worker);
                                self.worker_registry.register(worker_arc.clone());

                                // Notify PolicyRegistry about the new worker
                                let model_id = worker_arc.model_id();
                                let policy = self.policy_registry.on_worker_added(model_id, None);

                                // If this is a cache-aware policy, update it with all workers for this model
                                if policy.name() == "cache_aware" {
                                    if let Some(cache_aware) = policy
                                        .as_any()
                                        .downcast_ref::<crate::policies::CacheAwarePolicy>(
                                    ) {
                                        let model_workers =
                                            self.worker_registry.get_by_model_fast(model_id);
                                        cache_aware.init_workers(&model_workers);
                                    }
                                }

909
910
911
912
913
914
                                worker_added = true;
                            }
                            if !worker_added {
                                return Err(format!("No worker added for {}", worker_url));
                            }
                        } else {
915
                            if self.worker_registry.get_by_url(worker_url).is_some() {
916
917
918
                                return Err(format!("Worker {} already exists", worker_url));
                            }
                            info!("Added worker: {}", worker_url);
919

920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
                            // TODO: In IGW mode, fetch model_id from worker's /get_model_info endpoint
                            let new_worker =
                                BasicWorker::new(worker_url.to_string(), WorkerType::Regular)
                                    .with_circuit_breaker_config(
                                        self.circuit_breaker_config.clone(),
                                    );

                            let worker_arc = Arc::new(new_worker);
                            self.worker_registry.register(worker_arc.clone());

                            // Notify PolicyRegistry about the new worker
                            let model_id = worker_arc.model_id();
                            let policy = self.policy_registry.on_worker_added(model_id, None);

                            // If this is a cache-aware policy, add this worker to it
                            if policy.name() == "cache_aware" {
                                if let Some(cache_aware) = policy
                                    .as_any()
                                    .downcast_ref::<crate::policies::CacheAwarePolicy>(
                                ) {
                                    // Get all workers for this model
                                    let model_workers =
                                        self.worker_registry.get_by_model_fast(model_id);
                                    cache_aware.init_workers(&model_workers);
                                }
                            }
946
947
                        }

948
949
                        RouterMetrics::set_active_workers(self.worker_registry.get_all().len());

950
951
                        return Ok(format!("Successfully added worker: {}", worker_url));
                    } else {
952
953
                        debug!(
                            "Worker {} health check pending - status: {}",
954
955
956
957
958
959
960
961
962
                            worker_url,
                            res.status()
                        );
                        // if the url does not have http or https prefix, warn users
                        if !worker_url.starts_with("http://") && !worker_url.starts_with("https://")
                        {
                            warn!("The worker url {} does not have http or https prefix. Please add the prefix to the url.", worker_url);
                        }

963
964
965
966
                        tokio::time::sleep(Duration::from_secs(
                            self.worker_startup_check_interval_secs,
                        ))
                        .await;
967
968
969
970
                        continue;
                    }
                }
                Err(e) => {
971
                    debug!("Worker {} health check pending - error: {}", worker_url, e);
972
973
974
975
976
977

                    // if the url does not have http or https prefix, warn users
                    if !worker_url.starts_with("http://") && !worker_url.starts_with("https://") {
                        warn!("The worker url {} does not have http or https prefix. Please add the prefix to the url.", worker_url);
                    }

978
979
980
981
                    tokio::time::sleep(Duration::from_secs(
                        self.worker_startup_check_interval_secs,
                    ))
                    .await;
982
983
984
985
986
987
988
                    continue;
                }
            }
        }
    }

    pub fn remove_worker(&self, worker_url: &str) {
989
990
991
992
993
994
        if self.dp_aware {
            // remove dp-aware workers in a prefix-matching fashion
            // without contacting the remote worker
            let mut removed_workers: Vec<String> = Vec::new();
            let worker_url_prefix = format!("{}@", worker_url);

995
996
997
998
999
1000
1001
1002
1003
1004
            // Find and remove all workers with matching prefix
            let all_workers = self.worker_registry.get_all();
            for w in all_workers.iter() {
                if w.url().starts_with(&worker_url_prefix) {
                    // Get model_id before removing
                    let model_id = w.model_id().to_string();

                    if self.worker_registry.remove_by_url(w.url()).is_some() {
                        info!("Removed worker: {}", w.url());
                        removed_workers.push(w.url().to_string());
1005

1006
1007
                        // Notify PolicyRegistry about the removed worker
                        self.policy_registry.on_worker_removed(&model_id);
1008
                    } else {
1009
                        warn!("Worker {} not found, skipping removal", w.url());
1010
1011
1012
1013
                    }
                }
            }

1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
            RouterMetrics::set_active_workers(self.worker_registry.get_all().len());

            // If any models are using cache aware policy, remove the workers from the tree
            // Check each removed worker's model and get its policy
            for dp_url in removed_workers.iter() {
                if let Some(worker) = self.worker_registry.get_by_url(dp_url) {
                    let model_id = worker.model_id();
                    if let Some(policy) = self.policy_registry.get_policy(model_id) {
                        if let Some(cache_aware) = policy
                            .as_any()
                            .downcast_ref::<crate::policies::CacheAwarePolicy>()
                        {
                            cache_aware.remove_worker_by_url(dp_url);
                            info!("Removed worker from cache-aware tree: {}", dp_url);
                        }
                    }
1030
1031
1032
                }
            }
        } else {
1033
1034
1035
            // Get the worker first to extract model_id
            let model_id = if let Some(worker) = self.worker_registry.get_by_url(worker_url) {
                worker.model_id().to_string()
1036
1037
1038
            } else {
                warn!("Worker {} not found, skipping removal", worker_url);
                return;
1039
1040
1041
1042
1043
1044
1045
1046
1047
            };

            if self.worker_registry.remove_by_url(worker_url).is_some() {
                info!("Removed worker: {}", worker_url);

                // Notify PolicyRegistry about the removed worker
                self.policy_registry.on_worker_removed(&model_id);

                RouterMetrics::set_active_workers(self.worker_registry.get_all().len());
1048
1049
            }

1050
1051
1052
1053
1054
1055
1056
1057
1058
            // If the model is using cache aware policy, remove the worker from the tree
            if let Some(policy) = self.policy_registry.get_policy(&model_id) {
                if let Some(cache_aware) = policy
                    .as_any()
                    .downcast_ref::<crate::policies::CacheAwarePolicy>()
                {
                    cache_aware.remove_worker_by_url(worker_url);
                    info!("Removed worker from cache-aware tree: {}", worker_url);
                }
1059
1060
1061
1062
            }
        }
    }

1063
    async fn get_worker_load(&self, worker_url: &str) -> Option<isize> {
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
        let worker_url = if self.dp_aware {
            // Need to extract the URL from "http://host:port@dp_rank"
            let (worker_url_prefix, _dp_rank) = match Self::extract_dp_rank(worker_url) {
                Ok(tup) => tup,
                Err(e) => {
                    error!("Failed to extract dp_rank: {}", e);
                    return None;
                }
            };
            worker_url_prefix
        } else {
            worker_url
        };

1078
1079
        match self
            .client
1080
            .get(format!("{}/get_load", worker_url))
1081
1082
1083
            .send()
            .await
        {
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
            Ok(res) if res.status().is_success() => match res.bytes().await {
                Ok(bytes) => match serde_json::from_slice::<serde_json::Value>(&bytes) {
                    Ok(data) => data
                        .get("load")
                        .and_then(|v| v.as_i64())
                        .map(|v| v as isize),
                    Err(e) => {
                        debug!("Failed to parse load response from {}: {}", worker_url, e);
                        None
                    }
                },
                Err(e) => {
                    debug!("Failed to read load response from {}: {}", worker_url, e);
                    None
                }
            },
            Ok(res) => {
                debug!(
                    "Worker {} returned non-success status: {}",
                    worker_url,
                    res.status()
                );
                None
            }
            Err(e) => {
                debug!("Failed to get load from {}: {}", worker_url, e);
                None
            }
        }
    }

    // Background task to monitor worker loads
    async fn monitor_worker_loads(
        worker_urls: Vec<String>,
        tx: tokio::sync::watch::Sender<HashMap<String, isize>>,
        interval_secs: u64,
        policy: Arc<dyn LoadBalancingPolicy>,
1121
        client: Client,
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
    ) {
        let mut interval = tokio::time::interval(Duration::from_secs(interval_secs));

        loop {
            interval.tick().await;

            let mut loads = HashMap::new();
            for url in &worker_urls {
                if let Some(load) = Self::get_worker_load_static(&client, url).await {
                    loads.insert(url.clone(), load);
                }
            }

            if !loads.is_empty() {
                // Update policy with new loads
                policy.update_loads(&loads);

                // Send to watchers
                if let Err(e) = tx.send(loads) {
                    error!("Failed to send load update: {}", e);
                }
            }
        }
    }

    // Static version of get_worker_load for use in monitoring task
    async fn get_worker_load_static(client: &reqwest::Client, worker_url: &str) -> Option<isize> {
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
        let worker_url = if worker_url.contains("@") {
            // Need to extract the URL from "http://host:port@dp_rank"
            let (worker_url_prefix, _dp_rank) = match Self::extract_dp_rank(worker_url) {
                Ok(tup) => tup,
                Err(e) => {
                    debug!("Failed to extract dp_rank: {}", e);
                    return None;
                }
            };
            worker_url_prefix
        } else {
            worker_url
        };

1163
        match client.get(format!("{}/get_load", worker_url)).send().await {
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
            Ok(res) if res.status().is_success() => match res.bytes().await {
                Ok(bytes) => match serde_json::from_slice::<serde_json::Value>(&bytes) {
                    Ok(data) => data
                        .get("load")
                        .and_then(|v| v.as_i64())
                        .map(|v| v as isize),
                    Err(e) => {
                        debug!("Failed to parse load response from {}: {}", worker_url, e);
                        None
                    }
                },
                Err(e) => {
                    debug!("Failed to read load response from {}: {}", worker_url, e);
                    None
                }
            },
            Ok(res) => {
                debug!(
                    "Worker {} returned non-success status: {}",
                    worker_url,
                    res.status()
                );
                None
            }
            Err(e) => {
                debug!("Failed to get load from {}: {}", worker_url, e);
                None
            }
        }
    }
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212

    async fn build_rerank_response(
        req: &RerankRequest,
        response: Response,
    ) -> anyhow::Result<Response> {
        let (_, response_body) = response.into_parts();
        let body_bytes = to_bytes(response_body, usize::MAX).await?;
        let rerank_results = serde_json::from_slice::<Vec<RerankResult>>(&body_bytes)?;
        let mut rerank_response =
            RerankResponse::new(rerank_results, req.model.clone(), req.rid.clone());
        rerank_response.sort_by_score();
        if let Some(top_k) = req.top_k {
            rerank_response.apply_top_k(top_k);
        }
        if !req.return_documents {
            rerank_response.drop_documents();
        }
        Ok(Json(rerank_response).into_response())
    }
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
}

use async_trait::async_trait;

#[async_trait]
impl WorkerManagement for Router {
    async fn add_worker(&self, worker_url: &str) -> Result<String, String> {
        Router::add_worker(self, worker_url).await
    }

    fn remove_worker(&self, worker_url: &str) {
        Router::remove_worker(self, worker_url)
    }

    fn get_worker_urls(&self) -> Vec<String> {
        Router::get_worker_urls(self)
    }
}

1232
#[async_trait]
1233
1234
1235
1236
1237
impl RouterTrait for Router {
    fn as_any(&self) -> &dyn std::any::Any {
        self
    }

1238
    async fn health(&self, _req: Request<Body>) -> Response {
1239
        let workers = self.worker_registry.get_all();
1240
1241
1242
1243
1244
        let unhealthy_servers: Vec<_> = workers
            .iter()
            .filter(|w| !w.is_healthy())
            .map(|w| w.url().to_string())
            .collect();
1245

1246
1247
        if unhealthy_servers.is_empty() {
            (StatusCode::OK, "All servers healthy").into_response()
1248
        } else {
1249
1250
1251
1252
1253
            (
                StatusCode::SERVICE_UNAVAILABLE,
                format!("Unhealthy servers: {:?}", unhealthy_servers),
            )
                .into_response()
1254
1255
1256
        }
    }

1257
1258
    async fn health_generate(&self, req: Request<Body>) -> Response {
        self.proxy_get_request(req, "health_generate").await
1259
1260
    }

1261
1262
    async fn get_server_info(&self, req: Request<Body>) -> Response {
        self.proxy_get_request(req, "get_server_info").await
1263
1264
    }

1265
1266
    async fn get_models(&self, req: Request<Body>) -> Response {
        self.proxy_get_request(req, "v1/models").await
1267
1268
    }

1269
1270
    async fn get_model_info(&self, req: Request<Body>) -> Response {
        self.proxy_get_request(req, "get_model_info").await
1271
1272
1273
1274
    }

    async fn route_generate(
        &self,
1275
1276
        headers: Option<&HeaderMap>,
        body: &GenerateRequest,
1277
        model_id: Option<&str>,
1278
    ) -> Response {
1279
1280
        self.route_typed_request(headers, body, "/generate", model_id)
            .await
1281
1282
1283
1284
    }

    async fn route_chat(
        &self,
1285
1286
        headers: Option<&HeaderMap>,
        body: &ChatCompletionRequest,
1287
        model_id: Option<&str>,
1288
    ) -> Response {
1289
        self.route_typed_request(headers, body, "/v1/chat/completions", model_id)
1290
            .await
1291
1292
1293
1294
    }

    async fn route_completion(
        &self,
1295
1296
        headers: Option<&HeaderMap>,
        body: &CompletionRequest,
1297
        model_id: Option<&str>,
1298
    ) -> Response {
1299
        self.route_typed_request(headers, body, "/v1/completions", model_id)
1300
            .await
1301
1302
    }

1303
1304
1305
1306
    async fn route_responses(
        &self,
        headers: Option<&HeaderMap>,
        body: &ResponsesRequest,
1307
        model_id: Option<&str>,
1308
    ) -> Response {
1309
        self.route_typed_request(headers, body, "/v1/responses", model_id)
1310
1311
1312
            .await
    }

1313
1314
1315
1316
    async fn route_embeddings(&self, _headers: Option<&HeaderMap>, _body: Body) -> Response {
        todo!()
    }

1317
1318
1319
1320
1321
1322
    async fn route_rerank(
        &self,
        headers: Option<&HeaderMap>,
        body: &RerankRequest,
        model_id: Option<&str>,
    ) -> Response {
1323
1324
1325
        if let Err(e) = body.validate() {
            return (StatusCode::BAD_REQUEST, e).into_response();
        }
1326
1327
1328
        let response = self
            .route_typed_request(headers, body, "/v1/rerank", model_id)
            .await;
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
        if response.status().is_success() {
            match Self::build_rerank_response(body, response).await {
                Ok(rerank_response) => rerank_response,
                Err(e) => {
                    error!("Failed to build rerank response: {}", e);
                    return (
                        StatusCode::INTERNAL_SERVER_ERROR,
                        "Failed to build rerank response".to_string(),
                    )
                        .into_response();
                }
            }
        } else {
            response
        }
1344
1345
    }

1346
    async fn flush_cache(&self) -> Response {
1347
1348
1349
1350
1351
1352
        // Get all worker URLs
        let worker_urls = self.get_worker_urls();

        // Send requests to all workers concurrently without headers
        let mut tasks = Vec::new();
        for worker_url in &worker_urls {
1353
1354
1355
1356
1357
1358
            let worker_url = if self.dp_aware {
                // Need to extract the URL from "http://host:port@dp_rank"
                let (worker_url_prefix, _dp_rank) = match Self::extract_dp_rank(worker_url) {
                    Ok(tup) => tup,
                    Err(e) => {
                        error!("Failed to extract dp_rank: {}", e);
1359
1360
1361
1362
1363
                        return (
                            StatusCode::INTERNAL_SERVER_ERROR,
                            format!("Failed to extract dp_rank: {}", e),
                        )
                            .into_response();
1364
1365
1366
1367
1368
1369
                    }
                };
                worker_url_prefix
            } else {
                worker_url
            };
1370
            let request_builder = self.client.post(format!("{}/flush_cache", worker_url));
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
            tasks.push(request_builder.send());
        }

        // Wait for all responses
        let results = futures_util::future::join_all(tasks).await;

        // Check if all succeeded
        let all_success = results.iter().all(|r| {
            r.as_ref()
                .map(|res| res.status().is_success())
                .unwrap_or(false)
        });

        if all_success {
1385
            (StatusCode::OK, "Cache flushed on all servers").into_response()
1386
        } else {
1387
1388
1389
1390
1391
            (
                StatusCode::INTERNAL_SERVER_ERROR,
                "Cache flush failed on one or more servers",
            )
                .into_response()
1392
1393
1394
        }
    }

1395
    async fn get_worker_loads(&self) -> Response {
1396
1397
1398
1399
1400
        let urls = self.get_worker_urls();
        let mut loads = Vec::new();

        // Get loads from all workers
        for url in &urls {
1401
            let load = self.get_worker_load(url).await.unwrap_or(-1);
1402
1403
1404
1405
1406
1407
            loads.push(serde_json::json!({
                "worker": url,
                "load": load
            }));
        }

1408
        Json(serde_json::json!({
1409
1410
            "workers": loads
        }))
1411
        .into_response()
1412
1413
1414
1415
1416
1417
    }

    fn router_type(&self) -> &'static str {
        "regular"
    }

1418
    fn readiness(&self) -> Response {
1419
        // Regular router is ready if it has at least one healthy worker
1420
1421
1422
        let workers = self.worker_registry.get_all();
        let healthy_count = workers.iter().filter(|w| w.is_healthy()).count();
        let total_workers = workers.len();
1423
1424

        if healthy_count > 0 {
1425
            Json(serde_json::json!({
1426
1427
                "status": "ready",
                "healthy_workers": healthy_count,
1428
                "total_workers": total_workers
1429
            }))
1430
            .into_response()
1431
        } else {
1432
1433
1434
1435
1436
            (
                StatusCode::SERVICE_UNAVAILABLE,
                Json(serde_json::json!({
                    "status": "not_ready",
                    "reason": "no healthy workers available",
1437
                    "total_workers": total_workers
1438
1439
1440
                })),
            )
                .into_response()
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
        }
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use std::collections::HashMap;

    fn create_test_regular_router() -> Router {
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
        // Create registries
        let worker_registry = Arc::new(WorkerRegistry::new());
        let policy_registry = Arc::new(PolicyRegistry::new(
            crate::config::types::PolicyConfig::RoundRobin,
        ));

        // Register test workers
        let worker1 = BasicWorker::new("http://worker1:8080".to_string(), WorkerType::Regular);
        let worker2 = BasicWorker::new("http://worker2:8080".to_string(), WorkerType::Regular);
        worker_registry.register(Arc::new(worker1));
        worker_registry.register(Arc::new(worker2));

1463
1464
        let (_, rx) = tokio::sync::watch::channel(HashMap::new());
        Router {
1465
1466
            worker_registry,
            policy_registry,
1467
1468
            worker_startup_timeout_secs: 5,
            worker_startup_check_interval_secs: 1,
1469
1470
            dp_aware: false,
            api_key: None,
1471
            client: Client::new(),
1472
            retry_config: RetryConfig::default(),
1473
            circuit_breaker_config: CircuitBreakerConfig::default(),
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
            _worker_loads: Arc::new(rx),
            _load_monitor_handle: None,
        }
    }

    #[test]
    fn test_router_get_worker_urls_regular() {
        let router = create_test_regular_router();
        let urls = router.get_worker_urls();

        assert_eq!(urls.len(), 2);
        assert!(urls.contains(&"http://worker1:8080".to_string()));
        assert!(urls.contains(&"http://worker2:8080".to_string()));
    }

    #[test]
    fn test_select_first_worker_regular() {
        let router = create_test_regular_router();
        let result = router.select_first_worker();

        assert!(result.is_ok());
1495
1496
1497
        let url = result.unwrap();
        // DashMap doesn't guarantee order, so just check we get one of the workers
        assert!(url == "http://worker1:8080" || url == "http://worker2:8080");
1498
1499
    }

1500
1501
1502
1503
    #[tokio::test]
    async fn test_wait_for_healthy_workers_empty_list() {
        // Empty list will return error immediately
        let result = Router::wait_for_healthy_workers(&[], 1, 1).await;
1504
        assert!(result.is_err());
1505
        assert!(result.unwrap_err().contains("no workers provided"));
1506
1507
    }

1508
1509
    #[tokio::test]
    async fn test_wait_for_healthy_workers_invalid_urls() {
1510
1511
        // This test will timeout quickly since the URLs are invalid
        let result =
1512
            Router::wait_for_healthy_workers(&["http://nonexistent:8080".to_string()], 1, 1).await;
1513
1514
1515
1516
        assert!(result.is_err());
        assert!(result.unwrap_err().contains("Timeout"));
    }
}