router.rs 48.6 KB
Newer Older
1
use crate::config::types::{CircuitBreakerConfig as ConfigCircuitBreakerConfig, RetryConfig};
2
3
4
use crate::core::{
    is_retryable_status, CircuitBreakerConfig, HealthChecker, RetryExecutor, Worker, WorkerFactory,
};
5
use crate::metrics::RouterMetrics;
6
use crate::openai_api_types::{ChatCompletionRequest, CompletionRequest, GenerateRequest};
7
use crate::policies::LoadBalancingPolicy;
8
9
10
11
use crate::routers::{RouterTrait, WorkerManagement};
use axum::{
    body::Body,
    extract::Request,
12
    http::{header::CONTENT_LENGTH, header::CONTENT_TYPE, HeaderMap, HeaderValue, StatusCode},
13
14
15
16
    response::{IntoResponse, Response},
    Json,
};
use futures_util::StreamExt;
17
use reqwest::Client;
18
19
20
use std::collections::HashMap;
use std::sync::{Arc, RwLock};
use std::time::{Duration, Instant};
21
use tokio_stream::wrappers::UnboundedReceiverStream;
22
use tracing::{debug, error, info, warn};
23
pub fn copy_request_headers(req: &Request<Body>) -> Vec<(String, String)> {
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
    req.headers()
        .iter()
        .filter_map(|(name, value)| {
            value
                .to_str()
                .ok()
                .map(|v| (name.to_string(), v.to_string()))
        })
        .collect()
}

/// Regular router that uses injected load balancing policies
#[derive(Debug)]
pub struct Router {
    workers: Arc<RwLock<Vec<Box<dyn Worker>>>>,
    policy: Arc<dyn LoadBalancingPolicy>,
40
    client: Client,
41
42
    timeout_secs: u64,
    interval_secs: u64,
43
44
    dp_aware: bool,
    api_key: Option<String>,
45
    retry_config: RetryConfig,
46
    circuit_breaker_config: CircuitBreakerConfig,
47
48
49
50
51
52
    _worker_loads: Arc<tokio::sync::watch::Receiver<HashMap<String, isize>>>,
    _load_monitor_handle: Option<Arc<tokio::task::JoinHandle<()>>>,
    _health_checker: Option<HealthChecker>,
}

impl Router {
53
    /// Create a new router with injected policy and client
54
    pub async fn new(
55
56
        worker_urls: Vec<String>,
        policy: Arc<dyn LoadBalancingPolicy>,
57
        client: Client,
58
59
        timeout_secs: u64,
        interval_secs: u64,
60
61
        dp_aware: bool,
        api_key: Option<String>,
62
        retry_config: RetryConfig,
63
        circuit_breaker_config: ConfigCircuitBreakerConfig,
64
65
    ) -> Result<Self, String> {
        // Update active workers gauge
66
        RouterMetrics::set_active_workers(worker_urls.len());
67
68
69

        // Wait for workers to be healthy (skip if empty - for service discovery mode)
        if !worker_urls.is_empty() {
70
            Self::wait_for_healthy_workers(&worker_urls, timeout_secs, interval_secs).await?;
71
72
        }

73
74
75
76
77
78
79
80
        let worker_urls = if dp_aware {
            // worker address now in the format of "http://host:port@dp_rank"
            Self::get_dp_aware_workers(&worker_urls, &api_key)
                .map_err(|e| format!("Failed to get dp-aware workers: {}", e))?
        } else {
            worker_urls
        };

81
82
83
84
        // Convert config CircuitBreakerConfig to core CircuitBreakerConfig
        let core_cb_config = CircuitBreakerConfig {
            failure_threshold: circuit_breaker_config.failure_threshold,
            success_threshold: circuit_breaker_config.success_threshold,
85
86
            timeout_duration: Duration::from_secs(circuit_breaker_config.timeout_duration_secs),
            window_duration: Duration::from_secs(circuit_breaker_config.window_duration_secs),
87
88
        };

89
90
91
        // Create Worker trait objects from URLs
        let workers: Vec<Box<dyn Worker>> = worker_urls
            .iter()
92
93
94
            .map(|url| {
                WorkerFactory::create_regular_with_config(url.clone(), core_cb_config.clone())
            })
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
            .collect();

        // Initialize policy with workers if needed (e.g., for cache-aware)
        if let Some(cache_aware) = policy
            .as_any()
            .downcast_ref::<crate::policies::CacheAwarePolicy>()
        {
            cache_aware.init_workers(&workers);
        }

        let workers = Arc::new(RwLock::new(workers));
        let health_checker = crate::core::start_health_checker(Arc::clone(&workers), interval_secs);

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

        let load_monitor_handle = if policy.name() == "power_of_two" {
            let monitor_urls = worker_urls.clone();
            let monitor_interval = interval_secs;
            let policy_clone = Arc::clone(&policy);
116
            let client_clone = client.clone();
117
118

            Some(Arc::new(tokio::spawn(async move {
119
120
121
122
123
124
125
126
                Self::monitor_worker_loads(
                    monitor_urls,
                    tx,
                    monitor_interval,
                    policy_clone,
                    client_clone,
                )
                .await;
127
128
129
130
131
132
133
134
            })))
        } else {
            None
        };

        Ok(Router {
            workers,
            policy,
135
            client,
136
137
            timeout_secs,
            interval_secs,
138
139
            dp_aware,
            api_key,
140
            retry_config,
141
            circuit_breaker_config: core_cb_config,
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
            _worker_loads: worker_loads,
            _load_monitor_handle: load_monitor_handle,
            _health_checker: Some(health_checker),
        })
    }

    /// Get the current list of worker URLs
    pub fn get_worker_urls(&self) -> Vec<String> {
        self.workers
            .read()
            .unwrap()
            .iter()
            .map(|w| w.url().to_string())
            .collect()
    }

158
    pub async fn wait_for_healthy_workers(
159
160
161
162
        worker_urls: &[String],
        timeout_secs: u64,
        interval_secs: u64,
    ) -> Result<(), String> {
163
164
165
166
167
168
        if worker_urls.is_empty() {
            return Err(
                "Timeout waiting for workers to become healthy: no workers provided".to_string(),
            );
        }

169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
        // Perform health check asynchronously
        Self::wait_for_healthy_workers_async(worker_urls, timeout_secs, interval_secs).await
    }

    async fn wait_for_healthy_workers_async(
        worker_urls: &[String],
        timeout_secs: u64,
        interval_secs: u64,
    ) -> Result<(), String> {
        info!(
            "Waiting for {} workers to become healthy (timeout: {}s)",
            worker_urls.len(),
            timeout_secs
        );

184
        let start_time = std::time::Instant::now();
185
186
        let client = reqwest::Client::builder()
            .timeout(Duration::from_secs(2))
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
            .build()
            .map_err(|e| format!("Failed to create HTTP client: {}", e))?;

        loop {
            if start_time.elapsed() > Duration::from_secs(timeout_secs) {
                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",
                    timeout_secs, worker_urls
                );
                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",
                    timeout_secs, worker_urls
                ));
            }

202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
            // 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;

228
229
230
            let mut all_healthy = true;
            let mut unhealthy_workers = Vec::new();

231
232
233
234
            for result in results {
                match result {
                    Ok(None) => {
                        // Worker is healthy
235
                    }
236
                    Ok(Some((url, reason))) => {
237
                        all_healthy = false;
238
239
240
241
242
243
                        unhealthy_workers.push((url, reason));
                    }
                    Err(e) => {
                        all_healthy = false;
                        unhealthy_workers
                            .push(("unknown".to_string(), format!("task error: {}", e)));
244
245
246
247
248
                    }
                }
            }

            if all_healthy {
249
                info!("All {} workers are healthy", worker_urls.len());
250
251
                return Ok(());
            } else {
252
                debug!(
253
                    "Waiting for {} workers to become healthy ({} unhealthy: {:?})",
254
                    worker_urls.len(),
255
256
                    unhealthy_workers.len(),
                    unhealthy_workers
257
                );
258
                tokio::time::sleep(Duration::from_secs(interval_secs)).await;
259
260
261
262
            }
        }
    }

263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
    fn get_worker_dp_size(worker_url: &str, api_key: &Option<String>) -> Result<usize, String> {
        let sync_client = reqwest::blocking::Client::new();
        let mut req_builder = sync_client.get(&format!("{}/get_server_info", worker_url));
        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)
    }

319
320
321
322
323
324
325
326
327
    fn select_first_worker(&self) -> Result<String, String> {
        let workers_guard = self.workers.read().unwrap();
        if workers_guard.is_empty() {
            Err("No workers are available".to_string())
        } else {
            Ok(workers_guard[0].url().to_string())
        }
    }

328
    pub async fn send_health_check(&self, worker_url: &str) -> Response {
329
        let health_url = if self.dp_aware {
330
            // Need to extract the URL from "http://host:port@dp_rank"
331
332
            match Self::extract_dp_rank(worker_url) {
                Ok((worker_url_prefix, _dp_rank)) => worker_url_prefix,
333
                Err(e) => {
334
335
336
337
338
339
                    error!("Failed to extract dp_rank for health check: {}", e);
                    return (
                        StatusCode::INTERNAL_SERVER_ERROR,
                        format!("Failed to extract dp_rank: {}", e),
                    )
                        .into_response();
340
                }
341
            }
342
343
344
345
        } else {
            worker_url
        };

346
        let request_builder = self.client.get(format!("{}/health", health_url));
347
348
349

        let response = match request_builder.send().await {
            Ok(res) => {
350
351
                let status = StatusCode::from_u16(res.status().as_u16())
                    .unwrap_or(StatusCode::INTERNAL_SERVER_ERROR);
352
353

                match res.bytes().await {
354
                    Ok(body) => (status, body).into_response(),
355
356
                    Err(e) => {
                        error!(
357
                            worker_url = %health_url,
358
                            error = %e,
359
                            "Failed to read health response body"
360
                        );
361
362
363
364
365
                        (
                            StatusCode::INTERNAL_SERVER_ERROR,
                            format!("Failed to read response body: {}", e),
                        )
                            .into_response()
366
                    }
367
368
                }
            }
369
370
            Err(e) => {
                error!(
371
                    worker_url = %health_url,
372
                    error = %e,
373
                    "Failed to send health request to worker"
374
                );
375
376
377
378
379
                (
                    StatusCode::INTERNAL_SERVER_ERROR,
                    format!("Failed to send request to worker {}: {}", health_url, e),
                )
                    .into_response()
380
            }
381
382
        };

383
        // Don't record metrics for health checks
384
385
386
        response
    }

387
    // Helper method to proxy GET requests to the first available worker
388
    async fn proxy_get_request(&self, req: Request<Body>, endpoint: &str) -> Response {
389
390
391
392
        let headers = copy_request_headers(&req);

        match self.select_first_worker() {
            Ok(worker_url) => {
393
                let mut request_builder = self.client.get(format!("{}/{}", worker_url, endpoint));
394
                for (name, value) in headers {
395
396
                    let name_lc = name.to_lowercase();
                    if name_lc != "content-type" && name_lc != "content-length" {
397
398
399
                        request_builder = request_builder.header(name, value);
                    }
                }
400

401
402
403
404
405
406
407
408
409
410
411
                match request_builder.send().await {
                    Ok(res) => {
                        let status = StatusCode::from_u16(res.status().as_u16())
                            .unwrap_or(StatusCode::INTERNAL_SERVER_ERROR);
                        match res.bytes().await {
                            Ok(body) => (status, body).into_response(),
                            Err(e) => (
                                StatusCode::INTERNAL_SERVER_ERROR,
                                format!("Failed to read response: {}", e),
                            )
                                .into_response(),
412
413
                        }
                    }
414
415
416
417
418
                    Err(e) => (
                        StatusCode::INTERNAL_SERVER_ERROR,
                        format!("Request failed: {}", e),
                    )
                        .into_response(),
419
420
                }
            }
421
            Err(e) => (StatusCode::SERVICE_UNAVAILABLE, e).into_response(),
422
423
424
425
        }
    }

    // New method to route typed requests directly
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
    /// Select worker considering circuit breaker state
    fn select_worker_with_circuit_breaker(&self, text: Option<&str>) -> Option<Box<dyn Worker>> {
        let workers = self.workers.read().ok()?;
        let available: Vec<Box<dyn Worker>> = workers
            .iter()
            .filter(|w| w.is_available())
            .map(|w| w.clone_worker())
            .collect();
        if available.is_empty() {
            return None;
        }
        let idx = self.policy.select_worker(&available, text)?;
        Some(available[idx].clone_worker())
    }

441
442
443
444
    pub async fn route_typed_request<
        T: crate::openai_api_types::GenerationRequest + serde::Serialize + Clone,
    >(
        &self,
445
        headers: Option<&HeaderMap>,
446
447
        typed_req: &T,
        route: &str,
448
    ) -> Response {
449
        let start = Instant::now();
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
        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 {
                let worker = match self.select_worker_with_circuit_breaker(Some(&text)) {
                    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();
                    }
                };
468

469
                // Optional load tracking for cache-aware policy
470
                let load_incremented = if self.policy.name() == "cache_aware" {
471
472
473
                    worker.increment_load();
                    RouterMetrics::set_running_requests(worker.url(), worker.load());
                    true
474
475
476
477
478
479
                } else {
                    false
                };

                let response = self
                    .send_typed_request(
480
                        headers,
481
482
                        typed_req,
                        route,
483
                        worker.url(),
484
485
486
487
488
                        is_stream,
                        load_incremented,
                    )
                    .await;

489
490
491
492
                worker.record_outcome(response.status().is_success());
                response
            },
            // should_retry predicate
493
            |res, _attempt| is_retryable_status(res.status()),
494
495
496
497
498
499
500
            // on_backoff hook
            |delay, attempt| {
                RouterMetrics::record_retry(route);
                RouterMetrics::record_retry_backoff_duration(delay, attempt);
            },
            // on_exhausted hook
            || RouterMetrics::record_retries_exhausted(route),
501
        )
502
503
504
505
506
507
        .await;

        if response.status().is_success() {
            let duration = start.elapsed();
            RouterMetrics::record_request(route);
            RouterMetrics::record_generate_duration(duration);
508
        } else if !is_retryable_status(response.status()) {
509
            RouterMetrics::record_request_error(route, "non_retryable_error");
510
        }
511
512

        response
513
514
    }

515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
    // 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
            )),
        }
    }

532
533
534
    // Send typed request directly without conversion
    async fn send_typed_request<T: serde::Serialize>(
        &self,
535
        headers: Option<&HeaderMap>,
536
537
538
539
540
        typed_req: &T,
        route: &str,
        worker_url: &str,
        is_stream: bool,
        load_incremented: bool, // Whether load was incremented for this request
541
    ) -> Response {
542
543
544
545
546
        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);
547
548
549
550
551
                    return (
                        StatusCode::INTERNAL_SERVER_ERROR,
                        format!("Failed to extract dp_rank: {}", e),
                    )
                        .into_response();
552
553
554
555
556
557
558
                }
            };

            // Parse the request body
            let mut json_val = match serde_json::to_value(typed_req) {
                Ok(j) => j,
                Err(e) => {
559
560
561
562
563
                    return (
                        StatusCode::BAD_REQUEST,
                        format!("Convert into serde_json::Value failed: {}", e),
                    )
                        .into_response();
564
565
566
567
568
569
570
571
572
573
574
575
576
577
                }
            };

            // 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 {
578
579
580
581
582
                return (
                    StatusCode::BAD_REQUEST,
                    "Failed to insert the data_parallel_rank field into the request body",
                )
                    .into_response();
583
584
            }

585
            self.client
586
587
588
                .post(format!("{}{}", worker_url_prefix, route))
                .json(&json_val)
        } else {
589
            self.client
590
591
592
                .post(format!("{}{}", worker_url, route))
                .json(typed_req) // Use json() directly with typed request
        };
593

594
595
596
597
        // 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
598
                if *name != CONTENT_TYPE && *name != CONTENT_LENGTH {
599
600
                    request_builder = request_builder.header(name, value);
                }
601
602
603
604
605
606
            }
        }

        let res = match request_builder.send().await {
            Ok(res) => res,
            Err(e) => {
607
608
609
610
                error!(
                    "Failed to send typed request worker_url={} route={} error={}",
                    worker_url, route, e
                );
611
612
613
614
615
616

                // Decrement load on error if it was incremented
                if load_incremented {
                    if let Ok(workers_guard) = self.workers.read() {
                        if let Some(worker) = workers_guard.iter().find(|w| w.url() == worker_url) {
                            worker.decrement_load();
617
                            RouterMetrics::set_running_requests(&worker_url, worker.load());
618
619
620
621
                        }
                    }
                }

622
623
624
625
626
                return (
                    StatusCode::INTERNAL_SERVER_ERROR,
                    format!("Request failed: {}", e),
                )
                    .into_response();
627
628
629
            }
        };

630
631
        let status = StatusCode::from_u16(res.status().as_u16())
            .unwrap_or(StatusCode::INTERNAL_SERVER_ERROR);
632
633
634
635

        if !is_stream {
            // For non-streaming requests, get response first
            let response = match res.bytes().await {
636
                Ok(body) => (status, body).into_response(),
637
638
                Err(e) => {
                    let error_msg = format!("Failed to get response body: {}", e);
639
                    (StatusCode::INTERNAL_SERVER_ERROR, error_msg).into_response()
640
641
642
643
644
645
646
647
                }
            };

            // Decrement load counter for non-streaming requests if it was incremented
            if load_incremented && !is_stream {
                if let Ok(workers_guard) = self.workers.read() {
                    if let Some(worker) = workers_guard.iter().find(|w| w.url() == worker_url) {
                        worker.decrement_load();
648
                        RouterMetrics::set_running_requests(&worker_url, worker.load());
649
650
651
652
653
654
655
656
657
658
                    }
                }
            }

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

659
660
661
662
663
664
            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;
665
                let mut decremented = false;
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
                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]")
                            {
                                if let Ok(workers_guard) = workers.read() {
                                    if let Some(worker) =
                                        workers_guard.iter().find(|w| w.url() == &worker_url)
                                    {
                                        worker.decrement_load();
                                        RouterMetrics::set_running_requests(
                                            &worker_url,
                                            worker.load(),
                                        );
684
                                        decremented = true;
685
686
687
                                    }
                                }
                            }
688
689
690
691
692
693
694
695
696
697
                            if tx.send(Ok(bytes)).is_err() {
                                break;
                            }
                        }
                        Err(e) => {
                            let _ = tx.send(Err(format!("Stream error: {}", e)));
                            break;
                        }
                    }
                }
698
699
700
701
702
703
704
705
706
                if !decremented {
                    if let Ok(workers_guard) = workers.read() {
                        if let Some(worker) = workers_guard.iter().find(|w| w.url() == &worker_url)
                        {
                            worker.decrement_load();
                            RouterMetrics::set_running_requests(&worker_url, worker.load());
                        }
                    }
                }
707
708
709
710
711
712
713
714
715
716
717
            });

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

            let mut response = Response::new(body);
            *response.status_mut() = status;
            response
                .headers_mut()
                .insert(CONTENT_TYPE, HeaderValue::from_static("text/event-stream"));
            response
718
719
        } else {
            // For requests without load tracking, just stream
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
            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;
            response
                .headers_mut()
                .insert(CONTENT_TYPE, HeaderValue::from_static("text/event-stream"));
            response
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
        }
    }

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

        loop {
            if start_time.elapsed() > Duration::from_secs(self.timeout_secs) {
                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",
                    self.timeout_secs, worker_url
                );
                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",
                    self.timeout_secs, worker_url
                ));
            }

            match client.get(&format!("{}/health", worker_url)).send().await {
                Ok(res) => {
                    if res.status().is_success() {
                        let mut workers_guard = self.workers.write().unwrap();
776
777
778
779
780
781
782
783
784
785
786
787
788
                        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 {
                                if workers_guard.iter().any(|w| w.url() == dp_url) {
                                    warn!("Worker {} already exists", dp_url);
                                    continue;
                                }
                                info!("Added worker: {}", dp_url);
789
790
791
792
                                let new_worker = WorkerFactory::create_regular_with_config(
                                    dp_url.to_string(),
                                    self.circuit_breaker_config.clone(),
                                );
793
794
795
796
797
798
799
800
801
802
803
                                workers_guard.push(new_worker);
                                worker_added = true;
                            }
                            if !worker_added {
                                return Err(format!("No worker added for {}", worker_url));
                            }
                        } else {
                            if workers_guard.iter().any(|w| w.url() == worker_url) {
                                return Err(format!("Worker {} already exists", worker_url));
                            }
                            info!("Added worker: {}", worker_url);
804
805
806
807
                            let new_worker = WorkerFactory::create_regular_with_config(
                                worker_url.to_string(),
                                self.circuit_breaker_config.clone(),
                            );
808
                            workers_guard.push(new_worker);
809
                        }
810

811
                        RouterMetrics::set_active_workers(workers_guard.len());
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826

                        // If cache aware policy, initialize the worker in the tree
                        if let Some(cache_aware) =
                            self.policy
                                .as_any()
                                .downcast_ref::<crate::policies::CacheAwarePolicy>()
                        {
                            // Get updated workers after adding
                            drop(workers_guard);
                            let workers_guard = self.workers.read().unwrap();
                            cache_aware.init_workers(&workers_guard);
                        }

                        return Ok(format!("Successfully added worker: {}", worker_url));
                    } else {
827
828
                        debug!(
                            "Worker {} health check pending - status: {}",
829
830
831
832
833
834
835
836
837
838
839
840
841
842
                            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);
                        }

                        tokio::time::sleep(Duration::from_secs(self.interval_secs)).await;
                        continue;
                    }
                }
                Err(e) => {
843
                    debug!("Worker {} health check pending - error: {}", worker_url, e);
844
845
846
847
848
849
850
851
852
853
854
855
856
857

                    // 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);
                    }

                    tokio::time::sleep(Duration::from_secs(self.interval_secs)).await;
                    continue;
                }
            }
        }
    }

    pub fn remove_worker(&self, worker_url: &str) {
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
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
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
        if self.dp_aware {
            // remove dp-aware workers in a prefix-matching fashion
            // without contacting the remote worker
            let mut candidate_workers: Vec<String> = Vec::new();
            let mut removed_workers: Vec<String> = Vec::new();
            let worker_url_prefix = format!("{}@", worker_url);

            {
                // find the candidate workers to be removed
                let workers_guard = self.workers.read().unwrap();
                for w in workers_guard.iter() {
                    if w.url().starts_with(&worker_url_prefix) {
                        candidate_workers.push(w.url().to_string());
                    }
                }
            }

            {
                // do the removing on the worker_urls
                let mut workers_guard = self.workers.write().unwrap();
                for dp_url in candidate_workers.iter() {
                    if let Some(index) = workers_guard.iter().position(|w| w.url() == dp_url) {
                        workers_guard.remove(index);
                        info!("Removed worker: {}", dp_url);
                        removed_workers.push(dp_url.to_string());
                    } else {
                        warn!("Worker {} not found, skipping removal", dp_url);
                        continue;
                    }
                }
                RouterMetrics::set_active_workers(workers_guard.len());
            }

            // If cache aware policy, remove the workers from the tree
            if let Some(cache_aware) = self
                .policy
                .as_any()
                .downcast_ref::<crate::policies::CacheAwarePolicy>()
            {
                for dp_url in removed_workers.iter() {
                    cache_aware.remove_worker(dp_url);
                    info!("Removed worker from tree: {}", dp_url);
                }
            }
        } else {
            let mut workers_guard = self.workers.write().unwrap();
            if let Some(index) = workers_guard.iter().position(|w| w.url() == worker_url) {
                workers_guard.remove(index);
                info!("Removed worker: {}", worker_url);
                RouterMetrics::set_active_workers(workers_guard.len());
            } else {
                warn!("Worker {} not found, skipping removal", worker_url);
                return;
            }

            // If cache aware policy, remove the workers from the tree
            if let Some(cache_aware) = self
                .policy
                .as_any()
                .downcast_ref::<crate::policies::CacheAwarePolicy>()
            {
                cache_aware.remove_worker(worker_url);
                info!("Removed worker from tree: {}", worker_url);
            }
        }
    }

925
    async fn get_worker_load(&self, worker_url: &str) -> Option<isize> {
926
927
928
929
930
931
932
933
934
935
936
937
938
939
        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
        };

940
941
942
943
944
945
        match self
            .client
            .get(&format!("{}/get_load", worker_url))
            .send()
            .await
        {
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
            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>,
983
        client: Client,
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
    ) {
        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> {
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
        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
        };

1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
        match client.get(&format!("{}/get_load", worker_url)).send().await {
            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
            }
        }
    }
}

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)
    }
}

1075
#[async_trait]
1076
1077
1078
1079
1080
impl RouterTrait for Router {
    fn as_any(&self) -> &dyn std::any::Any {
        self
    }

1081
    async fn health(&self, _req: Request<Body>) -> Response {
1082
1083
1084
1085
1086
1087
        let workers = self.workers.read().unwrap();
        let unhealthy_servers: Vec<_> = workers
            .iter()
            .filter(|w| !w.is_healthy())
            .map(|w| w.url().to_string())
            .collect();
1088

1089
1090
        if unhealthy_servers.is_empty() {
            (StatusCode::OK, "All servers healthy").into_response()
1091
        } else {
1092
1093
1094
1095
1096
            (
                StatusCode::SERVICE_UNAVAILABLE,
                format!("Unhealthy servers: {:?}", unhealthy_servers),
            )
                .into_response()
1097
1098
1099
        }
    }

1100
1101
    async fn health_generate(&self, req: Request<Body>) -> Response {
        self.proxy_get_request(req, "health_generate").await
1102
1103
    }

1104
1105
    async fn get_server_info(&self, req: Request<Body>) -> Response {
        self.proxy_get_request(req, "get_server_info").await
1106
1107
    }

1108
1109
    async fn get_models(&self, req: Request<Body>) -> Response {
        self.proxy_get_request(req, "v1/models").await
1110
1111
    }

1112
1113
    async fn get_model_info(&self, req: Request<Body>) -> Response {
        self.proxy_get_request(req, "get_model_info").await
1114
1115
1116
1117
    }

    async fn route_generate(
        &self,
1118
1119
1120
        headers: Option<&HeaderMap>,
        body: &GenerateRequest,
    ) -> Response {
1121
        self.route_typed_request(headers, body, "/generate").await
1122
1123
1124
1125
    }

    async fn route_chat(
        &self,
1126
1127
1128
        headers: Option<&HeaderMap>,
        body: &ChatCompletionRequest,
    ) -> Response {
1129
        self.route_typed_request(headers, body, "/v1/chat/completions")
1130
            .await
1131
1132
1133
1134
    }

    async fn route_completion(
        &self,
1135
1136
1137
        headers: Option<&HeaderMap>,
        body: &CompletionRequest,
    ) -> Response {
1138
        self.route_typed_request(headers, body, "/v1/completions")
1139
            .await
1140
1141
    }

1142
    async fn flush_cache(&self) -> Response {
1143
1144
1145
1146
1147
1148
        // 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 {
1149
1150
1151
1152
1153
1154
            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);
1155
1156
1157
1158
1159
                        return (
                            StatusCode::INTERNAL_SERVER_ERROR,
                            format!("Failed to extract dp_rank: {}", e),
                        )
                            .into_response();
1160
1161
1162
1163
1164
1165
                    }
                };
                worker_url_prefix
            } else {
                worker_url
            };
1166
            let request_builder = self.client.post(format!("{}/flush_cache", worker_url));
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
            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 {
1181
            (StatusCode::OK, "Cache flushed on all servers").into_response()
1182
        } else {
1183
1184
1185
1186
1187
            (
                StatusCode::INTERNAL_SERVER_ERROR,
                "Cache flush failed on one or more servers",
            )
                .into_response()
1188
1189
1190
        }
    }

1191
    async fn get_worker_loads(&self) -> Response {
1192
1193
1194
1195
1196
        let urls = self.get_worker_urls();
        let mut loads = Vec::new();

        // Get loads from all workers
        for url in &urls {
1197
            let load = self.get_worker_load(url).await.unwrap_or(-1);
1198
1199
1200
1201
1202
1203
            loads.push(serde_json::json!({
                "worker": url,
                "load": load
            }));
        }

1204
        Json(serde_json::json!({
1205
1206
            "workers": loads
        }))
1207
        .into_response()
1208
1209
1210
1211
1212
1213
    }

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

1214
    fn readiness(&self) -> Response {
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
        // Regular router is ready if it has at least one healthy worker
        let healthy_count = self
            .workers
            .read()
            .unwrap()
            .iter()
            .filter(|w| w.is_healthy())
            .count();

        if healthy_count > 0 {
1225
            Json(serde_json::json!({
1226
1227
1228
1229
                "status": "ready",
                "healthy_workers": healthy_count,
                "total_workers": self.workers.read().unwrap().len()
            }))
1230
            .into_response()
1231
        } else {
1232
1233
1234
1235
1236
1237
1238
1239
1240
            (
                StatusCode::SERVICE_UNAVAILABLE,
                Json(serde_json::json!({
                    "status": "not_ready",
                    "reason": "no healthy workers available",
                    "total_workers": self.workers.read().unwrap().len()
                })),
            )
                .into_response()
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
        }
    }
}

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

    fn create_test_regular_router() -> Router {
        let workers = vec![
            WorkerFactory::create_regular("http://worker1:8080".to_string()),
            WorkerFactory::create_regular("http://worker2:8080".to_string()),
        ];
        let (_, rx) = tokio::sync::watch::channel(HashMap::new());
        Router {
            workers: Arc::new(RwLock::new(workers)),
            policy: Arc::new(RandomPolicy::new()),
            timeout_secs: 5,
            interval_secs: 1,
1262
1263
            dp_aware: false,
            api_key: None,
1264
            client: Client::new(),
1265
            retry_config: RetryConfig::default(),
1266
            circuit_breaker_config: CircuitBreakerConfig::default(),
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
            _worker_loads: Arc::new(rx),
            _load_monitor_handle: None,
            _health_checker: 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());
        assert_eq!(result.unwrap(), "http://worker1:8080");
    }

1292
1293
1294
1295
    #[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;
1296
        assert!(result.is_err());
1297
        assert!(result.unwrap_err().contains("no workers provided"));
1298
1299
    }

1300
1301
    #[tokio::test]
    async fn test_wait_for_healthy_workers_invalid_urls() {
1302
1303
        // This test will timeout quickly since the URLs are invalid
        let result =
1304
            Router::wait_for_healthy_workers(&["http://nonexistent:8080".to_string()], 1, 1).await;
1305
1306
1307
1308
        assert!(result.is_err());
        assert!(result.unwrap_err().contains("Timeout"));
    }
}