distributed.rs 16.4 KB
Newer Older
1
2
// SPDX-FileCopyrightText: Copyright (c) 2024-2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
// SPDX-License-Identifier: Apache-2.0
Ryan Olson's avatar
Ryan Olson committed
3
4

pub use crate::component::Component;
5
use crate::storage::key_value_store::{EtcdStore, KeyValueStore, MemoryStore};
6
use crate::transports::nats::DRTNatsClientPrometheusMetrics;
Ryan Olson's avatar
Ryan Olson committed
7
use crate::{
8
    ErrorContext, RuntimeCallback,
9
    component::{self, ComponentBuilder, Endpoint, InstanceSource, Namespace},
Ryan Olson's avatar
Ryan Olson committed
10
    discovery::DiscoveryClient,
11
    metrics::MetricsRegistry,
Ryan Olson's avatar
Ryan Olson committed
12
13
14
15
    service::ServiceClient,
    transports::{etcd, nats, tcp},
};

16
use super::utils::GracefulShutdownTracker;
17
use super::{Arc, DistributedRuntime, OK, OnceCell, Result, Runtime, SystemHealth, Weak, error};
18
use std::sync::OnceLock;
Ryan Olson's avatar
Ryan Olson committed
19
20
21

use derive_getters::Dissolve;
use figment::error;
22
23
use std::collections::HashMap;
use tokio::sync::Mutex;
24
use tokio_util::sync::CancellationToken;
Ryan Olson's avatar
Ryan Olson committed
25

26
27
28
29
30
31
32
33
34
35
impl MetricsRegistry for DistributedRuntime {
    fn basename(&self) -> String {
        "".to_string() // drt has no basename. Basename only begins with the Namespace.
    }

    fn parent_hierarchy(&self) -> Vec<String> {
        vec![] // drt is the root, so no parent hierarchy
    }
}

Ryan Olson's avatar
Ryan Olson committed
36
37
38
39
40
41
impl std::fmt::Debug for DistributedRuntime {
    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        write!(f, "DistributedRuntime")
    }
}

Ryan Olson's avatar
Ryan Olson committed
42
43
impl DistributedRuntime {
    pub async fn new(runtime: Runtime, config: DistributedConfig) -> Result<Self> {
44
        let (etcd_config, nats_config, is_static) = config.dissolve();
Ryan Olson's avatar
Ryan Olson committed
45
46
47

        let runtime_clone = runtime.clone();

48
49
50
        let (etcd_client, store) = if is_static {
            let store: Arc<dyn KeyValueStore> = Arc::new(MemoryStore::new());
            (None, store)
51
        } else {
52
53
54
55
            let etcd_client = etcd::Client::new(etcd_config.clone(), runtime_clone).await?;
            let store: Arc<dyn KeyValueStore> = Arc::new(EtcdStore::new(etcd_client.clone()));

            (Some(etcd_client), store)
56
        };
Ryan Olson's avatar
Ryan Olson committed
57

58
        let nats_client = nats_config.clone().connect().await?;
Ryan Olson's avatar
Ryan Olson committed
59

60
        // Start system status server for health and metrics if enabled in configuration
61
62
63
64
65
66
67
68
        let config = crate::config::RuntimeConfig::from_settings().unwrap_or_default();
        // IMPORTANT: We must extract cancel_token from runtime BEFORE moving runtime into the struct below.
        // This is because after moving, runtime is no longer accessible in this scope (ownership rules).
        let cancel_token = if config.system_server_enabled() {
            Some(runtime.clone().child_token())
        } else {
            None
        };
69
70
        let starting_health_status = config.starting_health_status.clone();
        let use_endpoint_health_status = config.use_endpoint_health_status.clone();
71
72
73
        let health_endpoint_path = config.system_health_path.clone();
        let live_endpoint_path = config.system_live_path.clone();
        let system_health = Arc::new(std::sync::Mutex::new(SystemHealth::new(
74
75
            starting_health_status,
            use_endpoint_health_status,
76
77
            health_endpoint_path,
            live_endpoint_path,
78
        )));
79

80
81
        let nats_client_for_metrics = nats_client.clone();

82
        let distributed_runtime = Self {
Ryan Olson's avatar
Ryan Olson committed
83
84
            runtime,
            etcd_client,
85
            store,
Ryan Olson's avatar
Ryan Olson committed
86
87
            nats_client,
            tcp_server: Arc::new(OnceCell::new()),
88
            system_status_server: Arc::new(OnceLock::new()),
Ryan Olson's avatar
Ryan Olson committed
89
            component_registry: component::Registry::new(),
90
            is_static,
91
            instance_sources: Arc::new(Mutex::new(HashMap::new())),
92
            hierarchy_to_metricsregistry: Arc::new(std::sync::RwLock::new(HashMap::<
93
                String,
94
                crate::MetricsRegistryEntry,
95
            >::new())),
96
            system_health,
97
98
        };

99
        let nats_client_metrics = DRTNatsClientPrometheusMetrics::new(
100
101
102
103
104
105
            &distributed_runtime,
            nats_client_for_metrics.client().clone(),
        )?;
        let mut drt_hierarchies = distributed_runtime.parent_hierarchy();
        drt_hierarchies.push(distributed_runtime.hierarchy());
        // Register a callback to update NATS client metrics
106
107
        let nats_client_callback = Arc::new({
            let nats_client_clone = nats_client_metrics.clone();
108
            move || {
109
                nats_client_clone.set_from_client_stats();
110
111
112
                Ok(())
            }
        });
113
        distributed_runtime.register_metrics_callback(drt_hierarchies, nats_client_callback);
114

115
116
117
118
119
120
121
        // Initialize the uptime gauge in SystemHealth
        distributed_runtime
            .system_health
            .lock()
            .unwrap()
            .initialize_uptime_gauge(&distributed_runtime)?;

122
        // Handle system status server initialization
123
        if let Some(cancel_token) = cancel_token {
124
            // System server is enabled - start both the state and HTTP server
125
126
127
            let host = config.system_host.clone();
            let port = config.system_port;

128
            // Start system status server (it creates SystemStatusState internally)
129
            match crate::system_status_server::spawn_system_status_server(
130
131
132
133
                &host,
                port,
                cancel_token,
                Arc::new(distributed_runtime.clone()),
134
135
136
            )
            .await
            {
137
                Ok((addr, handle)) => {
138
                    tracing::info!("System status server started successfully on {}", addr);
139

140
141
142
143
144
145
                    // Store system status server information
                    let system_status_server_info =
                        crate::system_status_server::SystemStatusServerInfo::new(
                            addr,
                            Some(handle),
                        );
146

147
                    // Initialize the system_status_server field
148
                    distributed_runtime
149
150
151
                        .system_status_server
                        .set(Arc::new(system_status_server_info))
                        .expect("System status server info should only be set once");
152
153
                }
                Err(e) => {
154
                    tracing::error!("System status server startup failed: {}", e);
155
                }
156
            }
157
        } else {
158
            // System server HTTP is disabled, but uptime metrics are still being tracked via SystemHealth
159
160
161
            tracing::debug!(
                "System status server HTTP endpoints disabled, but uptime metrics are being tracked"
            );
162
163
        }

164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
        // Start health check manager if enabled
        if config.health_check_enabled {
            let health_check_config = crate::health_check::HealthCheckConfig {
                canary_wait_time: std::time::Duration::from_secs(config.canary_wait_time_secs),
                request_timeout: std::time::Duration::from_secs(
                    config.health_check_request_timeout_secs,
                ),
            };

            // Start the health check manager (spawns per-endpoint monitoring tasks)
            match crate::health_check::start_health_check_manager(
                distributed_runtime.clone(),
                Some(health_check_config),
            )
            .await
            {
                Ok(()) => tracing::info!(
                    "Health check manager started (canary_wait_time: {}s, request_timeout: {}s)",
                    config.canary_wait_time_secs,
                    config.health_check_request_timeout_secs
                ),
                Err(e) => tracing::error!("Health check manager failed to start: {}", e),
            }
        }

189
        Ok(distributed_runtime)
Ryan Olson's avatar
Ryan Olson committed
190
191
192
    }

    pub async fn from_settings(runtime: Runtime) -> Result<Self> {
193
194
195
196
197
198
199
        let config = DistributedConfig::from_settings(false);
        Self::new(runtime, config).await
    }

    // Call this if you are using static workers that do not need etcd-based discovery.
    pub async fn from_settings_without_discovery(runtime: Runtime) -> Result<Self> {
        let config = DistributedConfig::from_settings(true);
Ryan Olson's avatar
Ryan Olson committed
200
201
202
203
204
205
206
        Self::new(runtime, config).await
    }

    pub fn runtime(&self) -> &Runtime {
        &self.runtime
    }

207
208
209
210
    pub fn primary_token(&self) -> CancellationToken {
        self.runtime.primary_token()
    }

211
212
213
214
    /// The etcd lease all our components will be attached to.
    /// Not available for static workers.
    pub fn primary_lease(&self) -> Option<etcd::Lease> {
        self.etcd_client.as_ref().map(|c| c.primary_lease())
Ryan Olson's avatar
Ryan Olson committed
215
216
217
218
219
220
221
222
    }

    pub fn shutdown(&self) {
        self.runtime.shutdown();
    }

    /// Create a [`Namespace`]
    pub fn namespace(&self, name: impl Into<String>) -> Result<Namespace> {
223
        Namespace::new(self.clone(), name.into(), self.is_static)
Ryan Olson's avatar
Ryan Olson committed
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
    }

    // /// Create a [`Component`]
    // pub fn component(
    //     &self,
    //     name: impl Into<String>,
    //     namespace: impl Into<String>,
    // ) -> Result<Component> {
    //     Ok(ComponentBuilder::from_runtime(self.clone())
    //         .name(name.into())
    //         .namespace(namespace.into())
    //         .build()?)
    // }

    pub(crate) fn discovery_client(&self, namespace: impl Into<String>) -> DiscoveryClient {
239
240
241
242
243
244
        DiscoveryClient::new(
            namespace.into(),
            self.etcd_client
                .clone()
                .expect("Attempt to get discovery_client on static DistributedRuntime"),
        )
Ryan Olson's avatar
Ryan Olson committed
245
246
247
248
249
250
    }

    pub(crate) fn service_client(&self) -> ServiceClient {
        ServiceClient::new(self.nats_client.clone())
    }

251
    pub async fn tcp_server(&self) -> Result<Arc<tcp::server::TcpStreamServer>> {
Ryan Olson's avatar
Ryan Olson committed
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
        Ok(self
            .tcp_server
            .get_or_try_init(async move {
                let options = tcp::server::ServerOptions::default();
                let server = tcp::server::TcpStreamServer::new(options).await?;
                OK(server)
            })
            .await?
            .clone())
    }

    pub fn nats_client(&self) -> nats::Client {
        self.nats_client.clone()
    }

267
268
269
270
271
    /// Get system status server information if available
    pub fn system_status_server_info(
        &self,
    ) -> Option<Arc<crate::system_status_server::SystemStatusServerInfo>> {
        self.system_status_server.get().cloned()
272
273
    }

274
    // todo(ryan): deprecate this as we move to Discovery traits and Component Identifiers
275
    pub fn etcd_client(&self) -> Option<etcd::Client> {
Ryan Olson's avatar
Ryan Olson committed
276
277
        self.etcd_client.clone()
    }
278

279
280
281
282
283
284
    /// An interface to store things. Will eventually replace `etcd_client`.
    /// Currently does key-value, but will grow to include whatever we need to store.
    pub fn store(&self) -> Arc<dyn KeyValueStore> {
        self.store.clone()
    }

285
286
287
    pub fn child_token(&self) -> CancellationToken {
        self.runtime.child_token()
    }
288

289
290
291
292
    pub(crate) fn graceful_shutdown_tracker(&self) -> Arc<GracefulShutdownTracker> {
        self.runtime.graceful_shutdown_tracker()
    }

293
294
295
    pub fn instance_sources(&self) -> Arc<Mutex<HashMap<Endpoint, Weak<InstanceSource>>>> {
        self.instance_sources.clone()
    }
296

297
298
    /// Add a Prometheus metric to a specific hierarchy's registry. Note that it is possible
    /// to register the same metric name multiple times, as long as the labels are different.
299
300
301
302
303
304
305
306
    pub fn add_prometheus_metric(
        &self,
        hierarchy: &str,
        prometheus_metric: Box<dyn prometheus::core::Collector>,
    ) -> anyhow::Result<()> {
        let mut registries = self.hierarchy_to_metricsregistry.write().unwrap();
        let entry = registries.entry(hierarchy.to_string()).or_default();

307
308
309
310
311
        // Try to register the metric
        entry
            .prometheus_registry
            .register(prometheus_metric)
            .map_err(|e| e.into())
312
313
314
    }

    /// Add a callback function to metrics registries for the given hierarchies
315
316
    // TODO: Rename to register_metrics_update_callback for consistency with Python API.
    //       Do this after we move the MetricsRegistry trait to composition pattern.
317
318
    pub fn register_metrics_callback(&self, hierarchies: Vec<String>, callback: RuntimeCallback) {
        let mut registries = self.hierarchy_to_metricsregistry.write().unwrap();
319
        for hierarchy in &hierarchies {
320
            registries
321
                .entry(hierarchy.clone())
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
                .or_default()
                .add_callback(callback.clone());
        }
    }

    /// Execute all callbacks for a given hierarchy key and return their results
    pub fn execute_metrics_callbacks(&self, hierarchy: &str) -> Vec<anyhow::Result<()>> {
        // Clone callbacks while holding read lock (fast operation)
        let callbacks = {
            let registries = self.hierarchy_to_metricsregistry.read().unwrap();
            registries
                .get(hierarchy)
                .map(|entry| entry.runtime_callbacks.clone())
        }; // Read lock released here

        // Execute callbacks without holding the lock
        match callbacks {
            Some(callbacks) => callbacks.iter().map(|callback| callback()).collect(),
            None => Vec::new(),
        }
    }

    /// Get all registered hierarchy keys. Private because it is only used for testing.
    fn get_registered_hierarchies(&self) -> Vec<String> {
        let registries = self.hierarchy_to_metricsregistry.read().unwrap();
        registries.keys().cloned().collect()
    }
Ryan Olson's avatar
Ryan Olson committed
349
350
351
352
353
354
}

#[derive(Dissolve)]
pub struct DistributedConfig {
    pub etcd_config: etcd::ClientOptions,
    pub nats_config: nats::ClientOptions,
355
    pub is_static: bool,
Ryan Olson's avatar
Ryan Olson committed
356
357
358
}

impl DistributedConfig {
359
    pub fn from_settings(is_static: bool) -> DistributedConfig {
Ryan Olson's avatar
Ryan Olson committed
360
361
362
        DistributedConfig {
            etcd_config: etcd::ClientOptions::default(),
            nats_config: nats::ClientOptions::default(),
363
            is_static,
Ryan Olson's avatar
Ryan Olson committed
364
365
        }
    }
Ryan Olson's avatar
Ryan Olson committed
366
367
368
369
370

    pub fn for_cli() -> DistributedConfig {
        let mut config = DistributedConfig {
            etcd_config: etcd::ClientOptions::default(),
            nats_config: nats::ClientOptions::default(),
371
            is_static: false,
Ryan Olson's avatar
Ryan Olson committed
372
373
374
375
376
377
        };

        config.etcd_config.attach_lease = false;

        config
    }
Ryan Olson's avatar
Ryan Olson committed
378
}
379

380
pub mod distributed_test_utils {
381
382
383
    //! Common test helper functions for DistributedRuntime tests
    // TODO: Use in-memory DistributedRuntime for tests instead of full runtime when available.

384
    /// Helper function to create a DRT instance for integration-only tests.
385
386
387
388
389
390
391
392
393
394
    /// Uses from_current to leverage existing tokio runtime
    /// Note: Settings are read from environment variables inside DistributedRuntime::from_settings_without_discovery
    #[cfg(feature = "integration")]
    pub async fn create_test_drt_async() -> crate::DistributedRuntime {
        let rt = crate::Runtime::from_current().unwrap();
        crate::DistributedRuntime::from_settings_without_discovery(rt)
            .await
            .unwrap()
    }
}
395

396
#[cfg(all(test, feature = "integration"))]
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
mod tests {
    use super::distributed_test_utils::create_test_drt_async;

    #[tokio::test]
    async fn test_drt_uptime_after_delay_system_disabled() {
        // Test uptime with system status server disabled
        temp_env::async_with_vars([("DYN_SYSTEM_ENABLED", Some("false"))], async {
            // Start a DRT
            let drt = create_test_drt_async().await;

            // Wait 50ms
            tokio::time::sleep(tokio::time::Duration::from_millis(50)).await;

            // Check that uptime is 50+ ms
            let uptime = drt.system_health.lock().unwrap().uptime();
            assert!(
                uptime >= std::time::Duration::from_millis(50),
                "Expected uptime to be at least 50ms, but got {:?}",
                uptime
            );

            println!(
                "✓ DRT uptime test passed (system disabled): uptime = {:?}",
                uptime
            );
        })
        .await;
    }

    #[tokio::test]
    async fn test_drt_uptime_after_delay_system_enabled() {
        // Test uptime with system status server enabled
        temp_env::async_with_vars([("DYN_SYSTEM_ENABLED", Some("true"))], async {
            // Start a DRT
            let drt = create_test_drt_async().await;

            // Wait 50ms
            tokio::time::sleep(tokio::time::Duration::from_millis(50)).await;

            // Check that uptime is 50+ ms
            let uptime = drt.system_health.lock().unwrap().uptime();
            assert!(
                uptime >= std::time::Duration::from_millis(50),
                "Expected uptime to be at least 50ms, but got {:?}",
                uptime
            );

            println!(
                "✓ DRT uptime test passed (system enabled): uptime = {:?}",
                uptime
            );
        })
        .await;
    }
}