| 1 | //! Fleet scheduler policy: leases, heartbeats, backpressure, and recovery. |
| 2 | |
| 3 | #![allow(dead_code)] |
| 4 | |
| 5 | use std::collections::{BTreeMap, BTreeSet}; |
| 6 | use std::path::Path; |
| 7 | use std::time::Duration; |
| 8 | |
| 9 | use anyhow::{Context, Result, anyhow}; |
| 10 | use chrono::{DateTime, SecondsFormat, Utc}; |
| 11 | use codewhale_protocol::fleet::*; |
| 12 | use serde_json::Value; |
| 13 | |
| 14 | use super::ledger::{FleetLedger, FleetLedgerState, FleetTaskLedgerStatus, FleetTaskState}; |
| 15 | |
| 16 | #[derive(Debug, Clone)] |
| 17 | pub struct FleetSchedulerPolicy { |
| 18 | pub max_workers_per_run: usize, |
| 19 | pub max_workers_per_host: usize, |
| 20 | pub max_workers_per_task_class: usize, |
| 21 | pub lease_seconds: u64, |
| 22 | pub heartbeat_timeout: Duration, |
| 23 | } |
| 24 | |
| 25 | impl Default for FleetSchedulerPolicy { |
| 26 | fn default() -> Self { |
| 27 | Self { |
| 28 | max_workers_per_run: 4, |
| 29 | max_workers_per_host: 4, |
| 30 | max_workers_per_task_class: 4, |
| 31 | lease_seconds: 300, |
| 32 | heartbeat_timeout: Duration::from_secs(120), |
| 33 | } |
| 34 | } |
| 35 | } |
| 36 | |
| 37 | #[derive(Debug, Clone, Default, PartialEq, Eq)] |
| 38 | pub struct FleetSchedulerReport { |
| 39 | pub launched: usize, |
| 40 | pub heartbeats: usize, |
| 41 | pub marked_stale: usize, |
| 42 | pub restarted: usize, |
| 43 | pub failed: usize, |
| 44 | pub cancelled: usize, |
| 45 | pub alerts: usize, |
| 46 | } |
| 47 | |
| 48 | #[derive(Debug)] |
| 49 | pub struct FleetScheduler { |
| 50 | ledger: FleetLedger, |
| 51 | policy: FleetSchedulerPolicy, |
| 52 | now: DateTime<Utc>, |
| 53 | } |
| 54 | |
| 55 | impl FleetScheduler { |
| 56 | pub fn open(workspace: impl AsRef<Path>, policy: FleetSchedulerPolicy) -> Result<Self> { |
| 57 | Ok(Self { |
| 58 | ledger: FleetLedger::open(workspace.as_ref())?, |
| 59 | policy, |
| 60 | now: Utc::now(), |
| 61 | }) |
| 62 | } |
| 63 | |
| 64 | pub fn set_now(&mut self, now: DateTime<Utc>) { |
| 65 | self.now = now; |
| 66 | } |
| 67 | |
| 68 | #[cfg(test)] |
| 69 | pub fn tick_run(&self, run_id: &FleetRunId) -> Result<FleetSchedulerReport> { |
| 70 | let mut report = FleetSchedulerReport::default(); |
| 71 | self.recover_unhealthy_work(run_id, &mut report, &mut |_, _, _, _| Ok(()))?; |
| 72 | self.launch_queued_work(run_id, &mut report)?; |
| 73 | self.refresh_run_status(run_id)?; |
| 74 | Ok(report) |
| 75 | } |
| 76 | |
| 77 | /// Resume reconciliation after a manager restart: detect orphaned/stale |
| 78 | /// in-flight leases left by a prior process and apply retry/escalation |
| 79 | /// policy, then recompute run status. |
| 80 | /// |
| 81 | /// Unlike `tick_run`, this launches no new queued work and does not |
| 82 | /// re-process tasks that already reached a terminal state, so it is safe |
| 83 | /// and idempotent to call on a fresh process: a task re-leased by an |
| 84 | /// earlier resume is no longer stale at the same instant, and a terminally |
| 85 | /// failed task is never failed or escalated twice. |
| 86 | #[cfg(test)] |
| 87 | pub fn resume_run(&self, run_id: &FleetRunId) -> Result<FleetSchedulerReport> { |
| 88 | self.resume_run_with_restart_callback(run_id, |_, _, _, _| Ok(())) |
| 89 | } |
| 90 | |
| 91 | pub(crate) fn resume_run_with_restart_callback( |
| 92 | &self, |
| 93 | run_id: &FleetRunId, |
| 94 | mut on_restarting: impl FnMut( |
| 95 | &FleetLedgerState, |
| 96 | &FleetTaskState, |
| 97 | &FleetTaskSpec, |
| 98 | &str, |
| 99 | ) -> Result<()>, |
| 100 | ) -> Result<FleetSchedulerReport> { |
| 101 | let mut report = FleetSchedulerReport::default(); |
| 102 | self.reconcile_stale_leases(run_id, &mut report, &mut on_restarting)?; |
| 103 | self.refresh_run_status(run_id)?; |
| 104 | Ok(report) |
| 105 | } |
| 106 | |
| 107 | pub fn cancel_run(&self, run_id: &FleetRunId, reason: &str) -> Result<FleetSchedulerReport> { |
| 108 | let state = self.ledger.rebuild_state()?; |
| 109 | let mut report = FleetSchedulerReport::default(); |
| 110 | for task in state |
| 111 | .tasks |
| 112 | .values() |
| 113 | .filter(|task| task.entry.run_id == *run_id) |
| 114 | { |
| 115 | if !matches!( |
| 116 | task.status, |
| 117 | FleetTaskLedgerStatus::Enqueued | FleetTaskLedgerStatus::Leased |
| 118 | ) { |
| 119 | continue; |
| 120 | } |
| 121 | if self.ledger.cancel_task_if_active( |
| 122 | &task.entry.run_id, |
| 123 | &task.entry.task_id, |
| 124 | None, |
| 125 | &self.timestamp(), |
| 126 | Some(reason), |
| 127 | Some("scheduler"), |
| 128 | )? { |
| 129 | report.cancelled += 1; |
| 130 | } |
| 131 | } |
| 132 | self.ledger |
| 133 | .update_run_status(run_id, FleetRunStatus::Cancelled, &self.timestamp())?; |
| 134 | Ok(report) |
| 135 | } |
| 136 | |
| 137 | fn recover_unhealthy_work( |
| 138 | &self, |
| 139 | run_id: &FleetRunId, |
| 140 | report: &mut FleetSchedulerReport, |
| 141 | on_restarting: &mut dyn FnMut( |
| 142 | &FleetLedgerState, |
| 143 | &FleetTaskState, |
| 144 | &FleetTaskSpec, |
| 145 | &str, |
| 146 | ) -> Result<()>, |
| 147 | ) -> Result<()> { |
| 148 | let state = self.ledger.rebuild_state()?; |
| 149 | for task in state |
| 150 | .tasks |
| 151 | .values() |
| 152 | .filter(|task| task.entry.run_id == *run_id) |
| 153 | { |
| 154 | let Some(task_spec) = task_spec_for(&state, task) else { |
| 155 | continue; |
| 156 | }; |
| 157 | match task.status { |
| 158 | FleetTaskLedgerStatus::Leased if self.task_is_stale(task, &state) => { |
| 159 | let worker_id = task |
| 160 | .leased_to |
| 161 | .clone() |
| 162 | .unwrap_or_else(|| "fleet-scheduler".to_string()); |
| 163 | let heartbeat_at = state |
| 164 | .heartbeats |
| 165 | .get(&worker_id) |
| 166 | .map(|heartbeat| heartbeat.timestamp.as_str()); |
| 167 | let latest_seq = state |
| 168 | .latest_seq |
| 169 | .get(&format!( |
| 170 | "{}:{}:{}", |
| 171 | worker_id, task.entry.run_id.0, task.entry.task_id |
| 172 | )) |
| 173 | .copied() |
| 174 | .unwrap_or(0); |
| 175 | let Some(stale_event) = self.ledger.append_event_if_lease_unchanged( |
| 176 | &task.entry.run_id, |
| 177 | &worker_id, |
| 178 | &task.entry.task_id, |
| 179 | task.entry.attempts, |
| 180 | latest_seq, |
| 181 | heartbeat_at, |
| 182 | &self.timestamp(), |
| 183 | FleetWorkerEventPayload::Stale { |
| 184 | last_heartbeat_at: heartbeat_at.map(str::to_string), |
| 185 | }, |
| 186 | )? |
| 187 | else { |
| 188 | continue; |
| 189 | }; |
| 190 | report.marked_stale += 1; |
| 191 | self.retry_or_fail( |
| 192 | &state, |
| 193 | task, |
| 194 | &task_spec, |
| 195 | &worker_id, |
| 196 | stale_event.seq, |
| 197 | heartbeat_at, |
| 198 | report, |
| 199 | on_restarting, |
| 200 | ) |
| 201 | .with_context(|| format!("recovering stale task {}", task.entry.task_id))?; |
| 202 | } |
| 203 | FleetTaskLedgerStatus::Failed => { |
| 204 | let worker_id = task |
| 205 | .leased_to |
| 206 | .clone() |
| 207 | .unwrap_or_else(|| "fleet-scheduler".to_string()); |
| 208 | let latest_seq = state |
| 209 | .latest_seq |
| 210 | .get(&format!( |
| 211 | "{}:{}:{}", |
| 212 | worker_id, task.entry.run_id.0, task.entry.task_id |
| 213 | )) |
| 214 | .copied() |
| 215 | .unwrap_or(0); |
| 216 | let heartbeat_at = state |
| 217 | .heartbeats |
| 218 | .get(&worker_id) |
| 219 | .map(|heartbeat| heartbeat.timestamp.as_str()); |
| 220 | self.retry_or_fail( |
| 221 | &state, |
| 222 | task, |
| 223 | &task_spec, |
| 224 | &worker_id, |
| 225 | latest_seq, |
| 226 | heartbeat_at, |
| 227 | report, |
| 228 | on_restarting, |
| 229 | ) |
| 230 | .with_context(|| format!("recovering failed task {}", task.entry.task_id))?; |
| 231 | } |
| 232 | _ => {} |
| 233 | } |
| 234 | } |
| 235 | Ok(()) |
| 236 | } |
| 237 | |
| 238 | /// Reconcile only orphaned/stale in-flight leases (the restart-recovery |
| 239 | /// subset of `recover_unhealthy_work`): a `Leased` task whose worker has |
| 240 | /// stopped heartbeating is marked stale and routed through the shared |
| 241 | /// retry/escalation budget. Terminal and healthy tasks are left untouched, |
| 242 | /// which keeps `resume_run` idempotent. |
| 243 | fn reconcile_stale_leases( |
| 244 | &self, |
| 245 | run_id: &FleetRunId, |
| 246 | report: &mut FleetSchedulerReport, |
| 247 | on_restarting: &mut dyn FnMut( |
| 248 | &FleetLedgerState, |
| 249 | &FleetTaskState, |
| 250 | &FleetTaskSpec, |
| 251 | &str, |
| 252 | ) -> Result<()>, |
| 253 | ) -> Result<()> { |
| 254 | let state = self.ledger.rebuild_state()?; |
| 255 | for task in state |
| 256 | .tasks |
| 257 | .values() |
| 258 | .filter(|task| task.entry.run_id == *run_id) |
| 259 | { |
| 260 | if !matches!(task.status, FleetTaskLedgerStatus::Leased) |
| 261 | || !self.task_is_stale(task, &state) |
| 262 | { |
| 263 | continue; |
| 264 | } |
| 265 | let Some(task_spec) = task_spec_for(&state, task) else { |
| 266 | continue; |
| 267 | }; |
| 268 | let worker_id = task |
| 269 | .leased_to |
| 270 | .clone() |
| 271 | .unwrap_or_else(|| "fleet-scheduler".to_string()); |
| 272 | let heartbeat_at = state |
| 273 | .heartbeats |
| 274 | .get(&worker_id) |
| 275 | .map(|heartbeat| heartbeat.timestamp.as_str()); |
| 276 | let latest_seq = state |
| 277 | .latest_seq |
| 278 | .get(&format!( |
| 279 | "{}:{}:{}", |
| 280 | worker_id, task.entry.run_id.0, task.entry.task_id |
| 281 | )) |
| 282 | .copied() |
| 283 | .unwrap_or(0); |
| 284 | let Some(stale_event) = self.ledger.append_event_if_lease_unchanged( |
| 285 | &task.entry.run_id, |
| 286 | &worker_id, |
| 287 | &task.entry.task_id, |
| 288 | task.entry.attempts, |
| 289 | latest_seq, |
| 290 | heartbeat_at, |
| 291 | &self.timestamp(), |
| 292 | FleetWorkerEventPayload::Stale { |
| 293 | last_heartbeat_at: heartbeat_at.map(str::to_string), |
| 294 | }, |
| 295 | )? |
| 296 | else { |
| 297 | continue; |
| 298 | }; |
| 299 | report.marked_stale += 1; |
| 300 | self.retry_or_fail( |
| 301 | &state, |
| 302 | task, |
| 303 | &task_spec, |
| 304 | &worker_id, |
| 305 | stale_event.seq, |
| 306 | heartbeat_at, |
| 307 | report, |
| 308 | on_restarting, |
| 309 | ) |
| 310 | .with_context(|| format!("resuming stale task {}", task.entry.task_id))?; |
| 311 | } |
| 312 | Ok(()) |
| 313 | } |
| 314 | |
| 315 | #[allow(clippy::too_many_arguments)] |
| 316 | fn retry_or_fail( |
| 317 | &self, |
| 318 | state: &FleetLedgerState, |
| 319 | task: &FleetTaskState, |
| 320 | task_spec: &FleetTaskSpec, |
| 321 | worker_id: &str, |
| 322 | expected_latest_seq: u64, |
| 323 | expected_heartbeat_at: Option<&str>, |
| 324 | report: &mut FleetSchedulerReport, |
| 325 | on_restarting: &mut dyn FnMut( |
| 326 | &FleetLedgerState, |
| 327 | &FleetTaskState, |
| 328 | &FleetTaskSpec, |
| 329 | &str, |
| 330 | ) -> Result<()>, |
| 331 | ) -> Result<()> { |
| 332 | let retry_policy = task_spec.retry_policy.clone().unwrap_or_default(); |
| 333 | if task.entry.attempts < retry_policy.max_attempts { |
| 334 | let lease_expires_at = self.lease_expires_at(); |
| 335 | if !self.ledger.restart_task_if_unchanged_with_callback( |
| 336 | &task.entry.run_id, |
| 337 | &task.entry.task_id, |
| 338 | worker_id, |
| 339 | task.status, |
| 340 | task.entry.attempts, |
| 341 | expected_latest_seq, |
| 342 | expected_heartbeat_at, |
| 343 | &self.timestamp(), |
| 344 | Some(&lease_expires_at), |
| 345 | task.entry.attempts, |
| 346 | || on_restarting(state, task, task_spec, worker_id), |
| 347 | )? { |
| 348 | return Ok(()); |
| 349 | } |
| 350 | report.restarted += 1; |
| 351 | return Ok(()); |
| 352 | } |
| 353 | |
| 354 | if task.status == FleetTaskLedgerStatus::Failed { |
| 355 | report.alerts += self.record_alerts( |
| 356 | &task.entry.run_id, |
| 357 | &task.entry.task_id, |
| 358 | worker_id, |
| 359 | task.entry.attempts, |
| 360 | task_spec, |
| 361 | FleetAlertEventClass::RestartExhausted, |
| 362 | )?; |
| 363 | return Ok(()); |
| 364 | } |
| 365 | |
| 366 | let terminal = self.ledger.append_terminal_event_if_lease_unchanged( |
| 367 | &task.entry.run_id, |
| 368 | worker_id, |
| 369 | &task.entry.task_id, |
| 370 | task.entry.attempts, |
| 371 | expected_latest_seq, |
| 372 | expected_heartbeat_at, |
| 373 | &self.timestamp(), |
| 374 | FleetWorkerEventPayload::Failed { |
| 375 | reason: format!( |
| 376 | "retry attempts exhausted after {} attempt(s)", |
| 377 | task.entry.attempts |
| 378 | ), |
| 379 | recoverable: false, |
| 380 | }, |
| 381 | )?; |
| 382 | if terminal.is_none() { |
| 383 | return Ok(()); |
| 384 | } |
| 385 | report.failed += 1; |
| 386 | report.alerts += self.record_alerts( |
| 387 | &task.entry.run_id, |
| 388 | &task.entry.task_id, |
| 389 | worker_id, |
| 390 | task.entry.attempts, |
| 391 | task_spec, |
| 392 | FleetAlertEventClass::RestartExhausted, |
| 393 | )?; |
| 394 | Ok(()) |
| 395 | } |
| 396 | |
| 397 | fn launch_queued_work( |
| 398 | &self, |
| 399 | run_id: &FleetRunId, |
| 400 | report: &mut FleetSchedulerReport, |
| 401 | ) -> Result<()> { |
| 402 | loop { |
| 403 | let state = self.ledger.rebuild_state()?; |
| 404 | let run = state |
| 405 | .runs |
| 406 | .get(&run_id.0) |
| 407 | .ok_or_else(|| anyhow!("Fleet run {} does not exist", run_id.0))?; |
| 408 | let active = active_tasks_for_run(&state, run_id); |
| 409 | if active.len() >= self.policy.max_workers_per_run { |
| 410 | return Ok(()); |
| 411 | } |
| 412 | let counts = active_counts(&state, run); |
| 413 | let Some((worker_id, task)) = self.next_launch(run, &state, &counts) else { |
| 414 | return Ok(()); |
| 415 | }; |
| 416 | let lease_expires_at = self.lease_expires_at(); |
| 417 | if !self.ledger.start_task_if_enqueued( |
| 418 | &task.entry.run_id, |
| 419 | &task.entry.task_id, |
| 420 | &worker_id, |
| 421 | &self.timestamp(), |
| 422 | Some(&lease_expires_at), |
| 423 | Some(self.policy.max_workers_per_run), |
| 424 | vec![ |
| 425 | FleetWorkerEventPayload::Leased { |
| 426 | lease_expires_at: Some(lease_expires_at.clone()), |
| 427 | }, |
| 428 | FleetWorkerEventPayload::Starting, |
| 429 | FleetWorkerEventPayload::Running, |
| 430 | ], |
| 431 | || Ok(()), |
| 432 | )? { |
| 433 | continue; |
| 434 | } |
| 435 | report.launched += 1; |
| 436 | report.heartbeats += 1; |
| 437 | } |
| 438 | } |
| 439 | |
| 440 | fn next_launch( |
| 441 | &self, |
| 442 | run: &FleetRun, |
| 443 | state: &FleetLedgerState, |
| 444 | counts: &ActiveCounts, |
| 445 | ) -> Option<(String, FleetTaskState)> { |
| 446 | let active_workers: BTreeSet<_> = active_tasks_for_run(state, &run.id) |
| 447 | .into_iter() |
| 448 | .filter_map(|task| task.leased_to) |
| 449 | .collect(); |
| 450 | let mut queued: Vec<_> = state |
| 451 | .tasks |
| 452 | .values() |
| 453 | .filter(|task| { |
| 454 | task.entry.run_id == run.id |
| 455 | && matches!(task.status, FleetTaskLedgerStatus::Enqueued) |
| 456 | }) |
| 457 | .cloned() |
| 458 | .collect(); |
| 459 | queued.sort_by_key(|task| { |
| 460 | ( |
| 461 | task.entry.priority, |
| 462 | task.entry.enqueued_at.clone(), |
| 463 | task.entry.task_id.clone(), |
| 464 | ) |
| 465 | }); |
| 466 | for task in queued { |
| 467 | let task_spec = run |
| 468 | .task_specs |
| 469 | .iter() |
| 470 | .find(|spec| spec.id == task.entry.task_id)?; |
| 471 | let task_class = task_class(task_spec); |
| 472 | if counts.by_task_class.get(&task_class).copied().unwrap_or(0) |
| 473 | >= self.policy.max_workers_per_task_class |
| 474 | { |
| 475 | continue; |
| 476 | } |
| 477 | for worker in &run.worker_specs { |
| 478 | if active_workers.contains(&worker.id) { |
| 479 | continue; |
| 480 | } |
| 481 | let host_key = host_key(worker); |
| 482 | if counts.by_host.get(&host_key).copied().unwrap_or(0) |
| 483 | >= self.policy.max_workers_per_host |
| 484 | { |
| 485 | continue; |
| 486 | } |
| 487 | return Some((worker.id.clone(), task)); |
| 488 | } |
| 489 | } |
| 490 | None |
| 491 | } |
| 492 | |
| 493 | fn task_is_stale(&self, task: &FleetTaskState, state: &FleetLedgerState) -> bool { |
| 494 | if let Some(worker_id) = task.leased_to.as_deref() |
| 495 | && let Some(heartbeat) = state.heartbeats.get(worker_id) |
| 496 | && let Ok(last) = DateTime::parse_from_rfc3339(&heartbeat.timestamp) |
| 497 | { |
| 498 | let age = self.now.signed_duration_since(last.with_timezone(&Utc)); |
| 499 | return age |
| 500 | .to_std() |
| 501 | .map_or(true, |age| age > self.policy.heartbeat_timeout); |
| 502 | } |
| 503 | if let Some(deadline) = task.entry.lease_deadline.as_deref() |
| 504 | && let Ok(deadline) = DateTime::parse_from_rfc3339(deadline) |
| 505 | { |
| 506 | return self.now > deadline.with_timezone(&Utc); |
| 507 | } |
| 508 | true |
| 509 | } |
| 510 | |
| 511 | fn record_alerts( |
| 512 | &self, |
| 513 | run_id: &FleetRunId, |
| 514 | task_id: &str, |
| 515 | worker_id: &str, |
| 516 | expected_attempts: u32, |
| 517 | task_spec: &FleetTaskSpec, |
| 518 | event_class: FleetAlertEventClass, |
| 519 | ) -> Result<usize> { |
| 520 | let Some(policy) = &task_spec.alert_policy else { |
| 521 | return Ok(0); |
| 522 | }; |
| 523 | if !alert_policy_matches(policy, event_class) { |
| 524 | return Ok(0); |
| 525 | } |
| 526 | let mut count = 0; |
| 527 | for (channel_index, channel) in policy.channels.iter().enumerate() { |
| 528 | let label = alert_channel_label(channel); |
| 529 | // A policy may contain multiple endpoints of the same kind. The |
| 530 | // run snapshot preserves channel order while redacting secrets, so |
| 531 | // a kind + ordinal key is stable, non-secret, and instance-unique. |
| 532 | let channel_key = format!("{label}#{channel_index}"); |
| 533 | if self.ledger.record_failed_attempt_alert_once( |
| 534 | run_id, |
| 535 | task_id, |
| 536 | worker_id, |
| 537 | expected_attempts, |
| 538 | label, |
| 539 | &channel_key, |
| 540 | &self.timestamp(), |
| 541 | )? { |
| 542 | count += 1; |
| 543 | } |
| 544 | } |
| 545 | Ok(count) |
| 546 | } |
| 547 | |
| 548 | fn refresh_run_status(&self, run_id: &FleetRunId) -> Result<()> { |
| 549 | let state = self.ledger.rebuild_state()?; |
| 550 | let mut has_open = false; |
| 551 | let mut has_failed = false; |
| 552 | let mut has_cancelled = false; |
| 553 | for task in state |
| 554 | .tasks |
| 555 | .values() |
| 556 | .filter(|task| task.entry.run_id == *run_id) |
| 557 | { |
| 558 | match task.status { |
| 559 | FleetTaskLedgerStatus::Enqueued | FleetTaskLedgerStatus::Leased => has_open = true, |
| 560 | FleetTaskLedgerStatus::Failed => has_failed = true, |
| 561 | FleetTaskLedgerStatus::Cancelled => has_cancelled = true, |
| 562 | FleetTaskLedgerStatus::Completed => {} |
| 563 | } |
| 564 | } |
| 565 | let status = if has_open { |
| 566 | FleetRunStatus::Running |
| 567 | } else if has_failed { |
| 568 | FleetRunStatus::Failed |
| 569 | } else if has_cancelled { |
| 570 | FleetRunStatus::Cancelled |
| 571 | } else { |
| 572 | FleetRunStatus::Completed |
| 573 | }; |
| 574 | self.ledger |
| 575 | .update_run_status(run_id, status, &self.timestamp()) |
| 576 | } |
| 577 | |
| 578 | fn append_worker_event( |
| 579 | &self, |
| 580 | run_id: &FleetRunId, |
| 581 | worker_id: &str, |
| 582 | task_id: &str, |
| 583 | payload: FleetWorkerEventPayload, |
| 584 | ) -> Result<FleetWorkerEvent> { |
| 585 | self.ledger |
| 586 | .append_event_next_seq(run_id, worker_id, task_id, &self.timestamp(), payload) |
| 587 | } |
| 588 | |
| 589 | fn timestamp(&self) -> String { |
| 590 | self.now.to_rfc3339_opts(SecondsFormat::Secs, true) |
| 591 | } |
| 592 | |
| 593 | fn lease_expires_at(&self) -> String { |
| 594 | (self.now + chrono::Duration::seconds(self.policy.lease_seconds as i64)) |
| 595 | .to_rfc3339_opts(SecondsFormat::Secs, true) |
| 596 | } |
| 597 | } |
| 598 | |
| 599 | #[derive(Debug, Default)] |
| 600 | struct ActiveCounts { |
| 601 | by_host: BTreeMap<String, usize>, |
| 602 | by_task_class: BTreeMap<String, usize>, |
| 603 | } |
| 604 | |
| 605 | fn active_counts(state: &FleetLedgerState, run: &FleetRun) -> ActiveCounts { |
| 606 | let mut counts = ActiveCounts::default(); |
| 607 | for task in active_tasks_for_run(state, &run.id) { |
| 608 | if let Some(worker_id) = task.leased_to.as_deref() |
| 609 | && let Some(worker) = run |
| 610 | .worker_specs |
| 611 | .iter() |
| 612 | .find(|worker| worker.id == worker_id) |
| 613 | { |
| 614 | *counts.by_host.entry(host_key(worker)).or_default() += 1; |
| 615 | } |
| 616 | if let Some(task_spec) = run |
| 617 | .task_specs |
| 618 | .iter() |
| 619 | .find(|spec| spec.id == task.entry.task_id) |
| 620 | { |
| 621 | *counts |
| 622 | .by_task_class |
| 623 | .entry(task_class(task_spec)) |
| 624 | .or_default() += 1; |
| 625 | } |
| 626 | } |
| 627 | counts |
| 628 | } |
| 629 | |
| 630 | fn active_tasks_for_run(state: &FleetLedgerState, run_id: &FleetRunId) -> Vec<FleetTaskState> { |
| 631 | state |
| 632 | .tasks |
| 633 | .values() |
| 634 | .filter(|task| { |
| 635 | task.entry.run_id == *run_id && matches!(task.status, FleetTaskLedgerStatus::Leased) |
| 636 | }) |
| 637 | .cloned() |
| 638 | .collect() |
| 639 | } |
| 640 | |
| 641 | fn task_spec_for(state: &FleetLedgerState, task: &FleetTaskState) -> Option<FleetTaskSpec> { |
| 642 | state |
| 643 | .runs |
| 644 | .get(&task.entry.run_id.0)? |
| 645 | .task_specs |
| 646 | .iter() |
| 647 | .find(|spec| spec.id == task.entry.task_id) |
| 648 | .cloned() |
| 649 | } |
| 650 | |
| 651 | fn host_key(worker: &FleetWorkerSpec) -> String { |
| 652 | match &worker.host { |
| 653 | FleetHostSpec::Local => "local".to_string(), |
| 654 | FleetHostSpec::Ssh { host, .. } => format!("ssh:{host}"), |
| 655 | FleetHostSpec::Docker { image, .. } => format!("docker:{image}"), |
| 656 | } |
| 657 | } |
| 658 | |
| 659 | fn task_class(task: &FleetTaskSpec) -> String { |
| 660 | task.metadata |
| 661 | .get("class") |
| 662 | .and_then(Value::as_str) |
| 663 | .filter(|value| !value.trim().is_empty()) |
| 664 | .unwrap_or("default") |
| 665 | .to_string() |
| 666 | } |
| 667 | |
| 668 | fn alert_channel_label(channel: &FleetAlertChannel) -> &'static str { |
| 669 | match channel { |
| 670 | FleetAlertChannel::Slack { .. } => "slack", |
| 671 | FleetAlertChannel::Webhook { .. } => "webhook", |
| 672 | FleetAlertChannel::PagerDuty { .. } => "pagerduty", |
| 673 | } |
| 674 | } |
| 675 | |
| 676 | fn alert_policy_matches(policy: &FleetAlertPolicy, class: FleetAlertEventClass) -> bool { |
| 677 | policy.events.is_empty() || policy.events.contains(&class) |
| 678 | } |
| 679 | |
| 680 | #[cfg(test)] |
| 681 | mod tests { |
| 682 | use super::*; |
| 683 | use std::sync::{Arc, Barrier}; |
| 684 | use std::thread; |
| 685 | use tempfile::TempDir; |
| 686 | |
| 687 | fn base_now() -> DateTime<Utc> { |
| 688 | DateTime::parse_from_rfc3339("2026-06-13T01:00:00Z") |
| 689 | .unwrap() |
| 690 | .with_timezone(&Utc) |
| 691 | } |
| 692 | |
| 693 | fn scheduler(tmp: &TempDir, max_workers: usize) -> FleetScheduler { |
| 694 | let mut scheduler = FleetScheduler::open( |
| 695 | tmp.path(), |
| 696 | FleetSchedulerPolicy { |
| 697 | max_workers_per_run: max_workers, |
| 698 | max_workers_per_host: max_workers, |
| 699 | max_workers_per_task_class: max_workers, |
| 700 | lease_seconds: 30, |
| 701 | heartbeat_timeout: Duration::from_secs(5), |
| 702 | }, |
| 703 | ) |
| 704 | .unwrap(); |
| 705 | scheduler.set_now(base_now()); |
| 706 | scheduler |
| 707 | } |
| 708 | |
| 709 | fn worker(id: &str) -> FleetWorkerSpec { |
| 710 | FleetWorkerSpec { |
| 711 | id: id.to_string(), |
| 712 | name: id.to_string(), |
| 713 | host: FleetHostSpec::Local, |
| 714 | trust_level: Some(FleetTrustLevel::Local), |
| 715 | labels: BTreeMap::new(), |
| 716 | capabilities: vec!["local".to_string()], |
| 717 | max_concurrent_tasks: Some(1), |
| 718 | } |
| 719 | } |
| 720 | |
| 721 | fn task(id: &str, max_attempts: u32) -> FleetTaskSpec { |
| 722 | FleetTaskSpec { |
| 723 | id: id.to_string(), |
| 724 | name: id.to_string(), |
| 725 | description: None, |
| 726 | objective: Some(format!("Schedule {id}")), |
| 727 | instructions: format!("do {id}"), |
| 728 | worker: None, |
| 729 | workspace: None, |
| 730 | input_files: Vec::new(), |
| 731 | context: Vec::new(), |
| 732 | budget: None, |
| 733 | tags: Vec::new(), |
| 734 | expected_artifacts: vec![FleetArtifactKind::Log], |
| 735 | scorer: None, |
| 736 | retry_policy: Some(FleetRetryPolicy { |
| 737 | max_attempts, |
| 738 | ..FleetRetryPolicy::default() |
| 739 | }), |
| 740 | alert_policy: None, |
| 741 | timeout_seconds: None, |
| 742 | metadata: BTreeMap::new(), |
| 743 | } |
| 744 | } |
| 745 | |
| 746 | fn create_run( |
| 747 | scheduler: &FleetScheduler, |
| 748 | run_id: &str, |
| 749 | tasks: Vec<FleetTaskSpec>, |
| 750 | workers: usize, |
| 751 | ) { |
| 752 | let run_id = FleetRunId::from(run_id); |
| 753 | scheduler |
| 754 | .ledger |
| 755 | .create_run(&FleetRun { |
| 756 | id: run_id.clone(), |
| 757 | name: "scheduler smoke".to_string(), |
| 758 | status: FleetRunStatus::Queued, |
| 759 | target: None, |
| 760 | workflow: None, |
| 761 | roles: Vec::new(), |
| 762 | max_workers: Some(workers), |
| 763 | usage_ceiling: None, |
| 764 | task_specs: tasks.clone(), |
| 765 | worker_specs: (1..=workers) |
| 766 | .map(|idx| worker(&format!("worker-{idx}"))) |
| 767 | .collect(), |
| 768 | labels: BTreeMap::new(), |
| 769 | security_policy: None, |
| 770 | created_at: scheduler.timestamp(), |
| 771 | updated_at: None, |
| 772 | completed_at: None, |
| 773 | }) |
| 774 | .unwrap(); |
| 775 | for task in tasks { |
| 776 | scheduler |
| 777 | .ledger |
| 778 | .enqueue(FleetInboxEntry { |
| 779 | run_id: run_id.clone(), |
| 780 | task_id: task.id, |
| 781 | priority: 0, |
| 782 | enqueued_at: scheduler.timestamp(), |
| 783 | lease_deadline: None, |
| 784 | attempts: 0, |
| 785 | }) |
| 786 | .unwrap(); |
| 787 | } |
| 788 | } |
| 789 | |
| 790 | fn ledger_text(scheduler: &FleetScheduler) -> String { |
| 791 | std::fs::read_to_string(scheduler.ledger.path()).unwrap() |
| 792 | } |
| 793 | |
| 794 | #[test] |
| 795 | fn fleet_scheduler_backpressure_prevents_over_launch() { |
| 796 | let tmp = TempDir::new().unwrap(); |
| 797 | let scheduler = scheduler(&tmp, 2); |
| 798 | create_run( |
| 799 | &scheduler, |
| 800 | "run-1", |
| 801 | vec![task("task-a", 3), task("task-b", 3), task("task-c", 3)], |
| 802 | 3, |
| 803 | ); |
| 804 | |
| 805 | let report = scheduler.tick_run(&FleetRunId::from("run-1")).unwrap(); |
| 806 | |
| 807 | assert_eq!(report.launched, 2); |
| 808 | let state = scheduler.ledger.rebuild_state().unwrap(); |
| 809 | assert_eq!( |
| 810 | state.tasks["run-1:task-a"].status, |
| 811 | FleetTaskLedgerStatus::Leased |
| 812 | ); |
| 813 | assert_eq!( |
| 814 | state.tasks["run-1:task-b"].status, |
| 815 | FleetTaskLedgerStatus::Leased |
| 816 | ); |
| 817 | assert_eq!( |
| 818 | state.tasks["run-1:task-c"].status, |
| 819 | FleetTaskLedgerStatus::Enqueued |
| 820 | ); |
| 821 | } |
| 822 | |
| 823 | #[test] |
| 824 | fn fleet_scheduler_lost_heartbeat_restarts_within_retry_limit() { |
| 825 | let tmp = TempDir::new().unwrap(); |
| 826 | let mut scheduler = scheduler(&tmp, 1); |
| 827 | create_run(&scheduler, "run-1", vec![task("task-a", 2)], 1); |
| 828 | scheduler.tick_run(&FleetRunId::from("run-1")).unwrap(); |
| 829 | scheduler.set_now(base_now() + chrono::Duration::seconds(10)); |
| 830 | |
| 831 | let report = scheduler.tick_run(&FleetRunId::from("run-1")).unwrap(); |
| 832 | |
| 833 | assert_eq!(report.marked_stale, 1); |
| 834 | assert_eq!(report.restarted, 1); |
| 835 | let state = scheduler.ledger.rebuild_state().unwrap(); |
| 836 | let task = &state.tasks["run-1:task-a"]; |
| 837 | assert_eq!(task.status, FleetTaskLedgerStatus::Leased); |
| 838 | assert_eq!(task.entry.attempts, 2); |
| 839 | let ledger = ledger_text(&scheduler); |
| 840 | assert!(ledger.contains("\"state\":\"stale\"")); |
| 841 | assert!(ledger.contains("\"state\":\"restarted\"")); |
| 842 | } |
| 843 | |
| 844 | #[test] |
| 845 | fn fleet_scheduler_restart_exhaustion_records_terminal_failure_and_alert() { |
| 846 | let tmp = TempDir::new().unwrap(); |
| 847 | let mut scheduler = scheduler(&tmp, 1); |
| 848 | let mut failing = task("task-a", 1); |
| 849 | failing.alert_policy = Some(FleetAlertPolicy { |
| 850 | events: vec![FleetAlertEventClass::RestartExhausted], |
| 851 | channels: vec![ |
| 852 | FleetAlertChannel::Slack { |
| 853 | webhook: FleetAlertEndpoint::inline("https://hooks.slack.invalid/secret-a"), |
| 854 | }, |
| 855 | FleetAlertChannel::Slack { |
| 856 | webhook: FleetAlertEndpoint::inline("https://hooks.slack.invalid/secret-b"), |
| 857 | }, |
| 858 | ], |
| 859 | after_attempts: Some(1), |
| 860 | after_minutes_stale: Some(1), |
| 861 | }); |
| 862 | create_run(&scheduler, "run-1", vec![failing], 1); |
| 863 | scheduler.tick_run(&FleetRunId::from("run-1")).unwrap(); |
| 864 | scheduler.set_now(base_now() + chrono::Duration::seconds(10)); |
| 865 | |
| 866 | let report = scheduler.tick_run(&FleetRunId::from("run-1")).unwrap(); |
| 867 | |
| 868 | assert_eq!(report.marked_stale, 1); |
| 869 | assert_eq!(report.restarted, 0); |
| 870 | assert_eq!(report.failed, 1); |
| 871 | assert_eq!(report.alerts, 2); |
| 872 | let state = scheduler.ledger.rebuild_state().unwrap(); |
| 873 | assert_eq!( |
| 874 | state.tasks["run-1:task-a"].status, |
| 875 | FleetTaskLedgerStatus::Failed |
| 876 | ); |
| 877 | assert_eq!(state.alerts.len(), 2); |
| 878 | assert_eq!(state.escalated_events.len(), 1); |
| 879 | let ledger = ledger_text(&scheduler); |
| 880 | assert!(ledger.contains("\"state\":\"failed\"")); |
| 881 | assert!(ledger.contains("\"record\":\"alert_sent\"")); |
| 882 | assert_eq!( |
| 883 | ledger |
| 884 | .lines() |
| 885 | .filter(|line| line.contains("\"record\":\"alert_sent\"")) |
| 886 | .count(), |
| 887 | 2 |
| 888 | ); |
| 889 | assert!(!ledger.contains("hooks.slack.invalid/secret")); |
| 890 | } |
| 891 | |
| 892 | #[test] |
| 893 | fn exhausted_failed_attempt_alerts_exactly_once_across_competing_schedulers() { |
| 894 | let tmp = TempDir::new().unwrap(); |
| 895 | let owner = scheduler(&tmp, 1); |
| 896 | let mut failing = task("task-a", 1); |
| 897 | failing.alert_policy = Some(FleetAlertPolicy { |
| 898 | events: vec![FleetAlertEventClass::RestartExhausted], |
| 899 | channels: vec![FleetAlertChannel::Slack { |
| 900 | webhook: FleetAlertEndpoint::inline("https://hooks.slack.invalid/secret"), |
| 901 | }], |
| 902 | after_attempts: Some(1), |
| 903 | after_minutes_stale: Some(1), |
| 904 | }); |
| 905 | create_run(&owner, "run-1", vec![failing], 1); |
| 906 | owner.tick_run(&FleetRunId::from("run-1")).unwrap(); |
| 907 | owner |
| 908 | .ledger |
| 909 | .append_terminal_event_if_leased( |
| 910 | &FleetRunId::from("run-1"), |
| 911 | "worker-1", |
| 912 | "task-a", |
| 913 | 1, |
| 914 | &owner.timestamp(), |
| 915 | FleetWorkerEventPayload::Failed { |
| 916 | reason: "worker failed before scheduler recovery".to_string(), |
| 917 | recoverable: true, |
| 918 | }, |
| 919 | ) |
| 920 | .unwrap() |
| 921 | .unwrap(); |
| 922 | |
| 923 | let root = tmp.path().to_path_buf(); |
| 924 | let barrier = Arc::new(Barrier::new(2)); |
| 925 | let handles = (0..2) |
| 926 | .map(|_| { |
| 927 | let root = root.clone(); |
| 928 | let barrier = Arc::clone(&barrier); |
| 929 | thread::spawn(move || { |
| 930 | let mut contender = FleetScheduler::open( |
| 931 | root, |
| 932 | FleetSchedulerPolicy { |
| 933 | max_workers_per_run: 1, |
| 934 | max_workers_per_host: 1, |
| 935 | max_workers_per_task_class: 1, |
| 936 | lease_seconds: 30, |
| 937 | heartbeat_timeout: Duration::from_secs(5), |
| 938 | }, |
| 939 | ) |
| 940 | .unwrap(); |
| 941 | contender.set_now(base_now() + chrono::Duration::seconds(10)); |
| 942 | barrier.wait(); |
| 943 | contender.tick_run(&FleetRunId::from("run-1")).unwrap() |
| 944 | }) |
| 945 | }) |
| 946 | .collect::<Vec<_>>(); |
| 947 | let reports = handles |
| 948 | .into_iter() |
| 949 | .map(|handle| handle.join().unwrap()) |
| 950 | .collect::<Vec<_>>(); |
| 951 | |
| 952 | assert_eq!(reports.iter().map(|report| report.alerts).sum::<usize>(), 1); |
| 953 | assert_eq!(reports.iter().map(|report| report.failed).sum::<usize>(), 0); |
| 954 | let state = owner.ledger.rebuild_state().unwrap(); |
| 955 | assert_eq!(state.alerts.len(), 1); |
| 956 | assert_eq!(state.escalated_events.len(), 1); |
| 957 | assert_eq!( |
| 958 | ledger_text(&owner) |
| 959 | .lines() |
| 960 | .filter(|line| line.contains("\"record\":\"alert_sent\"")) |
| 961 | .count(), |
| 962 | 1 |
| 963 | ); |
| 964 | |
| 965 | owner.ledger.compact().unwrap(); |
| 966 | let mut after_compaction = scheduler(&tmp, 1); |
| 967 | after_compaction.set_now(base_now() + chrono::Duration::seconds(20)); |
| 968 | let report = after_compaction |
| 969 | .tick_run(&FleetRunId::from("run-1")) |
| 970 | .unwrap(); |
| 971 | assert_eq!(report.alerts, 0); |
| 972 | assert_eq!( |
| 973 | after_compaction |
| 974 | .ledger |
| 975 | .rebuild_state() |
| 976 | .unwrap() |
| 977 | .alerts |
| 978 | .len(), |
| 979 | 1 |
| 980 | ); |
| 981 | } |
| 982 | |
| 983 | #[test] |
| 984 | fn fleet_scheduler_slow_provider_response_with_fresh_heartbeat_is_not_stale() { |
| 985 | let tmp = TempDir::new().unwrap(); |
| 986 | let mut scheduler = scheduler(&tmp, 1); |
| 987 | create_run(&scheduler, "run-1", vec![task("task-a", 2)], 1); |
| 988 | scheduler.tick_run(&FleetRunId::from("run-1")).unwrap(); |
| 989 | scheduler.set_now(base_now() + chrono::Duration::seconds(4)); |
| 990 | scheduler |
| 991 | .append_worker_event( |
| 992 | &FleetRunId::from("run-1"), |
| 993 | "worker-1", |
| 994 | "task-a", |
| 995 | FleetWorkerEventPayload::ModelWait { |
| 996 | model: Some("deepseek-v4-pro".to_string()), |
| 997 | }, |
| 998 | ) |
| 999 | .unwrap(); |
| 1000 | scheduler |
| 1001 | .ledger |
| 1002 | .heartbeat("worker-1", &scheduler.timestamp(), None, None) |
| 1003 | .unwrap(); |
| 1004 | |
| 1005 | let report = scheduler.tick_run(&FleetRunId::from("run-1")).unwrap(); |
| 1006 | |
| 1007 | assert_eq!(report.marked_stale, 0); |
| 1008 | assert_eq!(report.restarted, 0); |
| 1009 | let state = scheduler.ledger.rebuild_state().unwrap(); |
| 1010 | assert_eq!(state.tasks["run-1:task-a"].entry.attempts, 1); |
| 1011 | assert_eq!(state.workers["worker-1"], FleetWorkerStatus::Busy); |
| 1012 | } |
| 1013 | |
| 1014 | #[test] |
| 1015 | fn fleet_scheduler_cancel_run_interrupts_active_and_cancels_queued() { |
| 1016 | let tmp = TempDir::new().unwrap(); |
| 1017 | let scheduler = scheduler(&tmp, 1); |
| 1018 | create_run( |
| 1019 | &scheduler, |
| 1020 | "run-1", |
| 1021 | vec![task("task-a", 3), task("task-b", 3), task("task-c", 3)], |
| 1022 | 2, |
| 1023 | ); |
| 1024 | scheduler.tick_run(&FleetRunId::from("run-1")).unwrap(); |
| 1025 | |
| 1026 | let report = scheduler |
| 1027 | .cancel_run(&FleetRunId::from("run-1"), "operator") |
| 1028 | .unwrap(); |
| 1029 | |
| 1030 | assert_eq!(report.cancelled, 3); |
| 1031 | let state = scheduler.ledger.rebuild_state().unwrap(); |
| 1032 | for task in state.tasks.values() { |
| 1033 | assert_eq!(task.status, FleetTaskLedgerStatus::Cancelled); |
| 1034 | } |
| 1035 | let ledger = ledger_text(&scheduler); |
| 1036 | assert!(ledger.contains("\"state\":\"interrupted\"")); |
| 1037 | assert!(ledger.contains("\"state\":\"cancelled\"")); |
| 1038 | } |
| 1039 | } |
| 1040 |