runnerd_core/sizing.rs
1//! Right-sizing advice from the job history: the spread of each profile's and each job's sampled peaks, waits and
2//! slot use, and the `memory` and `slots` to set. Pure: the store reads the history, the daemon serves the result
3//! (`api::Sizing`). It never changes the config: a client applies advice through `PUT /v1/config`, which validates
4//! it like any other config, and under `[sizing] auto` the daemon applies what `auto::changes` picks the same way.
5
6use crate::admission::Key;
7use crate::admission::recent::Recent;
8use crate::api::{self, Advice, Peaks, Spread};
9use crate::config::{Admission, Config};
10use crate::observe::metrics::SAMPLE_INTERVAL;
11
12pub(crate) use crate::api::GRANULE;
13
14/// Without its container's `memory.events`, a peak at or above this share of its cap, in percent, likely met it.
15const AT_CAP: u64 = 95;
16
17/// Lowering a cap is advised once it saves at least this share of it, in percent.
18const WORTH_LOWERING: u64 = 10;
19
20/// Jobs that waited this long at p95, in seconds, or slots busy this share of the time, in percent, want more
21/// slots.
22const WAITED: u64 = 60;
23const BUSY: u32 = 50;
24
25/// One finished job of this daemon's runners, as sizing reads it. Times in milliseconds.
26#[derive(Debug, Clone, PartialEq, Eq)]
27pub struct Ran {
28 /// `owner/name`.
29 pub repository: String,
30 /// The workflow's name, when known.
31 pub workflow: Option<String>,
32 /// The job's name.
33 pub name: String,
34 /// The profile its runner ran in.
35 pub profile: String,
36 /// Its container's memory cap, when recorded.
37 pub memory_cap: Option<u64>,
38 /// Its container's CPU cap, millicores, when recorded.
39 pub cpu_cap: Option<u64>,
40 /// Its sampled memory peak, bytes, as `measure` counts memory.
41 pub memory_peak: u64,
42 /// Its sampled CPU peak, millicores.
43 pub cpu_peak: u64,
44 /// Its CPU time over its run time, millicores; `None` when not recorded.
45 pub cpu_average: Option<u64>,
46 /// When it was queued.
47 pub queued_at: u64,
48 /// When it started; `None` for a job lost before.
49 pub started_at: Option<u64>,
50 /// When its runner's container started: before `queued_at`, it was a warm runner.
51 pub runner_started_at: u64,
52 /// When it finished, or was found lost.
53 pub finished_at: Option<u64>,
54 /// Whether the kernel killed a process in its container for memory.
55 pub oom_killed: bool,
56 /// Whether its container's memory, or a descendant's, reached a limit with nothing left to reclaim
57 /// (`memory.events` `oom`); `None` when its events were not recorded.
58 pub hit_cap: Option<bool>,
59 /// Whether its container's own limit failed reclaim (`memory.events.local` `oom`), not a descendant's; `None`
60 /// when not recorded, as before schema 14.
61 pub own_cap: Option<bool>,
62 /// What its sampled memory counts.
63 pub measure: Measure,
64 /// Whether it completed with success: a failed or lost run may have stopped before its heaviest step.
65 pub succeeded: bool,
66 /// Whether it completed with `failure` or `timed_out`: not lost, not cancelled.
67 pub failed: bool,
68 /// Its workflow run on GitHub: a re-run of its failed jobs has the same.
69 pub run_id: u64,
70 /// The branch it ran for, when known.
71 pub branch: Option<String>,
72 /// What measured admission reserved for it; `None` under caps or for an adopted runner.
73 pub reserved: Option<u64>,
74 /// How long of its wait it waited because the host's memory could not take its runner, in ms: no more slots
75 /// would have started it sooner.
76 pub host_waited: u64,
77 /// How much of its wait passed while no runnerd ran, in ms.
78 pub unseen: u64,
79 /// The periods of its CPU quota it had work in (`api::JobMetrics::cpu_periods`); `None` where the kernel kept no
80 /// figure, as in runs from before schema 17.
81 pub cpu_periods: Option<u64>,
82 /// Of those, the periods its quota throttled it in (`api::JobMetrics::cpu_throttled`).
83 pub cpu_throttled: Option<u64>,
84}
85
86/// A job's wait for a runner as sizing, insights and the history's figures count it: from when GitHub queued it, or
87/// from when the history holds every job if that came later (`history_kept`), as for a job queued before runnerd
88/// first ran, to when it started, less what of it passed while no runnerd ran (ADR 0081). A job's own page shows its
89/// whole wait, as GitHub's does; arrivals, as `simulate` and warm runners model them, go by `queued_at`.
90#[derive(Debug, Clone, Copy, PartialEq, Eq)]
91pub struct Wait {
92 /// When GitHub queued it, in ms.
93 pub queued_at: u64,
94 /// When it started, in ms.
95 pub started_at: u64,
96 /// How much of it passed while no runnerd ran, in ms.
97 pub unseen: u64,
98}
99
100impl Wait {
101 /// In ms, 0 at least, with the history holding every job since `kept_since` (ms).
102 pub fn counted(self, kept_since: u64) -> u64 {
103 self.started_at
104 .saturating_sub(self.queued_at.max(kept_since))
105 .saturating_sub(self.unseen)
106 }
107}
108
109/// What a run's sampled memory counts.
110#[derive(Debug, Clone, Copy, PartialEq, Eq)]
111pub enum Measure {
112 /// `memory.current` less `inactive_file`, as runners started before schema version 8 were sampled: page cache
113 /// read twice (`active_file`) counts, so peaks overstate what a job needs.
114 WithActiveFile,
115 /// `memory.current` less the file LRU lists (`active_file`, `inactive_file`): an estimate of what a job holds
116 /// without the page cache the kernel can drop.
117 WithoutPageCache,
118}
119
120/// What a run says about its container's memory cap (ADR 0070).
121#[derive(Debug, Clone, Copy, PartialEq, Eq)]
122pub enum Capped {
123 /// The container's own limit failed reclaim: it needed more than the cap.
124 Met,
125 /// Recorded without the local counter: by the reading before it, a kill or a limit hit in the container, or
126 /// without events a peak within `AT_CAP` percent of the cap. Holds and floors as `Met`, never raises a cap.
127 Assumed,
128 /// A process in it was killed for memory while its own limit held: the slice or the host ran out.
129 Killed,
130 /// None of those. A peak near the cap with no `oom` at it is clear: a JVM fills the room it has.
131 Clear,
132}
133
134impl Capped {
135 /// Whether it says the job needed its cap: `Met` or `Assumed`.
136 pub fn met(self) -> bool {
137 matches!(self, Capped::Met | Capped::Assumed)
138 }
139
140 /// Whether it puts its job on its cap: anything but `Clear`.
141 pub fn holds(self) -> bool {
142 self != Capped::Clear
143 }
144}
145
146impl Ran {
147 /// Its wait, once it started.
148 pub fn wait(&self) -> Option<Wait> {
149 Some(Wait {
150 queued_at: self.queued_at,
151 started_at: self.started_at?,
152 unseen: self.unseen,
153 })
154 }
155
156 /// What it says about its container's `cap`. A kill that only a descendant's own limit explains, the counters
157 /// showing a limit hit below the container and none at it, is a service container the job capped itself: clear.
158 /// A kill no sample saw an `oom` before, its last sample within `AT_CAP` percent of the cap, met it in the last
159 /// interval.
160 pub fn capped(&self, cap: u64) -> Capped {
161 match self.own_cap {
162 Some(true) => Capped::Met,
163 // No sample saw `oom` at any level and its last came near the cap: its own limit killed it before the
164 // next sample could count it.
165 Some(false)
166 if self.oom_killed
167 && self.hit_cap == Some(false)
168 && self.memory_peak.saturating_mul(100) >= cap.saturating_mul(AT_CAP) =>
169 {
170 Capped::Met
171 }
172 Some(false) if self.oom_killed && self.hit_cap != Some(true) => Capped::Killed,
173 None if self.oom_killed
174 || self.hit_cap.unwrap_or_else(|| {
175 self.memory_peak.saturating_mul(100) >= cap.saturating_mul(AT_CAP)
176 }) =>
177 {
178 Capped::Assumed
179 }
180 Some(false) | None => Capped::Clear,
181 }
182 }
183}
184
185/// One runner container's time in a slot for a job. Times in milliseconds.
186#[derive(Debug, Clone, PartialEq, Eq)]
187pub struct Held {
188 /// Its profile.
189 pub profile: String,
190 /// When it started holding the slot for a job: its start, or when its job was queued if that came later, as for
191 /// a warm runner.
192 pub started_at: u64,
193 /// When it was removed; `None` while it runs.
194 pub removed_at: Option<u64>,
195}
196
197/// A profile's `cpus` as the daemon ran it from `at` (ms), by its `profile.changed` event: one when it started with
198/// the profile, and after each reload or restart that changed the profile's image, CPUs, memory or slots.
199#[derive(Debug, Clone, PartialEq, Eq)]
200pub struct CpusSet {
201 /// The profile.
202 pub profile: String,
203 /// Its `cpus` from then.
204 pub cpus: u32,
205 /// When, in milliseconds.
206 pub at: u64,
207}
208
209/// A profile's `memory` as the daemon ran it from `at` (ms), by its `profile.changed` event, as `CpusSet` has its
210/// `cpus`.
211#[derive(Debug, Clone, PartialEq, Eq)]
212pub struct MemorySet {
213 /// The profile.
214 pub profile: String,
215 /// Its `memory` from then, in bytes.
216 pub memory: u64,
217 /// When, in milliseconds.
218 pub at: u64,
219}
220
221/// What the store reads for sizing.
222#[derive(Debug, Clone, Default, PartialEq, Eq)]
223pub struct History {
224 /// Finished jobs with a recorded peak, cancelled ones left out.
225 pub jobs: Vec<Ran>,
226 /// Cancelled jobs, for what admission learns from how they ended: a kill for memory, a peak at the cap or past
227 /// what was reserved. Their peaks (0 when never sampled) are not read as a job's size.
228 pub cancelled: Vec<Ran>,
229 /// Runners that held a slot at some point after the history's start.
230 pub runners: Vec<Held>,
231 /// Auto sizing's changes, oldest first, from `CHANGES_READ` before now: what a revert reads (ADR 0070). Empty
232 /// where nothing reads them, as for admission.
233 pub changes: Vec<api::SizingChange>,
234 /// Each profile's `memory` as the daemon ran it, oldest first: since when a raise reads runs at it.
235 pub memory_set: Vec<MemorySet>,
236 /// From when, in ms, the history holds every job (`history_kept`): a wait before it, as of a job queued before
237 /// runnerd first ran, counts from then.
238 pub kept_since: u64,
239}
240
241/// What `advise` sizes against besides the history.
242#[derive(Debug, Clone, Copy)]
243pub struct Window {
244 /// The start asked for, in milliseconds, at most `now`.
245 pub since: u64,
246 /// Now, in milliseconds.
247 pub now: u64,
248 /// Seconds between two samples of a container.
249 pub sample_interval: u64,
250}
251
252impl Window {
253 /// The window `GET /v1/sizing` reads: from `since` (seconds), default 30 days before `now` (seconds), and
254 /// never after `now`.
255 pub fn new(since: Option<u64>, now: u64) -> Window {
256 Window {
257 since: since
258 .unwrap_or_else(|| now.saturating_sub(30 * 24 * 3600))
259 .min(now)
260 .saturating_mul(1000),
261 now: now.saturating_mul(1000),
262 sample_interval: SAMPLE_INTERVAL.as_secs(),
263 }
264 }
265}
266
267/// The advice for every profile of `config` from `history`, against the slice's `memory_max` as `config`'s
268/// admission policy budgets it. Peaks, and the memory advice, go by each job's sizing window over the history's last
269/// 30 days (`memory`, ADR 0070); `window.since` narrows only the waits, the busy share and the starts.
270pub fn advise(
271 config: &Config,
272 memory_max: Option<u64>,
273 history: &History,
274 window: Window,
275) -> api::Sizing {
276 let since = first_runner(history, window.since);
277 let caps = config.memory_caps().map_or(u64::MAX, |caps| caps.0);
278 let mirror = config.mirror_memory();
279 let rules = &config.sizing;
280 let recent = memory::windows(config, history, window.now);
281 let profiles = config
282 .profiles
283 .iter()
284 .map(|(name, profile)| {
285 let jobs: Vec<&Ran> = history
286 .jobs
287 .iter()
288 .filter(|job| job.profile == *name && job.queued_at >= window.since)
289 .collect();
290 let share = profile.memory.0.saturating_mul(u64::from(profile.slots));
291 let busy = busy(history, name, profile.slots, since, window.now);
292 let wait = waits(&jobs, history.kept_since);
293 let starts = starts(&jobs, history.kept_since);
294 let cap = profile.memory.0;
295 let took = memory::took(&history.memory_set, name, cap, window.now);
296 let sized = Sized {
297 cap,
298 cpus: u64::from(profile.cpus) * 1000,
299 slots: profile.slots,
300 share,
301 admission: config.host.admission,
302 room: memory_max.map(|max| match config.host.admission {
303 // What the slice leaves this profile once the other profiles and the mirror took theirs.
304 Admission::Caps => max.saturating_sub(caps.saturating_sub(share)),
305 // Measured admission reserves predicted peaks, not caps: the config check only holds one
306 // runner of each profile beside the mirror.
307 Admission::Measured => max.saturating_sub(mirror),
308 }),
309 wants_slots: wait.p95 >= WAITED || busy >= BUSY,
310 min_jobs: rules.min_jobs,
311 };
312 let raises = Raises {
313 history,
314 recent: &recent,
315 profile: name,
316 cap,
317 took,
318 margin: rules.margin,
319 };
320 let memory = job_memory(&raises, &jobs, &sized, window.now, busy);
321 let peaks = memory::peaks(
322 recent.jobs_of(name).flat_map(|(_, runs)| runs.window()),
323 memory.on_cap,
324 sized.cpus,
325 );
326 let current = api::Resize {
327 memory: cap,
328 slots: profile.slots,
329 cpus: None,
330 };
331 let revert = memory::revert(config, history, name, current);
332 let advice = match &revert {
333 Some((_, to)) => Advice::Raise {
334 memory: to.memory,
335 slots: to.slots,
336 },
337 None => sized.advice(
338 &peaks,
339 Needs {
340 raise: raises.need(None),
341 floor: memory.floor.max(memory::revert_floor(history, name)),
342 },
343 sized.wants_slots,
344 ),
345 };
346 api::ProfileSizing {
347 name: name.clone(),
348 memory: cap,
349 cpus: profile.cpus,
350 slots: profile.slots,
351 advice,
352 revert: revert.map(|(revert, _)| revert),
353 peaks,
354 wait,
355 starts,
356 busy,
357 wants_slots: sized.wants_slots,
358 jobs: memory.jobs,
359 }
360 })
361 .collect();
362 api::Sizing {
363 since: since / 1000,
364 until: window.now / 1000,
365 margin: rules.margin,
366 sample_interval: window.sample_interval,
367 min_jobs: rules.min_jobs,
368 memory_max,
369 admission: config.host.admission.policy(),
370 profiles,
371 auto: config.sizing.auto,
372 changes: Vec::new(),
373 auto_plan: None,
374 }
375}
376
377/// What a raise of one profile reads (`memory::raise_need`).
378struct Raises<'history> {
379 history: &'history History,
380 recent: &'history Recent,
381 profile: &'history str,
382 /// Its `memory`.
383 cap: u64,
384 /// Since when it has had it, in ms.
385 took: u64,
386 margin: u32,
387}
388
389impl Raises<'_> {
390 /// What a raise needs from the job `key`, or from every job for `None`.
391 fn need(&self, key: Option<&Key>) -> u64 {
392 memory::raise_need(
393 self.history,
394 self.recent,
395 self.profile,
396 self.cap,
397 self.took,
398 self.margin,
399 key,
400 )
401 }
402}
403
404/// What a profile's jobs say about its `memory`.
405struct JobMemory {
406 /// Each job's sizing.
407 jobs: Vec<api::JobSizing>,
408 /// How many of them are on their cap.
409 on_cap: u64,
410 /// The largest of their floors.
411 floor: u64,
412}
413
414/// Each job of `raises.profile` with runs in its window: its peaks, waits and advice; `jobs` are the profile's runs
415/// since the window's start, for the waits.
416fn job_memory(raises: &Raises<'_>, jobs: &[&Ran], sized: &Sized, now: u64, busy: u32) -> JobMemory {
417 let mut on_cap = 0;
418 let mut floor = 0;
419 let sizings = raises
420 .recent
421 .jobs_of(raises.profile)
422 .filter(|(_, runs)| runs.runs() > 0)
423 .map(|(key, runs)| {
424 let held = memory::on_cap(runs, now);
425 on_cap += u64::from(held);
426 let peaks = memory::peaks(runs.window(), u64::from(held), sized.cpus);
427 let needs = Needs {
428 raise: raises.need(Some(key)),
429 floor: memory::floor(runs, raises.margin),
430 };
431 floor = floor.max(needs.floor);
432 let runs_of: Vec<&Ran> = jobs
433 .iter()
434 .copied()
435 .filter(|job| Key::of(job) == *key)
436 .collect();
437 let wait = waits(&runs_of, raises.history.kept_since);
438 api::JobSizing {
439 repository: key.repository.clone(),
440 workflow: key.workflow.clone(),
441 name: key.name.clone(),
442 advice: sized.advice(&peaks, needs, wait.p95 >= WAITED || busy >= BUSY),
443 peaks,
444 wait,
445 }
446 })
447 .collect();
448 JobMemory {
449 jobs: sizings,
450 on_cap,
451 floor,
452 }
453}
454
455/// Where a history read from `start` (ms) begins: at its first runner when that started later, as a new daemon's
456/// history does.
457fn first_runner(history: &History, start: u64) -> u64 {
458 history
459 .runners
460 .iter()
461 .map(|runner| runner.started_at)
462 .min()
463 .map_or(start, |first| first.max(start))
464}
465
466/// The percent of the time from `since` to `now` (ms) that the profile's runners held its `slots`.
467fn busy(history: &History, profile: &str, slots: u32, since: u64, now: u64) -> u32 {
468 let span = now.saturating_sub(since).max(1);
469 let held: u64 = history
470 .runners
471 .iter()
472 .filter(|runner| runner.profile == profile)
473 .map(|runner| {
474 let end = runner.removed_at.unwrap_or(now).min(now);
475 end.saturating_sub(runner.started_at.max(since))
476 })
477 .sum();
478 u32::try_from(
479 (held.saturating_mul(100) / span.saturating_mul(u64::from(slots).max(1))).min(100),
480 )
481 .unwrap_or(100)
482}
483
484/// A profile's limits and the rules advice follows.
485struct Sized {
486 cap: u64,
487 /// Its `cpus`, in millicores: CPU peaks are read only of runs at them.
488 cpus: u64,
489 slots: u32,
490 /// What its slots take of the budget now: `memory × slots`.
491 share: u64,
492 admission: Admission,
493 /// The memory the slice holds for it, as `admission` budgets it: all its slots at their caps under `caps`, one
494 /// runner under `measured`; `None` for an unlimited slice.
495 room: Option<u64>,
496 /// Whether its jobs waited for a slot at p95, or its slots were busy, enough to want more.
497 wants_slots: bool,
498 min_jobs: u32,
499}
500
501/// What a set of jobs' runs say its `memory` needs (`memory`, ADR 0070).
502#[derive(Debug, Clone, Copy)]
503struct Needs {
504 /// What a raise goes to: from runs that met the current cap since the profile took it; 0 without one.
505 raise: u64,
506 /// What a lowering stays at or above: every windowed run's need, and any revert's floor.
507 floor: u64,
508}
509
510impl Sized {
511 /// A raise toward `needs.raise` when it is above the cap; a lowering toward `needs.floor` when that saves a tenth
512 /// and no job is on its cap; else keep, with `needs.floor` as what the jobs need, above the cap or not.
513 fn advice(&self, peaks: &Peaks, needs: Needs, wants_slots: bool) -> Advice {
514 if peaks.jobs < u64::from(self.min_jobs) {
515 return Advice::TooFew { jobs: peaks.jobs };
516 }
517 if needs.raise > self.cap {
518 let need = needs.raise;
519 let fit = |memory: u64| u32::try_from(memory / need).unwrap_or(u32::MAX);
520 let slots = match (self.admission, self.room) {
521 (Admission::Caps, Some(room)) => fit(room).min(self.slots),
522 (Admission::Measured, Some(room)) if room < need => 0,
523 (_, _) => self.slots,
524 };
525 if slots == 0 {
526 return Advice::SliceTooSmall { memory: need };
527 }
528 return Advice::Raise {
529 memory: need,
530 slots,
531 };
532 }
533 let need = needs.floor.max(GRANULE);
534 // A job on its cap may have been held back or killed by it: no lowering until it ran clear of it again.
535 if peaks.on_cap == 0
536 && need <= self.cap
537 && self.cap - need >= self.cap * WORTH_LOWERING / 100
538 {
539 let fit = |memory: u64| u32::try_from(memory / need).unwrap_or(u32::MAX);
540 // Under measured admission slots are not bought with caps: the advice line says when more may help.
541 return Advice::Lower {
542 memory: need,
543 slots: if wants_slots && self.admission == Admission::Caps {
544 fit(self.share).max(self.slots)
545 } else {
546 self.slots
547 },
548 };
549 }
550 Advice::Keep { memory: need }
551 }
552}
553
554/// The [waits](Wait) of the jobs that started, in seconds, less what they waited for the host's memory.
555fn waits(jobs: &[&Ran], kept_since: u64) -> Spread {
556 spread(
557 jobs.iter()
558 .filter_map(|job| {
559 job.wait()
560 .map(|wait| wait.counted(kept_since).saturating_sub(job.host_waited) / 1000)
561 })
562 .collect(),
563 )
564}
565
566/// The jobs that started on a warm runner and the others, and their median waits.
567fn starts(jobs: &[&Ran], kept_since: u64) -> api::Starts {
568 let (warm, cold): (Vec<&Ran>, Vec<&Ran>) = jobs
569 .iter()
570 .filter(|job| job.started_at.is_some())
571 .partition(|job| job.runner_started_at < job.queued_at);
572 let count = |jobs: &[&Ran]| u32::try_from(jobs.len()).unwrap_or(u32::MAX);
573 api::Starts {
574 warm: count(&warm),
575 warm_p50: waits(&warm, kept_since).p50,
576 cold: count(&cold),
577 cold_p50: waits(&cold, kept_since).p50,
578 }
579}
580
581/// Nearest-rank percentiles; all zero for no values.
582pub(crate) fn spread(mut values: Vec<u64>) -> Spread {
583 values.sort_unstable();
584 let rank = |percent: usize| {
585 let index = (values.len() * percent).div_ceil(100).saturating_sub(1);
586 values.get(index).copied().unwrap_or(0)
587 };
588 Spread {
589 p50: rank(50),
590 p95: rank(95),
591 p99: rank(99),
592 max: values.last().copied().unwrap_or(0),
593 }
594}
595
596pub mod auto;
597pub mod memory;
598
599#[cfg(test)]
600mod tests;