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