runnerd_core/api/sizing.rs
1//! Right-sizing as clients read it: per profile and per job (repository, workflow and job name), the spread of the
2//! job history's sampled memory and CPU peaks, how long jobs waited for a slot, how busy the slots were, and what
3//! to set each profile's `memory` and `slots` to. Plain data; the daemon computes it from the history. Sizes are
4//! bytes, CPU millicores, times seconds since the Unix epoch.
5
6use serde::{Deserialize, Serialize};
7
8use crate::api::AdmissionPolicy;
9use crate::config::ByteSize;
10
11/// `GET /v1/sizing` query parameters.
12#[derive(Debug, Clone, Default, PartialEq, Eq, Serialize, Deserialize)]
13pub struct SizingQuery {
14 /// Only jobs queued at or after this time; default 30 days ago.
15 pub since: Option<u64>,
16}
17
18/// `GET /v1/sizing`: what the job history says each configured profile needs.
19#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
20pub struct Sizing {
21 /// The start of the history read: `since`, or the first runner recorded after it.
22 pub since: u64,
23 /// When it was computed.
24 pub until: u64,
25 /// The percent added to a sampled peak, `[sizing] margin`: peaks are sampled every `sample_interval`
26 /// seconds, so a shorter spike between two samples goes unseen.
27 pub margin: u32,
28 /// Seconds between two samples of a job's container.
29 pub sample_interval: u64,
30 /// The fewest jobs a profile or job needs before its percentiles become advice, `[sizing] min_jobs`.
31 pub min_jobs: u32,
32 /// The slice's `MemoryMax`; `None` when it is unlimited.
33 pub memory_max: Option<u64>,
34 /// `[host] admission`, which the advice budgets `MemoryMax` by.
35 #[serde(default)]
36 pub admission: AdmissionPolicy,
37 /// Every configured profile, by name.
38 pub profiles: Vec<ProfileSizing>,
39 /// Whether the daemon applies advice itself, `[sizing] auto`.
40 #[serde(default)]
41 pub auto: bool,
42 /// The changes auto sizing made since `since`, oldest first.
43 #[serde(default)]
44 pub changes: Vec<SizingChange>,
45 /// Under auto sizing, what it does next and when. `None` while it is off, and from a daemon before it said.
46 #[serde(default, skip_serializing_if = "Option::is_none")]
47 pub auto_plan: Option<AutoPlan>,
48}
49
50/// What auto sizing will do: its rules, the changes its next look would make as the history and config are now, and
51/// the profiles a recent change holds.
52#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
53pub struct AutoPlan {
54 /// Seconds between two looks of the daemon's.
55 pub every: u64,
56 /// `[sizing] interval`: the least time between a change to one profile and one that lowers it, in seconds.
57 pub interval: u64,
58 /// `[sizing] raise_interval`: the least time between a change to one profile and one that only raises it, in
59 /// seconds. 0 from a daemon before it said: `interval` then holds every change.
60 #[serde(default)]
61 pub raise_interval: u64,
62 /// `[sizing] max_step`: the most one change moves a profile's `memory`, in percent of it.
63 pub max_step: u32,
64 /// The changes the next look would make, in the order it would make them; `at` is when this was worked out.
65 /// None of them is made while `blocked` says why.
66 pub next: Vec<SizingChange>,
67 /// The profiles a change in the last `interval` holds from some change, the hold that ends first first.
68 pub held: Vec<SizingHold>,
69 /// Why no look changes anything until the owner acts, as when the file on disk differs from the config runnerd
70 /// runs with. `None` when the next look may write `next`, and from a daemon before it said.
71 #[serde(default, skip_serializing_if = "Option::is_none")]
72 pub blocked: Option<String>,
73}
74
75/// A profile auto sizing changes nothing in before `until`, and lowers nothing in before `lower_until`.
76#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
77pub struct SizingHold {
78 /// The profile.
79 pub profile: String,
80 /// When its last change was made.
81 pub changed: u64,
82 /// When it may be raised again, in seconds since the Unix epoch.
83 pub until: u64,
84 /// When it may be lowered again; 0 from a daemon before it said, which held every change until `until`.
85 #[serde(default)]
86 pub lower_until: u64,
87}
88
89impl SizingHold {
90 /// When it may be lowered again.
91 pub fn lowered_after(&self) -> u64 {
92 self.lower_until.max(self.until)
93 }
94}
95
96/// A profile's `memory` and `slots` as auto sizing changed them in the config file.
97#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
98pub struct SizingChange {
99 /// When it was written.
100 pub at: u64,
101 /// The profile.
102 pub profile: String,
103 /// What they were: writing these back undoes it.
104 pub from: Resize,
105 /// What they are now.
106 pub to: Resize,
107 /// Why, as the advice said it then.
108 pub reason: String,
109 /// Whether it may not have been written: recorded before its file was written, and the config no longer said
110 /// whether it was when runnerd looked again. It still holds the interval, and `from` is still what it replaced.
111 #[serde(default)]
112 pub uncertain: bool,
113}
114
115/// One profile: its limits now, its jobs' peaks and waits, and the advice.
116#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
117pub struct ProfileSizing {
118 /// The profile.
119 pub name: String,
120 /// Its `memory` now: each container's hard limit.
121 pub memory: u64,
122 /// Its `cpus` now.
123 pub cpus: u32,
124 /// Its `slots` now.
125 pub slots: u32,
126 /// Its jobs' peaks.
127 pub peaks: Peaks,
128 /// Seconds from queued to started.
129 pub wait: Spread,
130 /// Its jobs by whether a warm runner took them, and how long each kind waited. Empty from a daemon before warm
131 /// runners.
132 #[serde(default)]
133 pub starts: Starts,
134 /// The percent of the time since `since` its slots held a runner for a job, from when the job was queued, over
135 /// its `slots` now.
136 pub busy: u32,
137 /// Whether its jobs waited a minute or more for a slot at p95, or its slots were busy half the time: more slots
138 /// would help.
139 #[serde(default)]
140 pub wants_slots: bool,
141 /// What to set `memory` and `slots` to. Sized for its largest peak, so no job gets less than it used.
142 pub advice: Advice,
143 /// Each job that ran in it, by repository, workflow and name.
144 pub jobs: Vec<JobSizing>,
145}
146
147/// A profile's jobs by whether their runner was already up when GitHub queued them (ADR 0050).
148#[derive(Debug, Clone, Copy, Default, PartialEq, Eq, Serialize, Deserialize)]
149pub struct Starts {
150 /// Jobs that started on a warm runner: one that started before the job was queued.
151 pub warm: u32,
152 /// Their median wait from queued to started, in seconds.
153 pub warm_p50: u64,
154 /// Jobs whose runner started after they were queued.
155 pub cold: u32,
156 /// Their median wait, in seconds.
157 pub cold_p50: u64,
158}
159
160/// The jobs of one name in one workflow of one repository.
161#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
162pub struct JobSizing {
163 /// `owner/name` on GitHub.
164 pub repository: String,
165 /// The workflow's name, when the history knew it.
166 pub workflow: Option<String>,
167 /// The job's name in the workflow.
168 pub name: String,
169 /// Its runs' peaks.
170 pub peaks: Peaks,
171 /// Seconds from queued to started.
172 pub wait: Spread,
173 /// What its profile's `memory` would be for this job alone, and the `slots` that would fit in the memory the
174 /// profile's slots take now.
175 pub advice: Advice,
176}
177
178/// The peaks of a set of jobs, each the highest of its samples.
179#[derive(Debug, Clone, Default, PartialEq, Eq, Serialize, Deserialize)]
180pub struct Peaks {
181 /// How many jobs: the number of peaks the percentiles are taken over.
182 pub jobs: u64,
183 /// Of those, how many met their container's memory cap: killed for memory, or out of memory at it with
184 /// nothing left to reclaim (for a run whose container's events were not recorded, a peak within 5% of it). They
185 /// may need more than they show.
186 pub at_cap: u64,
187 /// Memory without page cache, in bytes.
188 pub memory: Spread,
189 /// CPU, in millicores.
190 pub cpu: Spread,
191}
192
193/// Percentiles of a set of values, by nearest rank: each is one of the values, and with fewer than 100 values
194/// `p99` is the largest.
195#[derive(Debug, Clone, Copy, Default, PartialEq, Eq, Serialize, Deserialize)]
196pub struct Spread {
197 /// The median.
198 pub p50: u64,
199 /// The 95th percentile.
200 pub p95: u64,
201 /// The 99th percentile.
202 pub p99: u64,
203 /// The largest.
204 pub max: u64,
205}
206
207/// What a cap keeps at least over the largest peak, for page cache: 1G. Peaks leave page cache out, and a container
208/// whose memory past them leaves the kernel less than about 500M of page cache to reclaim at its cap reclaims badly,
209/// which throttles its sockets and slows its downloads (ADR 0055).
210pub const PAGE_CACHE_ROOM: u64 = 1 << 30;
211
212/// What to set a profile's `memory` and `slots` to. `memory` is the largest sampled peak plus the margin, or plus
213/// `PAGE_CACHE_ROOM` when that is more, rounded up to 256M: the cap kills a job that goes over it, so it is sized for
214/// the heaviest run, not a percentile.
215#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
216#[serde(rename_all = "kebab-case", tag = "kind")]
217pub enum Advice {
218 /// Fewer than `min_jobs` jobs: too few to advise on.
219 TooFew {
220 /// How many there are.
221 jobs: u64,
222 },
223 /// The cap fits, and lowering it would save less than a tenth or a run met its cap.
224 Keep {
225 /// What the jobs need.
226 memory: u64,
227 },
228 /// The cap is a tenth or more above what the jobs need, and no run met its cap.
229 Lower {
230 /// The new `memory`.
231 memory: u64,
232 /// The new `slots`: under `caps` admission, as many as fit in the memory the slots take now when jobs waited
233 /// for one; else the same.
234 slots: u32,
235 },
236 /// The jobs need more than the cap.
237 Raise {
238 /// The new `memory`.
239 memory: u64,
240 /// The new `slots`: the same, or under `caps` admission as many as the slice's `MemoryMax` still holds.
241 slots: u32,
242 },
243 /// The jobs need more than the cap, and the slice's `MemoryMax` does not hold one slot of what they need (under
244 /// `measured` admission, one runner beside the mirror): the slice has to grow first.
245 SliceTooSmall {
246 /// What the jobs need.
247 memory: u64,
248 },
249}
250
251impl Advice {
252 /// The `memory` and `slots` to write, when they differ from the profile's.
253 pub fn change(&self) -> Option<Resize> {
254 match *self {
255 Advice::Lower { memory, slots } | Advice::Raise { memory, slots } => Some(Resize {
256 memory,
257 slots,
258 cpus: None,
259 }),
260 Advice::TooFew { .. } | Advice::Keep { .. } | Advice::SliceTooSmall { .. } => None,
261 }
262 }
263}
264
265/// A profile's new `memory`, `slots` and, when auto sizing changed them, `cpus`.
266#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
267pub struct Resize {
268 /// Bytes.
269 pub memory: u64,
270 /// Slots.
271 pub slots: u32,
272 /// CPUs per slot; `None` leaves them as they are, as advice does and as auto sizing's changes recorded before it
273 /// changed them.
274 #[serde(default, skip_serializing_if = "Option::is_none")]
275 pub cpus: Option<u32>,
276}
277
278impl AutoPlan {
279 /// Its rules in words, as clients say them after `runnerd`: `looks every 5 minutes, raises a profile at most
280 /// once every 1 hour and lowers it at most once every 24 hours, moving its memory and CPUs by at most 25% at a
281 /// time`.
282 pub fn rules(&self) -> String {
283 if self.raise_interval == 0 || self.raise_interval >= self.interval {
284 return format!(
285 "looks every {} and changes a profile at most once every {}, moving its memory by at most {}% at a \
286 time",
287 period(self.every),
288 period(self.interval),
289 self.max_step
290 );
291 }
292 format!(
293 "looks every {}, raises a profile at most once every {} and lowers it at most once every {}, moving its \
294 memory and CPUs by at most {}% at a time",
295 period(self.every),
296 period(self.raise_interval),
297 period(self.interval),
298 self.max_step
299 )
300 }
301}
302
303/// A setting's length in words, as a config's periods read: `5 minutes`, `24 hours`, `90 s`.
304pub fn period(seconds: u64) -> String {
305 let plural = |count: u64, unit: &str| {
306 if count == 1 {
307 format!("{count} {unit}")
308 } else {
309 format!("{count} {unit}s")
310 }
311 };
312 match seconds {
313 0 => "0 s".to_owned(),
314 _ if seconds.is_multiple_of(3600) => plural(seconds / 3600, "hour"),
315 _ if seconds.is_multiple_of(60) => plural(seconds / 60, "minute"),
316 _ => format!("{seconds} s"),
317 }
318}
319
320impl SizingChange {
321 /// What it changed, such as `memory 9G → 6.75G, slots 2`, and `cpus 12 → 15` when it changed them.
322 pub fn line(&self) -> String {
323 let memory = if self.from.memory == self.to.memory {
324 format!("memory {}", ByteSize(self.to.memory))
325 } else {
326 format!(
327 "memory {} → {}",
328 ByteSize(self.from.memory),
329 ByteSize(self.to.memory)
330 )
331 };
332 let slots = if self.from.slots == self.to.slots {
333 format!("slots {}", self.to.slots)
334 } else {
335 format!("slots {} → {}", self.from.slots, self.to.slots)
336 };
337 let cpus = match (self.from.cpus, self.to.cpus) {
338 (Some(from), Some(to)) if from != to => format!(", cpus {from} → {to}"),
339 _ => String::new(),
340 };
341 if self.uncertain {
342 format!("{memory}, {slots}{cpus} (may not have been written)")
343 } else {
344 format!("{memory}, {slots}{cpus}")
345 }
346 }
347}
348
349impl Sizing {
350 /// One line on what to do with a profile and why, as `runnerctl sizing`, the dashboard and `runnerd doctor`
351 /// say it, such as `lower memory to 2.5G from 7G, its largest peak + 15%; 5 slots fit in the same 14G`.
352 /// Under `measured` admission, when jobs wanted more slots, it adds that `slots` is not bought with caps there, or
353 /// with auto sizing on, that it adds them.
354 pub fn advice_line(&self, profile: &ProfileSizing) -> String {
355 let line = self.memory_line(profile);
356 if self.admission == AdmissionPolicy::Measured
357 && profile.wants_slots
358 && !matches!(profile.advice, Advice::TooFew { .. })
359 {
360 let then = if self.auto {
361 "auto sizing adds one at a time while every profile's slots fit the slice at their jobs' peaks"
362 } else {
363 "measured admission reserves predicted peaks, not caps, so more slots may fit: raise slots by hand"
364 };
365 format!(
366 "{line}; jobs waited {}s for a slot (p95) and slots were busy {}%: {then}",
367 profile.wait.p95, profile.busy
368 )
369 } else {
370 line
371 }
372 }
373
374 /// `advice_line` without what it says about slots under measured admission.
375 pub(crate) fn memory_line(&self, profile: &ProfileSizing) -> String {
376 let size = |bytes: u64| ByteSize(bytes).to_string();
377 let peaks = &profile.peaks;
378 let need = if peaks.memory.max * u64::from(self.margin) / 100 >= PAGE_CACHE_ROOM {
379 format!("its largest peak + {}%", self.margin)
380 } else {
381 format!(
382 "its largest peak + {} for page cache",
383 size(PAGE_CACHE_ROOM)
384 )
385 };
386 let at_cap = format!(
387 "{} of {} jobs ran out of memory at their cap",
388 peaks.at_cap, peaks.jobs
389 );
390 match profile.advice {
391 Advice::TooFew { jobs } => format!(
392 "{jobs} jobs, too few to advise on (it takes {}); largest peak {} of a {} cap",
393 self.min_jobs,
394 size(peaks.memory.max),
395 size(profile.memory)
396 ),
397 Advice::Keep { .. } if peaks.at_cap > 0 => format!(
398 "keep {}: {at_cap}, so they may need more than they show",
399 size(profile.memory)
400 ),
401 Advice::Keep { memory } => {
402 format!("keep {}: {need} is {}", size(profile.memory), size(memory))
403 }
404 Advice::Lower { memory, slots } => {
405 let share = profile.memory * u64::from(profile.slots);
406 if self.admission == AdmissionPolicy::Measured {
407 return format!(
408 "lower memory to {} from {}, {need}",
409 size(memory),
410 size(profile.memory)
411 );
412 }
413 let then = if slots > profile.slots {
414 format!("{slots} slots fit in the same {}", size(share))
415 } else {
416 format!(
417 "frees {} of the budget",
418 size(share - memory * u64::from(slots))
419 )
420 };
421 format!(
422 "lower memory to {} from {}, {need}; {then}",
423 size(memory),
424 size(profile.memory)
425 )
426 }
427 Advice::Raise { memory, slots } => {
428 let why = if peaks.at_cap > 0 {
429 at_cap
430 } else {
431 format!("{need} is above the {} cap", size(profile.memory))
432 };
433 let slots = if slots < profile.slots {
434 format!("; {slots} of its {} slots fit the slice", profile.slots)
435 } else {
436 String::new()
437 };
438 format!("raise memory to {}: {why}{slots}", size(memory))
439 }
440 Advice::SliceTooSmall { memory } => format!(
441 "needs {memory_needed} ({need}), more than the slice holds for one slot: raise the slice's \
442 MemoryMax first",
443 memory_needed = size(memory)
444 ),
445 }
446 }
447}
448
449// The wire format is JSON, which only the host and client features read.
450#[cfg(all(test, any(feature = "host", feature = "client")))]
451mod tests;