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