runnerd-core 0.1.1

Config, domain logic and API types shared by runnerd and its clients
Documentation
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
//! Insights from the job history (ADR 0051): jobs that got slower and stayed so, jobs that fail and pass on a re-run,
//! jobs whose memory nears their profile's cap, jobs that find almost nothing in the build cache, repositories whose
//! downloads memory pressure throttled, jobs their CPU cap held back by the rule auto sizing reads (ADR 0087), jobs
//! held for the host's memory by programs outside CI (ADR 0100), and how long each repository's jobs wait, this week against the last when the history holds all of last week; and the
//! daily digest. What the host's memory came to while CI ran, and the warning when it ran low on a host credit
//! (ADR 0108), are `host`'s. Pure: the daemon reads the history, the config and the clock and passes them in. It sends nothing
//! and changes nothing.

use std::collections::{BTreeMap, BTreeSet};

use crate::api::{
    self, AutoNote, Digest, Insight, InsightKind, Insights, JobState, LongWait, RepositoryWaits,
    RunLink, Severity, WaitSpread, WeekWaits,
};
use crate::config::{Admission, ByteSize, Config};
use crate::sizing::auto::{CpuRun, RAISE_RUNS, cpu_step, unhelped_raise};
use crate::sizing::{HELD, Throttling, spread};

mod host;

const SECOND: u64 = 1000;
const DAY: u64 = 24 * 3600 * SECOND;
const WEEK: u64 = 7 * DAY;

/// The last runs a slower job's median is taken over; the baseline needs as many before them.
const RECENT: usize = 5;
/// A job is slower when its recent median is this share above its baseline's, in percent…
const SLOWER_PERCENT: u64 = 20;
/// …and at least this many seconds above it.
const SLOWER_SECONDS: u64 = 30;

/// The last runs whose memory peaks are read, and the fewest that say anything.
const CAP_RUNS: usize = 10;
const CAP_MIN_RUNS: usize = 5;
/// A job is near its cap when the 95th percentile of its peaks is this share of it, in percent.
const NEAR_CAP: u64 = 85;
/// Runs at `NEAR_CAP` or more that make it a warning.
const NEAR_CAP_WARNING: usize = 2;
/// Flaky runs that make it a warning.
const FLAKY_WARNING: u32 = 2;

/// A repository's waits are an insight when this week's p95 is this share above last week's, in percent, and a minute
/// above it, each week with this many jobs.
const WAITS_PERCENT: u64 = 50;
const WAITS_SECONDS: u64 = 60;
const WAITS_MIN_JOBS: u32 = 5;
/// The jobs a week needs for its p95 to be other than its longest wait, the nearest rank `spread` takes.
const WAITS_ONE_IN_20: u32 = 20;

/// A repository's runs the kernel throttled for memory pressure that make an insight.
const THROTTLED_MIN_RUNS: u32 = 2;

/// A queued job's wait for the host's memory that makes an insight, in milliseconds, while a job waits for it now.
const HOST_HELD_AFTER: u64 = 30 * 60 * SECOND;
/// The insight's key changes this often, so it is told once in each such window however long the wait lasts, in
/// milliseconds.
const HOST_HELD_EVERY: u64 = 6 * 3600 * SECOND;
/// How old the host's reading may be for the insight to hold, in milliseconds, and the least the engine's look at the
/// waiting jobs may be, which is three `[host] poll_interval`s when that is more: a daemon that stopped looking says
/// nothing about now.
const HOST_HELD_FRESH: u64 = 60 * SECOND;

/// One job the history closed, as insights read it. Times in milliseconds.
#[derive(Debug, Clone, PartialEq, Eq)]
pub struct Run {
    /// GitHub's job ID.
    pub id: u64,
    /// `owner/name`.
    pub repository: String,
    /// The workflow's name, when known.
    pub workflow: Option<String>,
    /// The job's name.
    pub name: String,
    /// The branch the run is for.
    pub branch: Option<String>,
    /// The job's page on GitHub.
    pub url: Option<String>,
    /// The workflow run.
    pub run_id: u64,
    /// The run's attempt, 1 for the first.
    pub run_attempt: u32,
    /// `Completed` or `Lost`.
    pub state: JobState,
    /// GitHub's conclusion, for a completed job.
    pub conclusion: Option<String>,
    /// When GitHub queued it.
    pub queued_at: u64,
    /// When it started.
    pub started_at: Option<u64>,
    /// When it finished, or was found lost.
    pub finished_at: Option<u64>,
    /// When its runner's container started: before `queued_at`, it found a warm runner.
    pub runner_started_at: Option<u64>,
    /// The profile its runner ran in.
    pub profile: Option<String>,
    /// Its sampled memory peak.
    pub memory_peak: Option<u64>,
    /// What the build cache served it, if its runner had the build cache.
    pub build_cache: Option<api::BuildCacheCounts>,
    /// Times the kernel throttled its container's sockets for memory pressure, when counted.
    pub sock_throttled: Option<u64>,
    /// Its container's CPU cap, millicores, when recorded.
    pub cpu_cap: Option<u64>,
    /// The periods of its CPU quota it had work in; `None` where the kernel kept no figure, as before schema 17.
    pub cpu_periods: Option<u64>,
    /// Of those, the periods its quota throttled it in.
    pub cpu_throttled: Option<u64>,
    /// How much of its wait passed while no runnerd ran.
    pub unseen: u64,
}

impl Run {
    fn finished(&self) -> u64 {
        self.finished_at.unwrap_or(0)
    }

    fn succeeded(&self) -> bool {
        self.state == JobState::Completed && self.conclusion.as_deref() == Some("success")
    }

    /// Failed as `GET /v1/stats` counts failures: failed, timed out or lost.
    fn failed(&self) -> bool {
        self.state == JobState::Lost || self.failed_itself()
    }

    /// Failed or timed out: what the job did, not what became of its runner.
    fn failed_itself(&self) -> bool {
        self.state == JobState::Completed
            && matches!(self.conclusion.as_deref(), Some("failure" | "timed_out"))
    }

    /// Seconds from start to finish.
    fn took(&self) -> Option<u64> {
        Some(self.finished_at?.saturating_sub(self.started_at?) / SECOND)
    }

    /// Its [wait](crate::sizing::Wait) in seconds, once it started.
    fn waited(&self, kept_since: u64) -> Option<u64> {
        let wait = crate::sizing::Wait {
            queued_at: self.queued_at,
            started_at: self.started_at?,
            unseen: self.unseen,
        };
        Some(wait.counted(kept_since) / SECOND)
    }

    fn throttling(&self) -> Throttling {
        Throttling {
            started_at: self.started_at,
            finished_at: self.finished_at,
            periods: self.cpu_periods,
            throttled: self.cpu_throttled,
        }
    }

    fn link(&self) -> RunLink {
        RunLink {
            job: self.id,
            run_id: self.run_id,
            run_attempt: self.run_attempt,
            branch: self.branch.clone(),
            url: self.url.clone(),
            at: self.finished() / SECOND,
        }
    }
}

/// The jobs waiting for the host's memory at one moment, and the reading that says what holds it. What the daemon
/// knows of now, not of the history.
#[derive(Debug, Clone, PartialEq, Eq)]
pub struct HostHold {
    /// The host's memory as last read.
    pub memory: api::HostMemory,
    /// When the engine last saw these jobs wait for it, in ms.
    pub seen: u64,
    /// The jobs waiting for it then that are still queued.
    pub jobs: Vec<HeldJob>,
}

/// A queued job waiting for the host's memory.
#[derive(Debug, Clone, PartialEq, Eq)]
pub struct HeldJob {
    /// `owner/name`.
    pub repository: String,
    /// The workflow's name, when known.
    pub workflow: Option<String>,
    /// The job's name.
    pub name: String,
    /// How long it has waited for the host's memory, in ms (`jobs.host_waited`).
    pub waited: u64,
}

/// A reading of the host's memory as a `budget.changed` event stored it (ADR 0108).
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub struct HostReading {
    /// When it was stored, in ms.
    pub at: u64,
    /// The host's `MemTotal`, in bytes.
    pub total: u64,
    /// Its `MemAvailable`.
    pub available: u64,
    /// Its `SwapFree`, on a host with swap.
    pub swap_free: Option<u64>,
    /// What the host's limit took off the plain count then: 0 for no credit, as before 0.0.30.
    pub credit: u64,
    /// Whether a runner ran: the budget counted more than the mirror.
    pub running: bool,
}

/// What insights read from the job history.
#[derive(Debug, Clone, Default, PartialEq, Eq)]
pub struct History {
    /// The jobs it closed in the days read.
    pub runs: Vec<Run>,
    /// From when it holds every job that ended, in ms (`history_kept`): before that, pruning or the daemon's first
    /// start left only some, so no week before it is compared.
    pub kept_since: u64,
    /// The jobs held for the host's memory now; `None` when none is, or the daemon has no reading of it.
    pub host: Option<HostHold>,
    /// The host's readings the budget's events stored, oldest first.
    pub readings: Vec<HostReading>,
}

/// A job as insights tell jobs apart: repository, workflow and name.
type JobKey<'run> = (&'run str, Option<&'run str>, &'run str);

fn key_of(run: &Run) -> JobKey<'_> {
    (&run.repository, run.workflow.as_deref(), &run.name)
}

/// The insights of the `days` before `now` (ms), and the waits of the 14 days before it. `history` may hold more:
/// only the runs finished by `now` count.
pub fn insights(history: &History, config: &Config, days: u32, now: u64) -> Insights {
    let runs = &history.runs;
    let since = now.saturating_sub(u64::from(days) * DAY);
    let window: Vec<&Run> = runs
        .iter()
        .filter(|run| {
            matches!(run.state, JobState::Completed | JobState::Lost)
                && run
                    .finished_at
                    .is_some_and(|at| (since..=now).contains(&at))
        })
        .collect();
    let mut jobs: BTreeMap<JobKey<'_>, Vec<&Run>> = BTreeMap::new();
    for run in &window {
        jobs.entry(key_of(run)).or_default().push(run);
    }
    let mut found = Vec::new();
    for (key, runs) in &mut jobs {
        runs.sort_by_key(|run| (run.finished(), run.id));
        found.extend(slower(*key, runs));
        found.extend(flaky(*key, runs));
        found.extend(near_cap(*key, runs, config));
        found.extend(cold_cache(*key, runs));
        found.extend(cpu_held(*key, runs, config));
    }
    found.extend(throttled(&window));
    found.extend(
        history
            .host
            .as_ref()
            .and_then(|hold| host_held(hold, config, now)),
    );
    let host_days = host::days(&history.readings, since, now);
    found.extend(host::low(&host_days));
    let waits = waits(history, now);
    found.extend(waits.iter().filter_map(waits_insight));
    found.sort_by(|a, b| {
        (a.severity, &a.repository, &a.workflow, &a.job, &a.key).cmp(&(
            b.severity,
            &b.repository,
            &b.workflow,
            &b.job,
            &b.key,
        ))
    });
    Insights {
        since: since / SECOND,
        until: now / SECOND,
        insights: found,
        waits,
        host: host_days,
    }
}

/// The 24 hours before `now` (ms): what finished and failed, the insights that hold at `now` and did not 24 hours
/// before, and the longest waits.
pub fn digest(history: &History, config: &Config, now: u64) -> Digest {
    let runs = &history.runs;
    let since = now.saturating_sub(DAY);
    let finished: Vec<&Run> = runs
        .iter()
        .filter(|run| {
            matches!(run.state, JobState::Completed | JobState::Lost)
                && run.finished_at.is_some_and(|at| at > since && at <= now)
        })
        .collect();
    let before: BTreeSet<String> = insights(history, config, api::INSIGHT_DAYS, since)
        .insights
        .into_iter()
        .map(|insight| insight.key)
        .collect();
    let new = insights(history, config, api::INSIGHT_DAYS, now)
        .insights
        .into_iter()
        .filter(|insight| !before.contains(&insight.key))
        .collect();
    let mut started: Vec<&Run> = runs
        .iter()
        .filter(|run| run.started_at.is_some_and(|at| at > since && at <= now))
        .collect();
    started.sort_by_key(|run| (std::cmp::Reverse(run.waited(history.kept_since)), run.id));
    Digest {
        since: since / SECOND,
        until: now / SECOND,
        finished: finished.len() as u64,
        failed: finished.iter().filter(|run| run.failed()).count() as u64,
        new,
        waits: started
            .into_iter()
            .take(3)
            .map(|run| LongWait {
                job: run.id,
                repository: run.repository.clone(),
                workflow: run.workflow.clone(),
                name: run.name.clone(),
                waited: run.waited(history.kept_since).unwrap_or(0),
                url: run.url.clone(),
            })
            .collect(),
        savings: None,
        host: host::hazard(&history.readings, since, now),
    }
}

/// A job has waited for the host's memory for 30 minutes or more, and `hold` says a job waits for it at `now` (ms):
/// the host's memory is short, and the programs outside CI that hold it are the owner's to close. Worded here once
/// for `runnerctl`, the dashboard, the TUI and notifications, with the longest waiter's repository. It names the
/// programs `HostMemory::outside_line` names and says closing them frees only part of what is held, since the figure
/// also counts the kernel and compressed swap.
///
/// Holds from the host's reading and the engine's look at the waiting jobs: the reading no older than a minute, the
/// look no older than that or three `[host] poll_interval`s, neither from after `now`, the second of the look (a plan
/// in that second is not from after it). With no reading, or a stale one, nothing is said. Its key is the same for 6
/// hours and then changes, so a notification target tells it once in each window of them (told keys are stored, so a
/// restart repeats none).
fn host_held(hold: &HostHold, config: &Config, now: u64) -> Option<Insight> {
    let reading = hold.memory.at.saturating_mul(SECOND);
    // The engine plans once a `poll_interval`, so its look is as old as three of them before it is missed.
    let poll = u64::try_from(config.host.poll_interval.as_millis()).unwrap_or(u64::MAX);
    let plan_fresh = HOST_HELD_FRESH.max(poll.saturating_mul(3));
    let fresh =
        |at: u64, within: u64| at / SECOND <= now / SECOND && now.saturating_sub(at) <= within;
    if !fresh(reading, HOST_HELD_FRESH) || !fresh(hold.seen, plan_fresh) {
        return None;
    }
    let longest = hold.jobs.iter().max_by_key(|job| job.waited)?;
    if longest.waited < HOST_HELD_AFTER {
        return None;
    }
    let jobs = u32::try_from(hold.jobs.len()).unwrap_or(u32::MAX);
    let waited = longest.waited / SECOND;
    let sentence = Insight::host_held_sentence(
        jobs,
        waited,
        &subject((
            &longest.repository,
            longest.workflow.as_deref(),
            &longest.name,
        )),
        &hold.memory,
    );
    Some(Insight {
        key: format!("host-held:{}", now / HOST_HELD_EVERY),
        severity: Severity::Warning,
        repository: longest.repository.clone(),
        workflow: None,
        job: None,
        sentence,
        kind: InsightKind::HostHeld { jobs, waited },
        runs: Vec::new(),
    })
}

/// `repository: workflow / job`, or `repository: job` when the workflow is not known.
fn subject((repository, workflow, name): JobKey<'_>) -> String {
    match workflow {
        Some(workflow) => format!("{repository}: {workflow} / {name}"),
        None => format!("{repository}: {name}"),
    }
}

fn key_text(kind: &str, (repository, workflow, name): JobKey<'_>) -> String {
    format!("{kind}:{repository}/{}/{name}", workflow.unwrap_or(""))
}

fn insight(
    key: JobKey<'_>,
    kind: &str,
    severity: Severity,
    sentence: String,
    found: InsightKind,
    runs: Vec<RunLink>,
) -> Insight {
    Insight {
        key: key_text(kind, key),
        severity,
        repository: key.0.to_owned(),
        workflow: key.1.map(str::to_owned),
        job: Some(key.2.to_owned()),
        sentence,
        kind: found,
        runs,
    }
}

/// The median: the mean of the middle two for an even count.
fn median(values: &[u64]) -> u64 {
    let mut values = values.to_vec();
    values.sort_unstable();
    let middle = values.len() / 2;
    match values.len() {
        0 => 0,
        len if len % 2 == 1 => values[middle],
        _ => u64::midpoint(values[middle - 1], values[middle]),
    }
}

/// A job whose successful runs on `main` or `master`, or else its busiest branch, got slower and stayed so.
fn slower(key: JobKey<'_>, runs: &[&Run]) -> Option<Insight> {
    let timed_on = |branch: &str| -> Vec<(&Run, u64)> {
        runs.iter()
            .filter(|run| run.succeeded() && run.branch.as_deref() == Some(branch))
            .filter_map(|run| Some((*run, run.took()?)))
            .collect()
    };
    let mut branches: BTreeMap<&str, usize> = BTreeMap::new();
    for run in runs.iter().filter(|run| run.succeeded()) {
        if let Some(branch) = &run.branch {
            *branches.entry(branch).or_default() += 1;
        }
    }
    // The default branch when it has runs enough, so a pull request that adds tests is not a regression; else the
    // most runs, and of equal counts the first by name.
    let busiest = branches
        .iter()
        .max_by(|a, b| a.1.cmp(b.1).then(b.0.cmp(a.0)))
        .map(|(branch, _)| *branch)?;
    let (branch, timed) = ["main", "master"]
        .into_iter()
        .map(|branch| (branch, timed_on(branch)))
        .find(|(_, timed)| timed.len() >= 2 * RECENT)
        .unwrap_or_else(|| (busiest, timed_on(busiest)));
    if timed.len() < 2 * RECENT {
        return None;
    }
    let (baseline, recent) = timed.split_at(timed.len() - RECENT);
    let took = |runs: &[(&Run, u64)]| runs.iter().map(|(_, took)| *took).collect::<Vec<_>>();
    let before = median(&took(baseline));
    let now = median(&took(recent));
    let threshold = (before + before * SLOWER_PERCENT / 100).max(before + SLOWER_SECONDS);
    // Every recent run slower than the baseline's median: a job that flips between two speeds, such as build cache
    // hits and misses, moves its median with a few more slow runs in a row without getting slower.
    if now < threshold || recent.iter().any(|(_, took)| *took <= before) {
        return None;
    }
    // The first slow run after which no two runs in a row were fast.
    let mut since = None;
    let mut fast_in_a_row = 0;
    for (run, took) in &timed {
        if *took >= threshold {
            fast_in_a_row = 0;
            since = since.or(Some(*run));
        } else {
            fast_in_a_row += 1;
            if fast_in_a_row >= 2 {
                since = None;
            }
        }
    }
    let since = since?;
    let sentence = format!(
        "{} on {branch} takes {} now, up from {}.",
        subject(key),
        span(now),
        span(before)
    );
    let links = recent.iter().rev().map(|(run, _)| run.link()).collect();
    Some(insight(
        key,
        "slower",
        Severity::Warning,
        sentence,
        InsightKind::Slower {
            branch: Some(branch.to_owned()),
            before,
            now,
            since: Some(since.link()),
        },
        links,
    ))
}

/// A job that failed and then passed on a re-run of the same run.
fn flaky(key: JobKey<'_>, runs: &[&Run]) -> Option<Insight> {
    let mut by_run: BTreeMap<u64, Vec<&Run>> = BTreeMap::new();
    for run in runs {
        by_run.entry(run.run_id).or_default().push(run);
    }
    let mut failed_first: Vec<&Run> = by_run
        .values()
        .filter_map(|attempts| {
            attempts.iter().copied().find(|failed| {
                failed.failed_itself()
                    && attempts
                        .iter()
                        .any(|later| later.run_attempt > failed.run_attempt && later.succeeded())
            })
        })
        .collect();
    if failed_first.is_empty() {
        return None;
    }
    failed_first.sort_by_key(|run| std::cmp::Reverse((run.finished(), run.id)));
    let flaky = u32::try_from(failed_first.len()).unwrap_or(u32::MAX);
    let total = u32::try_from(by_run.len()).unwrap_or(u32::MAX);
    let sentence = format!(
        "{} failed and then passed on a re-run in {flaky} of its {total} runs.",
        subject(key)
    );
    // Once in the window may be chance: a notice; twice, a warning.
    let severity = if flaky >= FLAKY_WARNING {
        Severity::Warning
    } else {
        Severity::Notice
    };
    Some(insight(
        key,
        "flaky",
        severity,
        sentence,
        InsightKind::Flaky { flaky, total },
        failed_first.iter().take(5).map(|run| run.link()).collect(),
    ))
}

/// A job whose last runs' memory peaks near its profile's `memory`: the profile of its last measured run, whose
/// runs alone count, as a job a repository serves in several profiles peaks apart in each.
fn near_cap(key: JobKey<'_>, runs: &[&Run], config: &Config) -> Option<Insight> {
    let measured = runs
        .iter()
        .rev()
        .filter(|run| run.memory_peak.is_some() && run.profile.is_some());
    let name = measured.clone().next()?.profile.as_deref()?;
    let measured: Vec<&Run> = measured
        .filter(|run| run.profile.as_deref() == Some(name))
        .take(CAP_RUNS)
        .copied()
        .collect();
    if measured.len() < CAP_MIN_RUNS {
        return None;
    }
    let profile = config.profiles.get(name)?;
    let cap = profile.memory.0;
    let peaks: Vec<u64> = measured.iter().filter_map(|run| run.memory_peak).collect();
    let near = |peak: &u64| peak.saturating_mul(100) >= cap.saturating_mul(NEAR_CAP);
    let peak = spread(peaks.clone()).p95;
    if cap == 0 || !near(&peak) {
        return None;
    }
    let near_runs = peaks.iter().filter(|peak| near(peak)).count();
    let auto = if !config.sizing.auto {
        AutoNote::Off
    } else if profile
        .max_memory
        .is_some_and(|max| profile.memory.0 >= max.0)
    {
        AutoNote::AtMax
    } else {
        AutoNote::Raises
    };
    let what = match auto {
        AutoNote::Raises => format!("Auto sizing raises {name} when a run meets the cap."),
        AutoNote::AtMax => {
            format!("{name} is at its max_memory, so auto sizing cannot raise it.")
        }
        AutoNote::Off => format!("Auto sizing is off: raise {name}'s memory, or turn it on."),
    };
    let sentence = format!(
        "{} peaks at {}, {}% of {name}'s {}. {what}",
        subject(key),
        ByteSize(peak),
        peak.saturating_mul(100) / cap,
        ByteSize(cap)
    );
    // Auto sizing deals with it, and one spike may be chance: a notice. The owner acts on the rest.
    let severity = if auto == AutoNote::Raises || near_runs < NEAR_CAP_WARNING {
        Severity::Notice
    } else {
        Severity::Warning
    };
    Some(insight(
        key,
        "near-cap",
        severity,
        sentence,
        InsightKind::NearCap {
            profile: name.to_owned(),
            peak,
            cap,
            auto,
        },
        measured.iter().take(5).map(|run| run.link()).collect(),
    ))
}

/// A job's last runs that could write to the build cache, each with this many lookups at least…
const COLD_RUNS: usize = 5;
const COLD_MIN_RUNS: usize = 3;
const COLD_MIN_LOOKUPS: u64 = 10;
/// …found at most this share of their lookups, in percent: the runs before them stored what they built, so the
/// next should find most of it.
const COLD_PERCENT: u64 = 10;

fn cold_cache(key: JobKey<'_>, runs: &[&Run]) -> Option<Insight> {
    let writers: Vec<&Run> = runs
        .iter()
        .rev()
        .filter(|run| {
            run.build_cache
                .is_some_and(|cache| cache.writer && cache.lookups() >= COLD_MIN_LOOKUPS)
        })
        .take(COLD_RUNS)
        .copied()
        .collect();
    if writers.len() < COLD_MIN_RUNS {
        return None;
    }
    let mut total = api::BuildCacheCounts::default();
    for run in &writers {
        total.add(&run.build_cache.unwrap_or_default());
    }
    let percent = total.hit_percent()?;
    if percent > COLD_PERCENT {
        return None;
    }
    let branch = writers.first()?.branch.clone();
    let sentence = format!(
        "{} got {percent}% of its build cache lookups from the cache{} over its last {} runs: its tasks' inputs \
         differ from run to run.",
        subject(key),
        branch
            .as_deref()
            .map(|branch| format!(" on {branch}"))
            .unwrap_or_default(),
        writers.len()
    );
    Some(insight(
        key,
        "cold-cache",
        Severity::Notice,
        sentence,
        InsightKind::ColdCache {
            branch,
            hits: total.hits,
            lookups: total.lookups(),
        },
        writers.iter().map(|run| run.link()).collect(),
    ))
}

/// A job whose CPU cap held back `RAISE_RUNS` or more of its last `CAP_RUNS` runs that measured it, as auto sizing
/// reads a run (`Throttling::held`): successful runs, as sizing reads only those, in the profile and at the cap of its
/// last such run, the profile's `cpus` today. A run without a throttling figure is unknown, neither held nor free.
/// Says what the owner can do: under `measured` admission auto sizing raises `cpus` while each raise helps, under
/// `caps` it does not for throttling, as each CPU takes budget there. A raise that did not help the job is told as
/// this job's evidence only: auto sizing may still raise the profile for another job, or on the runs' averages.
fn cpu_held(key: JobKey<'_>, runs: &[&Run], config: &Config) -> Option<Insight> {
    let measured = runs.iter().rev().filter(|run| {
        run.succeeded() && run.profile.is_some() && run.throttling().held().is_some()
    });
    let last = measured.clone().next()?;
    let (name, cap) = (last.profile.as_deref()?, last.cpu_cap?);
    let profile = config.profiles.get(name)?;
    // Runs at a cap the profile no longer has say nothing of the one it has: its next runs will.
    if u64::from(profile.cpus) * 1000 != cap {
        return None;
    }
    let at_cap: Vec<&Run> = measured
        .filter(|run| run.profile.as_deref() == Some(name) && run.cpu_cap == Some(cap))
        .take(CAP_RUNS)
        .copied()
        .collect();
    let held: Vec<&Run> = at_cap
        .iter()
        .filter(|run| run.throttling().held() == Some(true))
        .copied()
        .collect();
    if (held.len() as u64) < RAISE_RUNS {
        return None;
    }
    let cpus = profile.cpus;
    let auto = config.host.admission == Admission::Measured && config.sizing.auto;
    let raised_from = auto
        .then(|| {
            let ran: Vec<CpuRun> = runs
                .iter()
                .filter(|run| run.succeeded() && run.profile.as_deref() == Some(name))
                .map(|run| CpuRun {
                    cap: run.cpu_cap,
                    queued_at: run.queued_at,
                    throttling: run.throttling(),
                })
                .collect();
            let now: Vec<Throttling> = at_cap.iter().map(|run| run.throttling()).collect();
            unhelped_raise(&ran, cap, &now)
        })
        .flatten()
        .map(|unhelped| u32::try_from(unhelped.from / 1000).unwrap_or(u32::MAX));
    let set = (!auto).then(|| cpus.saturating_add(cpu_step(config, cpus)));
    let what = match (raised_from, set) {
        (Some(from), _) => format!(
            "The raise of {name} from {from} to {cpus} CPUs did not make it faster or less throttled, as when its \
             tools start a worker per CPU, so its runs ask for no further raise; runnerctl sizing says what auto \
             sizing does next."
        ),
        (None, Some(set)) if config.host.admission == Admission::Measured => {
            format!("Auto sizing is off: set {name}'s cpus to {set}, or turn it on.")
        }
        (None, Some(set)) => format!(
            "Under caps admission each CPU takes budget, so auto sizing does not raise cpus for this: set {name}'s \
             cpus to {set} if runnerctl sizing shows the budget has room."
        ),
        (None, None) => format!(
            "Auto sizing raises {name}'s cpus while each raise makes it faster or less throttled and the slice has \
             room; runnerctl sizing says when it cannot."
        ),
    };
    let sentence = format!(
        "{} was held back by {name}'s {cpus} CPUs in {} of its last {} runs, its quota throttling it in {HELD}% of \
         its run or more. {what}",
        subject(key),
        held.len(),
        at_cap.len()
    );
    // Auto sizing deals with it: a notice. The owner acts on the rest.
    let severity = if auto && raised_from.is_none() {
        Severity::Notice
    } else {
        Severity::Warning
    };
    Some(insight(
        key,
        "cpu-held",
        severity,
        sentence,
        InsightKind::CpuHeld {
            profile: name.to_owned(),
            cpus,
            held: u32::try_from(held.len()).unwrap_or(u32::MAX),
            total: u32::try_from(at_cap.len()).unwrap_or(u32::MAX),
            raised_from,
            set,
        },
        held.iter().take(5).map(|run| run.link()).collect(),
    ))
}

/// One insight per repository the kernel throttled `THROTTLED_MIN_RUNS` or more of the runs of in `window`.
fn throttled(window: &[&Run]) -> Vec<Insight> {
    let mut repositories: BTreeMap<&str, Vec<&Run>> = BTreeMap::new();
    for run in window {
        repositories.entry(&run.repository).or_default().push(run);
    }
    repositories
        .into_iter()
        .filter_map(|(repository, mut runs)| {
            runs.sort_by_key(|run| std::cmp::Reverse((run.finished(), run.id)));
            let hit: Vec<&Run> = runs
                .iter()
                .filter(|run| run.sock_throttled.is_some_and(|count| count > 0))
                .copied()
                .collect();
            let throttled = u32::try_from(hit.len()).unwrap_or(u32::MAX);
            if throttled < THROTTLED_MIN_RUNS {
                return None;
            }
            let total = u32::try_from(runs.len()).unwrap_or(u32::MAX);
            Some(Insight {
                key: format!("throttled:{repository}"),
                severity: Severity::Warning,
                repository: repository.to_owned(),
                workflow: None,
                job: None,
                sentence: format!(
                    "{repository}: memory pressure slowed the downloads of {throttled} of its {total} jobs: the \
                     host, or a job at its cap, was short of memory."
                ),
                kind: InsightKind::Throttled { throttled, total },
                runs: hit.iter().take(5).map(|run| run.link()).collect(),
            })
        })
        .collect()
}

/// Each repository's waits in the 7 days before `now` (ms) and the 7 before those, each counted from no earlier than
/// when the history holds every job. The week before stays empty when it starts before that: it would hold a sliver
/// of it, such as the daemon's first minutes. This week is what the history has of it either way.
fn waits(history: &History, now: u64) -> Vec<RepositoryWaits> {
    let before_kept = now.saturating_sub(2 * WEEK) < history.kept_since;
    let mut weeks: BTreeMap<&str, [Vec<&Run>; 2]> = BTreeMap::new();
    for run in &history.runs {
        let Some(started) = run.started_at.filter(|at| *at <= now) else {
            continue;
        };
        let week = match now - started {
            back if back < WEEK => 0,
            back if back < 2 * WEEK && !before_kept => 1,
            _ => continue,
        };
        weeks.entry(&run.repository).or_default()[week].push(run);
    }
    weeks
        .into_iter()
        .map(|(repository, [now, before])| RepositoryWaits {
            repository: repository.to_owned(),
            now: week_waits(&now, history.kept_since),
            before: week_waits(&before, history.kept_since),
        })
        .collect()
}

fn week_waits(runs: &[&Run], kept_since: u64) -> WeekWaits {
    let warm = |run: &&&Run| run.runner_started_at.is_some_and(|at| at <= run.queued_at);
    let of = |runs: Vec<&&Run>| {
        let waited: Vec<u64> = runs
            .iter()
            .filter_map(|run| run.waited(kept_since))
            .collect();
        let jobs = u32::try_from(waited.len()).unwrap_or(u32::MAX);
        let spread = (!waited.is_empty()).then(|| spread(waited));
        WaitSpread {
            jobs,
            p50: spread.as_ref().map(|spread| spread.p50),
            p95: spread.as_ref().map(|spread| spread.p95),
        }
    };
    WeekWaits {
        all: of(runs.iter().collect()),
        warm: of(runs.iter().filter(warm).collect()),
        cold: of(runs.iter().filter(|run| !warm(run)).collect()),
    }
}

fn waits_insight(waits: &RepositoryWaits) -> Option<Insight> {
    if waits.now.all.jobs < WAITS_MIN_JOBS || waits.before.all.jobs < WAITS_MIN_JOBS {
        return None;
    }
    let (before, now) = (waits.before.all.p95?, waits.now.all.p95?);
    let threshold = (before + before * WAITS_PERCENT / 100).max(before + WAITS_SECONDS);
    if now < threshold {
        return None;
    }
    Some(Insight {
        key: format!("waits:{}", waits.repository),
        severity: Severity::Notice,
        repository: waits.repository.clone(),
        workflow: None,
        job: None,
        sentence: if waits.now.all.jobs < WAITS_ONE_IN_20 {
            format!(
                "{}: the longest wait for a runner this week is {}, up from {} the week before.",
                waits.repository,
                span(now),
                span(before)
            )
        } else {
            format!(
                "{}: 1 job in 20 waits {} or more for a runner this week, up from {} the week before.",
                waits.repository,
                span(now),
                span(before)
            )
        },
        kind: InsightKind::Waits { before, now },
        runs: Vec::new(),
    })
}

/// A duration as a sentence reads it: `45s`, `4m 30s`, `1h 5m`.
fn span(seconds: u64) -> String {
    match seconds {
        0..60 => format!("{seconds}s"),
        60..3600 if seconds.is_multiple_of(60) => format!("{}m", seconds / 60),
        60..3600 => format!("{}m {}s", seconds / 60, seconds % 60),
        _ => format!("{}h {}m", seconds / 3600, seconds % 3600 / 60),
    }
}

#[cfg(test)]
mod tests;