-
Notifications
You must be signed in to change notification settings - Fork 9
/
graph.go
624 lines (519 loc) · 17.3 KB
/
graph.go
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
// Copyright 2022 Namespace Labs Inc; All rights reserved.
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
package compute
import (
"context"
"errors"
"fmt"
"reflect"
"sync"
"time"
"go.opentelemetry.io/otel/attribute"
"go.opentelemetry.io/otel/trace"
"namespacelabs.dev/foundation/internal/compute/cache"
"namespacelabs.dev/foundation/internal/console"
"namespacelabs.dev/foundation/internal/executor"
"namespacelabs.dev/foundation/internal/fnerrors"
"namespacelabs.dev/foundation/schema"
"namespacelabs.dev/foundation/std/tasks"
)
var (
// Configurable globally only for now.
CachingEnabled = true
// If enabled, does not use cached contents, but still verifies that if we do have
// cached contents, they match what we produced.
VerifyCaching = false
)
const (
outputCachingInformation = true
cleanerFuncLogLevel = 2
)
type contextKey string
var (
_graphKey = contextKey("fn.workspace.graph")
)
type Orch struct {
cache cache.Cache
origctx context.Context
exec executor.ExecutorLike
bestEffort executor.ExecutorLike
mu sync.Mutex
promises map[string]*Promise[any]
cleaners []cleaner
}
type cleaner struct {
ev *tasks.ActionEvent
f func(context.Context) error
}
func On(ctx context.Context) *Orch {
v := ctx.Value(_graphKey)
if v == nil {
return nil
}
return v.(*Orch)
}
var errNoResult = errors.New("no result")
func NoResult[V any]() (Result[V], error) { return Result[V]{}, errNoResult }
func startComputing(ctx context.Context, g *Orch, c UntypedComputable) *Promise[any] {
if g == nil {
// We panic because this is unexpected.
panic("no graph in context")
}
if c == nil {
return ErrPromise[any](fnerrors.InternalError("Computable is required"))
}
return startComputingWithOpts(ctx, g, c.prepareCompute(c))
}
func startComputingWithOpts(ctx context.Context, g *Orch, opts computeInstance) *Promise[any] {
if opts.IsPrecomputed {
// Caller guarantees to not block.
v, err := opts.Compute(ctx, Resolved{})
if err != nil {
return ErrPromise[any](err)
}
var r ResultWithTimestamp[any]
r.Value = v
if digestible, ok := opts.Computable.(Digestible); ok {
r.Digest, err = digestible.ComputeDigest(ctx)
if err != nil {
return ErrPromise[any](err)
}
}
return valueFuture(r)
}
computeInputs := opts.Inputs()
if computeInputs.named != nil {
promise, err := tasks.Return(ctx, opts.Action(), func(ctx context.Context) (*Promise[any], error) {
return startComputing(ctx, On(ctx), computeInputs.named), nil
})
if err != nil {
return ErrPromise[any](err)
}
return promise
}
// XXX not really happy with passing ctx here, rather than using the top-level graph
// context. The attribution will be odd. However, submitting another task at this
// point is also a stretch.
inputs, err := computeInputs.computeDigest(ctx, opts.Computable, true)
if err != nil {
return ErrPromise[any](err)
}
// Any computation which yields is keyed on the same inputs shares its output.
if inputs.Digest.IsSet() {
p, isRunning := ensurePromise(g, opts.Computable, inputs)
if !isRunning {
deferCompute(g, p, opts, inputs)
}
return p
} else {
if opts.IsGlobal {
panic(fmt.Sprintf("%s: global node that doesn't have stable inputs", reflect.TypeOf(opts.Computable).String()))
}
}
opts.State.promise.mu.Lock()
compute := false
if !opts.State.running {
initializePromise(&opts.State.promise, opts.Computable, tasks.NewActionID().String())
opts.State.running = true
compute = true
}
opts.State.promise.mu.Unlock()
// If another path is already computing the value, waiting on the returned promise will still block.
if compute {
// The return value is ignored here because the promise will be resolved
// by waitCompute, and thus it's value will be returned below.
_ = waitCompute(ctx, g, &opts.State.promise, opts, inputs)
}
return &opts.State.promise
}
func ensurePromise(g *Orch, c hasAction, inputs *computedInputs) (*Promise[any], bool) {
g.mu.Lock()
key := inputs.Digest.String()
p, isComputing := g.promises[key]
if !isComputing {
p = makePromise[any](c, key)
g.promises[key] = p
}
g.mu.Unlock()
if isComputing {
// A computation is already running for this inputs; we must return the same
// instance so we re-use the same action ID; this will provide us with the
// a link between future (and their waiting time), and the actual computation
// as the promise stores the action ID.
// XXX re-do this to just use the digest values themselves for action tracking.
return p, true
}
// Returns false if we're not currently computing this Computable.
return p, false
}
func deferCompute(g *Orch, p *Promise[any], opts computeInstance, inputs *computedInputs) {
g.exec.Go(func(ctx context.Context) error {
return waitCompute(ctx, g, p, opts, inputs)
})
}
func waitCompute(ctx context.Context, g *Orch, p *Promise[any], opts computeInstance, inputs *computedInputs) error {
cacheable, shouldCache := opts.CacheInfo()
// Computables are cacheable (if they don't opt-out, and rely on deterministic inputs and outputs).
// The cache is a simple a content-addressible filesystem, where a digest of the output points to
// its contents. A separate index is kept that maps "inputs" digest to output digest. There are two
// types of input digests: "complete" and "incomplete". "complete" digests are produced by computing
// recursively the digest of each dependency, provided it itself is complete. A leaf Computable
// produces "complete" digests if all of its inputs are known and deterministic ahead of time. On
// the other hand "incomplete" digests are computed using the digest of the output of a Computable
// which doesn't have deterministic inputs. To ensuring we minimize cost while loading, two index
// entries are maintained pointing at the output: a "complete" one if available, and the
// "incomplete" one.
var resolved *Resolved
var hits []cacheHit
ev := opts.Action()
name, _ := tasks.NameOf(ev)
if err := ev.ID(p.actionID).RunWithOpts(ctx, tasks.RunOpts{
Wait: func(ctx context.Context) (bool, error) {
// If we've already calculated an inputs' digest, then attempt to load from the cache
// directly. If not, we'll need to wait on our dependencies to determine whether a
// complete digest is available then.
hit := checkCache(ctx, g, opts, cacheable, shouldCache, inputs, p)
if VerifyCaching {
hits = append(hits, hit)
}
if hit.VerifiedHit {
return true, nil
}
// If we come in through the "digest-compute" path, then we've already computed the results.
results, err := waitDeps(ctx, g, name, inputs.computable)
if err != nil {
return false, err
}
if outputCachingInformation {
addOutputsToSpan(ctx, results)
}
// Compute a new "inputs" digest based on the resolved future outputs. This
// provides a stable identifier we can cache on. Used below as well in `deferStore`.
if err := inputs.Finalize(results); err != nil {
return false, err
}
if outputCachingInformation {
span := trace.SpanFromContext(ctx)
span.SetAttributes(attribute.Stringer("fn.inputs.postcompute.digest", inputs.PostComputeDigest))
}
if shouldCache && inputs.PostComputeDigest.IsSet() {
// Errors are ignored in cache loading.
if hit, err := checkLoadCache(ctx, "cache.load.post", g, opts, cacheable, inputs.PostComputeDigest, p); err == nil && hit.Hit {
if VerifyCaching {
hit.Inputs = inputs
hits = append(hits, hit)
}
if hit.VerifiedHit {
return true, nil
}
}
}
resolved = &Resolved{
results: results,
}
return false, nil
},
Run: func(ctx context.Context) error {
res, err := compute(ctx, g, p.actionID, opts, cacheable, shouldCache, inputs, *resolved)
if err != nil {
return err
}
if VerifyCaching {
verifyCacheHits(ctx, opts.Computable, hits, res.Digest)
}
return p.resolve(res, nil)
},
}); err != nil {
return p.fail(err)
}
return nil
}
func compute(ctx context.Context, g *Orch, actionID tasks.ActionID, opts computeInstance, cacheable *cacheable, shouldCache bool, inputs *computedInputs, resolved Resolved) (ResultWithTimestamp[any], error) {
started := time.Now()
v, err := opts.Compute(ctx, resolved)
if err != nil {
return ResultWithTimestamp[any]{}, err
}
completed := time.Now()
var digester ComputeDigestFunc
if digester == nil && cacheable != nil {
digester = cacheable.ComputeDigest
}
d, err := computeOutputDigest(ctx, digester, v)
if err != nil {
d = schema.Digest{} // Ignore errors, but don't cache.
if VerifyCaching {
fmt.Fprintf(console.Errors(ctx), "VerifyCache: failed to compute digest for %q: %v", typeStr(opts.Computable), err)
}
}
if shouldCache && d.IsSet() {
deferStore(ctx, g, opts.Computable, cacheable, d, completed, v, inputs)
}
if outputCachingInformation {
trace.SpanFromContext(ctx).SetAttributes(attribute.Stringer("fn.output.digest", d))
}
return ResultWithTimestamp[any]{
Result: Result[any]{
Digest: d,
NonDeterministic: opts.NonDeterministic,
Value: v,
},
ActionID: actionID,
Started: started,
Completed: completed,
}, nil
}
func checkCache(ctx context.Context, g *Orch, opts computeInstance, cacheable *cacheable, shouldCache bool, inputs *computedInputs, p *Promise[any]) cacheHit {
if outputCachingInformation {
addInputsToSpan(ctx, opts.Inputs(), inputs, shouldCache)
}
if !shouldCache || !inputs.Digest.IsSet() {
return cacheHit{}
}
// Errors are ignored in cache loading.
if hit, err := checkLoadCache(ctx, "cache.load.pre", g, opts, cacheable, inputs.Digest, p); err == nil && hit.Hit {
if VerifyCaching {
hit.Inputs = inputs
}
return hit
}
return cacheHit{}
}
func waitDeps(ctx context.Context, g *Orch, desc string, computable map[string]UntypedComputable) (map[string]ResultWithTimestamp[any], error) {
if len(computable) == 0 {
return nil, nil
}
var rmu sync.Mutex // Protects resolved and digests.
// We wait in parallel to create N actions so that the full dependency
// graph is also visible in the action log. This is a bit wasteful though
// and should be rethinked.
eg := executor.Newf(ctx, "compute.wait-deps(%s, %d deps)", desc, len(computable))
results := map[string]ResultWithTimestamp[any]{}
for k, d := range computable {
k := k // Close k.
d := d // Close d.
eg.Go(func(ctx context.Context) error {
res, err := startComputing(ctx, g, d).Future().Wait(ctx)
if err != nil {
// Make sure this is reported as one of the dependencies failing, instead of this
// computation. This will provide for better error reporting.
return fnerrors.DependencyFailed(k, reflect.TypeOf(d).String(), err)
}
rmu.Lock()
results[k] = res
rmu.Unlock()
return nil
})
}
// XXX think through this, we're throwing the same errors all over the place.
// Probably just want a "dependency didn't compute" error here which feels like
// a cancellation.
err := eg.Wait()
return results, err
}
func computeOutputDigest(ctx context.Context, digester ComputeDigestFunc, v interface{}) (schema.Digest, error) {
if digester != nil {
return digester(ctx, v)
}
if cd, ok := v.(Digestible); ok {
return cd.ComputeDigest(ctx)
}
return schema.Digest{}, nil
}
func (g *Orch) Detach(ev *tasks.ActionEvent, f func(context.Context) error) {
g.DetachWith(Detach{
Action: ev,
Do: f,
})
}
type Detach struct {
Action *tasks.ActionEvent
Do func(context.Context) error
}
func (g *Orch) DetachWith(d Detach) {
if g == nil {
// We panic because this is unexpected.
panic("running outside of a compute.Do block")
}
g.exec.Go(func(ctx context.Context) error {
err := d.Action.Run(ctx, d.Do)
if errors.Is(err, context.Canceled) {
return nil
}
if err != nil {
fmt.Fprintf(console.Warnings(ctx), "detach failed: %v\n", err)
return nil // Ignore errors.
}
return err
})
}
func (g *Orch) BestEffort(ev *tasks.ActionEvent, do func(context.Context) error) {
if g == nil {
// We panic because this is unexpected.
panic("running outside of a compute.Do block")
}
g.bestEffort.Go(func(ctx context.Context) error {
err := ev.Run(ctx, do)
if errors.Is(err, context.Canceled) {
return nil
}
if err != nil {
fmt.Fprintf(console.Warnings(ctx), "best effort task failed: %v\n", err)
}
return nil // Ignore errors.
})
}
func (g *Orch) Cleanup(ev *tasks.ActionEvent, f func(context.Context) error) {
// XXX check if Cleanup() is called after we're done.
g.mu.Lock()
g.cleaners = append(g.cleaners, cleaner{ev, f})
g.mu.Unlock()
}
func (g *Orch) Call(callback func(context.Context) error) error {
errCh := make(chan error)
g.exec.Go(func(ctx context.Context) error {
defer close(errCh)
errCh <- callback(ctx)
return nil // We never fail the parent computation.
})
err, ok := <-errCh
if !ok {
return fnerrors.New("call was canceled?")
}
return err
}
func WithGraphLifecycle[V any](ctx context.Context, f func(context.Context) (V, error)) (V, error) {
g := On(ctx)
if g == nil {
var empty V
return empty, fnerrors.New("no graph in context")
}
return f(g.origctx)
}
func Cache(ctx context.Context) cache.Cache {
return On(ctx).cache
}
func AttachOrch(parent context.Context, orch *Orch) context.Context {
return context.WithValue(parent, _graphKey, orch)
}
func Do(parent context.Context, do func(context.Context) error) error {
parentOrch := On(parent)
if tasks.SinkFrom(parent) == nil {
panic("compute: action sink required in the context")
}
var c cache.Cache
if parentOrch != nil {
c = parentOrch.cache
} else {
var err error
c, err = cache.Local()
if err != nil {
return err
}
}
return DoWithCache(parent, c, do)
}
func DoWithCache(parent context.Context, cache cache.Cache, do func(context.Context) error) error {
g := &Orch{
cache: cache,
promises: map[string]*Promise[any]{},
}
ctx := AttachOrch(parent, g)
exec := executor.New(ctx, "compute.Do")
g.origctx = ctx
g.exec = exec
g.bestEffort = executor.New(ctx, "compute.Do.BestEffort")
// We execute do in the executor instead of directly, to ensure that error
// propagation is correct; i.e. if a separate branch ends up failing, we
// should see that error rather than the context cancelation that do() would
// otherwise.
exec.Go(do)
// Importantly, call `wait` before returning to make sure that any deferred work gets concluded.
errResult := exec.Wait()
_ = g.bestEffort.CancelAndWait()
g.mu.Lock()
cleaners := g.cleaners
g.cleaners = nil
g.mu.Unlock()
// XXX parallelize cleanups.
// Importantly, graph is not present in the context when calling a cleaner function. And we always
// run cleaners, regardless of errors above.
for _, c := range cleaners {
if err := c.ev.LogLevel(cleanerFuncLogLevel).Run(parent, c.f); err != nil {
if errResult == nil {
errResult = err
}
}
}
return errResult
}
func Get[V any](ctx context.Context, c Computable[V]) (ResultWithTimestamp[V], error) {
promise := startComputing(ctx, On(ctx), c)
r, err := promise.Future().Wait(ctx)
if err != nil {
return ResultWithTimestamp[V]{}, err
}
typed, ok := r.Value.(V)
if !ok {
panic(fmt.Sprintf("Computable of type %s produced wrong type %s", reflect.TypeOf(c), reflect.TypeOf(r.Value).String()))
}
var rwt ResultWithTimestamp[V]
rwt.Value = typed
rwt.Digest = r.Digest
rwt.Cached = r.Cached
rwt.NonDeterministic = r.NonDeterministic
rwt.Started = r.Started
rwt.Completed = r.Completed
return rwt, nil
}
func GetValue[V any](ctx context.Context, c Computable[V]) (V, error) {
v, err := Get(ctx, c)
return v.Value, err
}
func addInputsToSpan(ctx context.Context, in *In, inputs *computedInputs, shouldCache bool) {
span := trace.SpanFromContext(ctx)
span.SetAttributes(attribute.Bool("fn.input.cacheable", in.cacheable))
for _, input := range in.ins {
span.SetAttributes(attribute.Bool(fmt.Sprintf("fn.input.%s.undetermined", input.Name), input.Undetermined))
}
for _, input := range inputs.digests {
digest := input.Digest
if digest == "" {
digest = "<unset>"
}
span.SetAttributes(attribute.String(fmt.Sprintf("fn.input.%s.digest", input.Name), digest))
}
span.SetAttributes(attribute.Bool("fn.inputs.nonDeterministic", inputs.nonDeterministic))
span.SetAttributes(attribute.Bool("fn.shouldCache", shouldCache))
}
func addOutputsToSpan(ctx context.Context, results map[string]ResultWithTimestamp[any]) {
span := trace.SpanFromContext(ctx)
for k, res := range results {
span.SetAttributes(attribute.Stringer(fmt.Sprintf("fn.output.%s.digest", k), res.Digest))
}
}
func verifyCacheHits(ctx context.Context, c UntypedComputable, hits []cacheHit, d schema.Digest) {
for _, hit := range hits {
if hit.Hit && hit.OutputDigest != d {
console.WriteJSON(console.Errors(ctx),
fmt.Sprintf("VerifyCache: found non-determinism evaluating %q", typeStr(c)),
map[string]interface{}{
"expected": hit.OutputDigest,
"got": d,
"matching": hit.Input,
"inputs.digest": hit.Inputs.Digest,
"inputs.postDigest": hit.Inputs.PostComputeDigest,
"inputs.digests": hit.Inputs.digests,
"inputs.nonDeterministic": hit.Inputs.nonDeterministic,
})
_ = Explain(ctx, console.Debug(ctx), c)
}
}
}
func typeStr(v interface{}) string {
if v == nil {
return "(nil)"
}
return reflect.TypeOf(v).String()
}