-
Notifications
You must be signed in to change notification settings - Fork 1.1k
/
autoscale.go
451 lines (395 loc) · 15.4 KB
/
autoscale.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
/*
Copyright 2020 The Knative Authors
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
*/
package e2e
import (
"context"
"fmt"
"math"
"net/http"
"net/url"
"strconv"
"strings"
"testing"
"time"
"github.com/davecgh/go-spew/spew"
vegeta "github.com/tsenart/vegeta/v12/lib"
"golang.org/x/sync/errgroup"
appsv1 "k8s.io/api/apps/v1"
corev1 "k8s.io/api/core/v1"
"k8s.io/apimachinery/pkg/api/resource"
metav1 "k8s.io/apimachinery/pkg/apis/meta/v1"
pkgTest "knative.dev/pkg/test"
"knative.dev/pkg/test/spoof"
"knative.dev/serving/pkg/apis/autoscaling"
resourcenames "knative.dev/serving/pkg/reconciler/revision/resources/names"
rtesting "knative.dev/serving/pkg/testing/v1"
"knative.dev/serving/test"
v1test "knative.dev/serving/test/v1"
)
const (
// Concurrency must be high enough to avoid the problems with sampling
// but not high enough to generate scheduling problems.
containerConcurrency = 6
rpsTarget = 10
targetUtilization = 0.7
successRateSLO = 0.999
autoscaleSleep = 500
autoscaleTestImageName = "autoscale"
)
// TestContext includes context for autoscaler testing.
type TestContext struct {
t *testing.T
clients *test.Clients
names *test.ResourceNames
resources *v1test.ResourceObjects
autoscaler *AutoscalerOptions
}
// AutoscalerOptions holds autoscaling parameters for knative service.
type AutoscalerOptions struct {
Class string
Metric string
TargetUtilization float64
Target int
}
// Clients returns the clients of the TestContext.
func (ctx *TestContext) Clients() *test.Clients {
return ctx.clients
}
// Resources returns the resources of the TestContext.
func (ctx *TestContext) Resources() *v1test.ResourceObjects {
return ctx.resources
}
// SetResources sets the resources of the TestContext to the given values.
func (ctx *TestContext) SetResources(resources *v1test.ResourceObjects) {
ctx.resources = resources
}
// Names returns the resource names of the TestContext.
func (ctx *TestContext) Names() *test.ResourceNames {
return ctx.names
}
// SetNames set the resource names of the TestContext to the given values.
func (ctx *TestContext) SetNames(names *test.ResourceNames) {
ctx.names = names
}
func getVegetaTarget(domain string, paramName string, paramValue int, https bool) vegeta.Target {
scheme := "http"
if https {
scheme = "https"
}
return vegeta.Target{
Method: http.MethodGet,
URL: fmt.Sprintf("%s://%s?%s=%d", scheme, domain, paramName, paramValue),
}
}
func generateTraffic(
ctx *TestContext,
attacker *vegeta.Attacker,
pacer vegeta.Pacer,
stopChan chan struct{},
target vegeta.Target) error {
// The 0 duration means that the attack will only be controlled by the `Stop` function.
results := attacker.Attack(vegeta.NewStaticTargeter(target), pacer, 0, "load-test")
defer attacker.Stop()
var (
totalRequests int32
successfulRequests int32
)
for {
select {
case <-stopChan:
ctx.t.Logf("Stopping generateTraffic")
successRate := float64(1)
if totalRequests > 0 {
successRate = float64(successfulRequests) / float64(totalRequests)
}
if successRate < successRateSLO {
return fmt.Errorf("request success rate under SLO: total = %d, errors = %d, rate = %f, SLO = %f",
totalRequests, totalRequests-successfulRequests, successRate, successRateSLO)
}
return nil
case res := <-results:
totalRequests++
if res.Code != http.StatusOK {
ctx.t.Logf("Status = %d, want: 200", res.Code)
ctx.t.Logf("URL: %s Start: %s End: %s Duration: %v Error: %s Body:\n%s",
res.URL, res.Timestamp.Format(time.RFC3339), res.End().Format(time.RFC3339), res.Latency, res.Error, string(res.Body))
continue
}
successfulRequests++
}
}
}
func newVegetaHTTPClient(ctx *TestContext, url *url.URL) *http.Client {
vegetaTransportDefaults := func(transport *http.Transport) *http.Transport {
transport.MaxIdleConnsPerHost = vegeta.DefaultConnections
transport.MaxConnsPerHost = vegeta.DefaultMaxConnections
return transport
}
spoof, err := pkgTest.NewSpoofingClient(
context.Background(),
ctx.Clients().KubeClient,
ctx.t.Logf,
url.Hostname(),
test.ServingFlags.ResolvableDomain,
test.AddRootCAtoTransport(context.Background(), ctx.t.Logf, ctx.Clients(), test.ServingFlags.HTTPS),
vegetaTransportDefaults,
)
if err != nil {
ctx.t.Fatal("Error creating spoofing client:", err)
}
return spoof.Client
}
func generateTrafficAtFixedConcurrency(ctx *TestContext, concurrency int, stopChan chan struct{}) error {
pacer := vegeta.ConstantPacer{} // Sends requests as quickly as possible, capped by MaxWorkers below.
attacker := vegeta.NewAttacker(
vegeta.Timeout(0), // No timeout is enforced at all.
vegeta.Workers(uint64(concurrency)),
vegeta.MaxWorkers(uint64(concurrency)),
vegeta.Client(newVegetaHTTPClient(ctx, ctx.resources.Route.Status.URL.URL())),
)
target := getVegetaTarget(ctx.resources.Route.Status.URL.URL().Hostname(), "sleep", autoscaleSleep, test.ServingFlags.HTTPS)
ctx.t.Logf("Maintaining %d concurrent requests.", concurrency)
return generateTraffic(ctx, attacker, pacer, stopChan, target)
}
func generateTrafficAtFixedRPS(ctx *TestContext, rps int, stopChan chan struct{}) error {
pacer := vegeta.ConstantPacer{Freq: rps, Per: time.Second}
attacker := vegeta.NewAttacker(
vegeta.Timeout(0),
vegeta.Client(newVegetaHTTPClient(ctx, ctx.resources.Route.Status.URL.URL())),
)
target := getVegetaTarget(ctx.resources.Route.Status.URL.URL().Hostname(), "sleep", autoscaleSleep, test.ServingFlags.HTTPS)
ctx.t.Logf("Maintaining %v RPS.", rps)
return generateTraffic(ctx, attacker, pacer, stopChan, target)
}
func toPercentageString(f float64) string {
return strconv.FormatFloat(f*100, 'f', -1, 64)
}
// SetupSvc creates a new service, with given service options.
// It returns a TestContext that has resources, K8s clients and other needed
// data points.
func SetupSvc(t *testing.T, aopts *AutoscalerOptions, topts test.Options, fopts ...rtesting.ServiceOption) *TestContext {
t.Helper()
clients := test.Setup(t, topts)
t.Log("Creating a new Route and Configuration")
names := &test.ResourceNames{
Service: test.ObjectNameForTest(t),
Image: autoscaleTestImageName,
}
resources, err := v1test.CreateServiceReady(t, clients, names,
append([]rtesting.ServiceOption{
rtesting.WithConfigAnnotations(map[string]string{
autoscaling.ClassAnnotationKey: aopts.Class,
autoscaling.MetricAnnotationKey: aopts.Metric,
autoscaling.TargetAnnotationKey: strconv.Itoa(aopts.Target),
autoscaling.TargetUtilizationPercentageKey: toPercentageString(aopts.TargetUtilization),
// Reduce the amount of historical data we need before scaling down to account for
// the fact that the chaosduck will only let a bucket leader live for ~30s. This
// value still allows the chaosduck to make us failover, but is low enough that we
// should not need to survive multiple rounds of chaos in order to scale a
// revision down.
autoscaling.WindowAnnotationKey: "30s",
}),
rtesting.WithResourceRequirements(corev1.ResourceRequirements{
Requests: corev1.ResourceList{
corev1.ResourceCPU: resource.MustParse("30m"),
corev1.ResourceMemory: resource.MustParse("20Mi"),
},
}),
}, fopts...)...)
if err != nil {
t.Fatalf("Failed to create initial Service: %v: %v", names.Service, err)
}
if _, err := pkgTest.CheckEndpointState(
context.Background(),
clients.KubeClient,
t.Logf,
names.URL,
spoof.MatchesAllOf(spoof.IsStatusOK),
"CheckingEndpointAfterCreate",
test.ServingFlags.ResolvableDomain,
test.AddRootCAtoTransport(context.Background(), t.Logf, clients, test.ServingFlags.HTTPS),
); err != nil {
t.Fatalf("Error probing %s: %v", names.URL.Hostname(), err)
}
return &TestContext{
t: t,
clients: clients,
names: names,
resources: resources,
autoscaler: aopts,
}
}
func assertScaleDown(ctx *TestContext) {
deploymentName := resourcenames.Deployment(ctx.resources.Revision)
if err := WaitForScaleToZero(ctx.t, deploymentName, ctx.clients); err != nil {
ctx.t.Fatalf("Unable to observe the Deployment named %s scaling down: %v", deploymentName, err)
}
// Account for the case where scaling up uses all available pods.
ctx.t.Logf("Wait for all pods to terminate.")
if err := pkgTest.WaitForPodListState(
context.Background(),
ctx.clients.KubeClient,
func(p *corev1.PodList) (bool, error) {
for i := range p.Items {
pod := &p.Items[i]
if strings.Contains(pod.Name, deploymentName) &&
!strings.Contains(pod.Status.Reason, "Evicted") {
return false, nil
}
}
return true, nil
},
"WaitForAvailablePods", test.ServingFlags.TestNamespace); err != nil {
ctx.t.Fatalf("Waiting for Pod.List to have no non-Evicted pods of %q: %v", deploymentName, err)
}
ctx.t.Logf("The Revision should remain ready after scaling to zero.")
if err := v1test.CheckRevisionState(ctx.clients.ServingClient, ctx.names.Revision, v1test.IsRevisionReady); err != nil {
ctx.t.Fatalf("The Revision %s did not stay Ready after scaling down to zero: %v", ctx.names.Revision, err)
}
ctx.t.Logf("Scaled down.")
}
func numberOfReadyPods(ctx *TestContext) (float64, *appsv1.Deployment, error) {
n := resourcenames.Deployment(ctx.resources.Revision)
deploy, err := ctx.clients.KubeClient.AppsV1().Deployments(test.ServingFlags.TestNamespace).Get(
context.Background(), n, metav1.GetOptions{})
if err != nil {
return 0, nil, fmt.Errorf("failed to get deployment %s: %w", n, err)
}
if isInRollout(deploy) {
// Ref: #11092
// The deployment was updated and the update is being rolled out so we defensively
// pick the desired replicas to assert the autoscaling decisions.
// TODO: Drop this once we solved the underscale issue.
ctx.t.Logf("Deployment is being rolled, picking spec.replicas=%d", *deploy.Spec.Replicas)
return float64(*deploy.Spec.Replicas), deploy, nil
}
// Otherwise we pick the ready pods to assert maximum consistency for ramp up tests.
return float64(deploy.Status.ReadyReplicas), deploy, nil
}
func checkPodScale(ctx *TestContext, targetPods, minPods, maxPods float64, done <-chan time.Time, quick bool) error {
// Short-circuit traffic generation once we exit from the check logic.
ticker := time.NewTicker(2 * time.Second)
defer ticker.Stop()
originalMaxPods := maxPods
for {
select {
case <-ticker.C:
// Each 2 second, check that the number of pods is at least `minPods`. `minPods` is increasing
// to verify that the number of pods doesn't go down while we are scaling up.
got, d, err := numberOfReadyPods(ctx)
if err != nil {
return err
}
if isInRollout(d) {
// Ref: #11092
// Allow for a higher scale if the deployment is being rolled as that
// might be skewing metrics in the autoscaler.
maxPods = math.Ceil(originalMaxPods * 1.2)
}
mes := fmt.Sprintf("revision %q #replicas: %v, want at least: %v", ctx.resources.Revision.Name, got, minPods)
ctx.t.Logf(mes)
// verify that the number of pods doesn't go down while we are scaling up.
if got < minPods {
return fmt.Errorf("interim scale didn't fulfill constraints: %s\ndeployment state: %s", mes, spew.Sdump(d))
}
// A quick test succeeds when the number of pods scales up to `targetPods`
// (and, as an extra check, no more than `maxPods`).
if quick && got >= targetPods && got <= maxPods {
ctx.t.Logf("Quick Mode: got %v >= %v", got, targetPods)
return nil
}
if minPods < targetPods-1 {
// Increase `minPods`, but leave room to reduce flakiness.
minPods = math.Min(got, targetPods) - 1
}
case <-done:
// The test duration is over. Do a last check to verify that the number of pods is at `targetPods`
// (with a little room for de-flakiness).
got, d, err := numberOfReadyPods(ctx)
if err != nil {
return fmt.Errorf("failed to fetch number of ready pods: %w", err)
}
if isInRollout(d) {
// Ref: #11092
// Allow for a higher scale if the deployment is being rolled as that
// might be skewing metrics in the autoscaler.
maxPods = math.Ceil(originalMaxPods * 1.2)
}
mes := fmt.Sprintf("revision %q #replicas: %v, want between [%v, %v]", ctx.resources.Revision.Name, got, targetPods-1, maxPods)
ctx.t.Logf(mes)
if got < targetPods-1 || got > maxPods {
return fmt.Errorf("final scale didn't fulfill constraints: %s\ndeployment state: %s", mes, spew.Sdump(d))
}
return nil
}
}
}
// AssertAutoscaleUpToNumPods asserts the number of pods gets scaled to targetPods.
// It supports two test modes: quick, and not quick.
// 1. Quick mode: succeeds when the number of pods meets targetPods.
// 2. Not Quick (sustaining) mode: succeeds when the number of pods gets scaled to targetPods and
// sustains there until the `done` channel sends a signal.
func AssertAutoscaleUpToNumPods(ctx *TestContext, curPods, targetPods float64, done <-chan time.Time, quick bool) {
ctx.t.Helper()
wait := AutoscaleUpToNumPods(ctx, curPods, targetPods, done, quick)
if err := wait(); err != nil {
ctx.t.Fatal(err)
}
}
// AutoscaleUpToNumPods starts the traffic for AssertAutoscaleUpToNumPods and returns
// a function to wait for which will return any error from test execution.
// Starting the routines is separated from waiting for easy re-use in other
// places (e.g. upgrade tests).
func AutoscaleUpToNumPods(ctx *TestContext, curPods, targetPods float64, done <-chan time.Time, quick bool) func() error {
ctx.t.Helper()
// Relax the bounds to reduce the flakiness caused by sampling in the autoscaling algorithm.
// Also adjust the values by the target utilization values.
minPods := math.Floor(curPods/ctx.autoscaler.TargetUtilization) - 1
maxPods := math.Ceil(math.Ceil(targetPods/ctx.autoscaler.TargetUtilization) * 1.1)
stopChan := make(chan struct{})
var grp errgroup.Group
grp.Go(func() error {
switch ctx.autoscaler.Metric {
case autoscaling.RPS:
return generateTrafficAtFixedRPS(ctx, int(targetPods*float64(ctx.autoscaler.Target)), stopChan)
default:
return generateTrafficAtFixedConcurrency(ctx, int(targetPods*float64(ctx.autoscaler.Target)), stopChan)
}
})
grp.Go(func() error {
defer close(stopChan)
return checkPodScale(ctx, targetPods, minPods, maxPods, done, quick)
})
return grp.Wait
}
// isInRollout is a loose copy of the kubectl function handling rollouts.
// See: https://github.com/kubernetes/kubectl/blob/0149779a03735a5d483115ca4220a7b6c861430c/pkg/polymorphichelpers/rollout_status.go#L75-L91
func isInRollout(deploy *appsv1.Deployment) bool {
if deploy.Generation > deploy.Status.ObservedGeneration {
// Waiting for update to be observed.
return true
}
if deploy.Spec.Replicas != nil && deploy.Status.UpdatedReplicas < *deploy.Spec.Replicas {
// Not enough replicas updated yet.
return true
}
if deploy.Status.Replicas > deploy.Status.UpdatedReplicas {
// Old replicas are being terminated.
return true
}
if deploy.Status.AvailableReplicas < deploy.Status.UpdatedReplicas {
// Not enough available yet.
return true
}
return false
}