Implement support of horizontal scaling rules for Autoscaling #27119

vboulineau · 2024-06-27T12:58:38Z

What does this PR do?

Implement support of horizontal scaling rules for Autoscaling, and add test suite for horizontal scaling.

Motivation

Feature

Additional Notes

Dedicated commit to set model fields to private, making sure all updates go through functions to track modifications.

Possible Drawbacks / Trade-offs

Describe how to test/QA your changes

QA already done as part of autoscaling release.

pr-commenter · 2024-06-27T16:59:43Z

Test changes on VM

Use this command from test-infra-definitions to manually test this PR changes on a VM:

inv create-vm --pipeline-id=38457337 --os-family=ubuntu

Note: This applies to commit 862d440

pr-commenter · 2024-06-27T17:27:31Z

Regression Detector

Regression Detector Results

Run ID: 6f4f6688-ca96-4a68-a494-48aed8829fe8 Metrics dashboard Target profiles

Baseline: c37797d
Comparison: 2c43fcd

Performance changes are noted in the perf column of each table:

✅ = significantly better comparison variant performance
❌ = significantly worse comparison variant performance
➖ = no significant change in performance

No significant changes in experiment optimization goals

Confidence level: 90.00%
Effect size tolerance: |Δ mean %| ≥ 5.00%

There were no significant changes in experiment optimization goals at this confidence level and effect size tolerance.

Fine details of change detection per experiment

perf	experiment	goal	Δ mean %	Δ mean % CI	links
➖	basic_py_check	% cpu utilization	+3.08	[+0.40, +5.76]	Logs
➖	tcp_syslog_to_blackhole	ingress throughput	+2.34	[-10.56, +15.23]	Logs
➖	uds_dogstatsd_to_api_cpu	% cpu utilization	+0.81	[-0.08, +1.70]	Logs
➖	file_tree	memory utilization	+0.44	[+0.39, +0.49]	Logs
➖	uds_dogstatsd_to_api	ingress throughput	-0.00	[-0.00, +0.00]	Logs
➖	tcp_dd_logs_filter_exclude	ingress throughput	-0.00	[-0.01, +0.01]	Logs
➖	idle	memory utilization	-0.38	[-0.42, -0.35]	Logs
➖	pycheck_1000_100byte_tags	% cpu utilization	-0.50	[-5.02, +4.01]	Logs
➖	otel_to_otel_logs	ingress throughput	-0.87	[-1.68, -0.07]	Logs

Explanation

A regression test is an A/B test of target performance in a repeatable rig, where "performance" is measured as "comparison variant minus baseline variant" for an optimization goal (e.g., ingress throughput). Due to intrinsic variability in measuring that goal, we can only estimate its mean value for each experiment; we report uncertainty in that value as a 90.00% confidence interval denoted "Δ mean % CI".

For each experiment, we decide whether a change in performance is a "regression" -- a change worth investigating further -- if all of the following criteria are true:

Its estimated |Δ mean %| ≥ 5.00%, indicating the change is big enough to merit a closer look.
Its 90.00% confidence interval "Δ mean % CI" does not contain zero, indicating that if our statistical model is accurate, there is at least a 90.00% chance there is a difference in performance between baseline and comparison variants.
Its configuration does not mark it "erratic".

pkg/clusteragent/autoscaling/processor.go

AliDatadog · 2024-07-05T09:10:20Z

pkg/clusteragent/autoscaling/workload/controller_horizontal.go

@@ -55,15 +57,24 @@ func newHorizontalReconciler(clock clock.Clock, eventRecorder record.EventRecord
 }

 func (hr *horizontalController) sync(ctx context.Context, podAutoscaler *datadoghq.DatadogPodAutoscaler, autoscalerInternal *model.PodAutoscalerInternal) (autoscaling.ProcessResult, error) {
+	// If we have no Spec, nothing to do
+	if autoscalerInternal.Spec() == nil {
+		return autoscaling.NoRequeue, nil


It would be a silent failure. Is it something we expect ? Should we return an error ?

It's not really possible in practice or only temporarily. In any case once the Spec is updated, it will requeued.

pkg/clusteragent/autoscaling/workload/controller_horizontal.go

AliDatadog · 2024-07-05T09:47:52Z

pkg/clusteragent/autoscaling/workload/controller_horizontal.go

+) (*datadoghq.DatadogPodAutoscalerHorizontalAction, time.Duration, error) {
+	// Check if we scaling has been disabled enabled explicitly
+	if currentDesiredReplicas == 0 {
+		return nil, 0, errors.New("scaling disabled as current replicas is set to 0")


I am curious about this case, can it happen unintentionally ? Should we retry later ?

We always retry all autoscalers every 5m if nothing happens in the meantime.

AliDatadog · 2024-07-05T10:00:26Z

pkg/clusteragent/autoscaling/workload/controller_horizontal.go

-	allowed, reason := isScalingAllowed(autoscalerInternal, source, scaleDirection)
+	// Checking if scaling constraints allow this scaling
+	autoscalerSpec := autoscalerInternal.Spec()
+	allowed, reason := isScalingAllowed(autoscalerSpec, source, scaleDirection)


Suppose we have a desired value that is less than minReplicas and upscaling is disabled, it means that the desired replicas will stay below minReplicas.

IMO we should always honor the boundaries, even if the policy forbids upscaling or downscaling. As a user, I would definitely expect that. What do you think ?

I think it makes sense to have it this way. If you disable downscaling for an incident and manually scale above maxReplicas, I would not want the controller to downscale it anyway.

AliDatadog · 2024-07-05T10:25:46Z

pkg/clusteragent/autoscaling/workload/controller_horizontal.go

+	earliestTimestamp := currentTime.Add(-periodDuration)
+
+	for _, event := range events {
+		if event.Time.Time.After(earliestTimestamp) {
+			if numEvents == 0 {
+				// Record when the oldest event will be out of the window
+				expireIn = event.Time.Sub(earliestTimestamp)
+			}
+
+			numEvents++
+			diff := event.ToReplicas - event.FromReplicas
+			if diff > 0 {
+				added += diff
+			} else {
+				removed += -diff
+			}
+		}
+	}


Suggested change

earliestTimestamp := currentTime.Add(-periodDuration)

for _, event := range events {

if event.Time.Time.After(earliestTimestamp) {

if numEvents == 0 {

// Record when the oldest event will be out of the window

expireIn = event.Time.Sub(earliestTimestamp)

}

numEvents++

diff := event.ToReplicas - event.FromReplicas

if diff > 0 {

added += diff

} else {

removed += -diff

}

}

}

earliestTimestamp := currentTime.Add(-periodDuration)

events = lo.Filter(events, func(event datadoghq.DatadogPodAutoscalerHorizontalAction, _ int) bool {

return event.Time.Time.After(earliestTimestamp)

})

numEvents = int32(len(events))

for i, event := range events {

if i == 0 {

// Record when the oldest event will be out of the window

expireIn = event.Time.Sub(earliestTimestamp)

}

diff := event.ToReplicas - event.FromReplicas

if diff > 0 {

added += diff

} else {

removed -= diff

}

}

But I loop matching events twice instead of once. I'd rather keep the current simple one-pass loop.

pkg/clusteragent/autoscaling/workload/controller_horizontal.go

AliDatadog · 2024-07-05T11:06:24Z

pkg/clusteragent/autoscaling/workload/controller_horizontal.go

+	if targetDesiredReplicas > maxReplicasFromRules {
+		return maxReplicasFromRules, minExpireIn, fmt.Sprintf("desired replica count limited to %d (originally %d) due to scaling policy", maxReplicasFromRules, targetDesiredReplicas)
+	}
+	return targetDesiredReplicas, 0, ""


I think we should re-check the boundaries here. If we look at the past N events and at event N+1 the boundaries changed, then the Nth value can be outside the boundaries and will be taken into account in maxReplicasFromRules.

targetDesiredReplicas is already bounded to max/minReplicas before calling scaleUp/Down. Rules do not account for min/max, so the result from maxReplicasFromRules is the same regardless of min/max.

AliDatadog · 2024-07-05T11:11:34Z

pkg/clusteragent/autoscaling/workload/controller_horizontal.go

+	if targetDesiredReplicas > maxReplicasFromRules {
+		return maxReplicasFromRules, minExpireIn, fmt.Sprintf("desired replica count limited to %d (originally %d) due to scaling policy", maxReplicasFromRules, targetDesiredReplicas)
+	}
+	return targetDesiredReplicas, 0, ""


Why do we return 0 here ? Shouldn't we reevaluate after minExpireIn in case the next event limits us ?

We don't requeue after a specific delay as we already fully reached target. The next impact will be when we receive a new reccomendation, which will trigger a requeue anyway.

pkg/clusteragent/autoscaling/workload/model/pod_autoscaler_test_utils.go

AliDatadog · 2024-07-05T12:55:14Z

pkg/clusteragent/autoscaling/workload/model/pod_autoscaler_test.go

+func TestAddHorizontalAction(t *testing.T) {
+	testTime := time.Now()
+
+	// Test no retention, should move back to keep a single action
+	var horizontalEventsRetention time.Duration
+	horizontalLastActions := []datadoghq.DatadogPodAutoscalerHorizontalAction{
+		{
+			Time: metav1.Time{Time: testTime.Add(-10 * time.Minute)},
+		},
+		{
+			Time: metav1.Time{Time: testTime.Add(-8 * time.Minute)},
+		},
+	}
+	addedAction1 := &datadoghq.DatadogPodAutoscalerHorizontalAction{
+		Time: metav1.Time{Time: testTime},
+	}
+	horizontalLastActions = addHorizontalAction(testTime, horizontalEventsRetention, horizontalLastActions, addedAction1)
+	assert.Equal(t, []datadoghq.DatadogPodAutoscalerHorizontalAction{*addedAction1}, horizontalLastActions)
+
+	// Add another event, should still keep one
+	horizontalLastActions = addHorizontalAction(testTime, horizontalEventsRetention, horizontalLastActions, addedAction1)
+	assert.Equal(t, []datadoghq.DatadogPodAutoscalerHorizontalAction{*addedAction1}, horizontalLastActions)
+
+	// 15 minutes retention, should keep everything
+	horizontalEventsRetention = 15 * time.Minute
+	horizontalLastActions = []datadoghq.DatadogPodAutoscalerHorizontalAction{
+		{
+			Time: metav1.Time{Time: testTime.Add(-10 * time.Minute)},
+		},
+		{
+			Time: metav1.Time{Time: testTime.Add(-8 * time.Minute)},
+		},
+	}
+	// Adding two fake events
+	horizontalLastActions = addHorizontalAction(testTime, horizontalEventsRetention, horizontalLastActions, addedAction1)
+	horizontalLastActions = addHorizontalAction(testTime, horizontalEventsRetention, horizontalLastActions, addedAction1)
+	assert.Equal(t, []datadoghq.DatadogPodAutoscalerHorizontalAction{
+		{
+			Time: metav1.Time{Time: testTime.Add(-10 * time.Minute)},
+		},
+		{
+			Time: metav1.Time{Time: testTime.Add(-8 * time.Minute)},
+		},
+		*addedAction1,
+		*addedAction1,
+	}, horizontalLastActions)
+
+	// Moving time forward, should keep only the last two events
+	testTime = testTime.Add(10 * time.Minute)
+	addedAction2 := &datadoghq.DatadogPodAutoscalerHorizontalAction{
+		Time: metav1.Time{Time: testTime},
+	}
+	horizontalLastActions = addHorizontalAction(testTime, horizontalEventsRetention, horizontalLastActions, addedAction2)
+	assert.Equal(t, []datadoghq.DatadogPodAutoscalerHorizontalAction{
+		*addedAction1,
+		*addedAction1,
+		*addedAction2,
+	}, horizontalLastActions)
+}


Suggested change

func TestAddHorizontalAction(t *testing.T) {

testTime := time.Now()

// Test no retention, should move back to keep a single action

var horizontalEventsRetention time.Duration

horizontalLastActions := []datadoghq.DatadogPodAutoscalerHorizontalAction{

{

Time: metav1.Time{Time: testTime.Add(-10 * time.Minute)},

},

{

Time: metav1.Time{Time: testTime.Add(-8 * time.Minute)},

},

}

addedAction1 := &datadoghq.DatadogPodAutoscalerHorizontalAction{

Time: metav1.Time{Time: testTime},

}

horizontalLastActions = addHorizontalAction(testTime, horizontalEventsRetention, horizontalLastActions, addedAction1)

assert.Equal(t, []datadoghq.DatadogPodAutoscalerHorizontalAction{*addedAction1}, horizontalLastActions)

// Add another event, should still keep one

horizontalLastActions = addHorizontalAction(testTime, horizontalEventsRetention, horizontalLastActions, addedAction1)

assert.Equal(t, []datadoghq.DatadogPodAutoscalerHorizontalAction{*addedAction1}, horizontalLastActions)

// 15 minutes retention, should keep everything

horizontalEventsRetention = 15 * time.Minute

horizontalLastActions = []datadoghq.DatadogPodAutoscalerHorizontalAction{

{

Time: metav1.Time{Time: testTime.Add(-10 * time.Minute)},

},

{

Time: metav1.Time{Time: testTime.Add(-8 * time.Minute)},

},

}

// Adding two fake events

horizontalLastActions = addHorizontalAction(testTime, horizontalEventsRetention, horizontalLastActions, addedAction1)

horizontalLastActions = addHorizontalAction(testTime, horizontalEventsRetention, horizontalLastActions, addedAction1)

assert.Equal(t, []datadoghq.DatadogPodAutoscalerHorizontalAction{

{

Time: metav1.Time{Time: testTime.Add(-10 * time.Minute)},

},

{

Time: metav1.Time{Time: testTime.Add(-8 * time.Minute)},

},

*addedAction1,

*addedAction1,

}, horizontalLastActions)

// Moving time forward, should keep only the last two events

testTime = testTime.Add(10 * time.Minute)

addedAction2 := &datadoghq.DatadogPodAutoscalerHorizontalAction{

Time: metav1.Time{Time: testTime},

}

horizontalLastActions = addHorizontalAction(testTime, horizontalEventsRetention, horizontalLastActions, addedAction2)

assert.Equal(t, []datadoghq.DatadogPodAutoscalerHorizontalAction{

*addedAction1,

*addedAction1,

*addedAction2,

}, horizontalLastActions)

}

func TestAddHorizontalAction(t *testing.T) {

testTime := time.Now()

actionNow := &datadoghq.DatadogPodAutoscalerHorizontalAction{

Time: metav1.Time{Time: testTime},

}

actionIn10Min := &datadoghq.DatadogPodAutoscalerHorizontalAction{

Time: metav1.Time{Time: testTime.Add(10 * time.Minute)},

}

testCases := []struct {

description string

testTime time.Time

horizontalEventsRetention time.Duration

horizontalLastActions []datadoghq.DatadogPodAutoscalerHorizontalAction

addedAction *datadoghq.DatadogPodAutoscalerHorizontalAction

expectedActions []datadoghq.DatadogPodAutoscalerHorizontalAction

}{

{

description: "No retention, keep single action",

testTime: testTime,

horizontalLastActions: []datadoghq.DatadogPodAutoscalerHorizontalAction{

{Time: metav1.Time{Time: testTime.Add(-10 * time.Minute)}},

{Time: metav1.Time{Time: testTime.Add(-8 * time.Minute)}},

},

addedAction: actionNow,

expectedActions: []datadoghq.DatadogPodAutoscalerHorizontalAction{*actionNow},

},

{

description: "Add another event, still keep one",

testTime: testTime,

horizontalLastActions: []datadoghq.DatadogPodAutoscalerHorizontalAction{

*actionNow,

},

addedAction: actionNow,

expectedActions: []datadoghq.DatadogPodAutoscalerHorizontalAction{*actionNow},

},

{

description: "15 minutes retention, keep everything",

testTime: testTime,

horizontalEventsRetention: 15 * time.Minute,

horizontalLastActions: []datadoghq.DatadogPodAutoscalerHorizontalAction{

{Time: metav1.Time{Time: testTime.Add(-10 * time.Minute)}},

{Time: metav1.Time{Time: testTime.Add(-8 * time.Minute)}},

},

addedAction: actionNow,

expectedActions: []datadoghq.DatadogPodAutoscalerHorizontalAction{

{Time: metav1.Time{Time: testTime.Add(-10 * time.Minute)}},

{Time: metav1.Time{Time: testTime.Add(-8 * time.Minute)}},

*actionNow,

},

},

{

description: "Moving time forward, keep only last two events",

testTime: testTime.Add(10 * time.Minute),

horizontalEventsRetention: 15 * time.Minute,

horizontalLastActions: []datadoghq.DatadogPodAutoscalerHorizontalAction{

*actionNow,

*actionNow,

},

addedAction: actionIn10Min,

expectedActions: []datadoghq.DatadogPodAutoscalerHorizontalAction{

*actionNow,

*actionNow,

*actionIn10Min,

},

},

}

for _, tc := range testCases {

t.Run(tc.description, func(t *testing.T) {

result := addHorizontalAction(tc.testTime, tc.horizontalEventsRetention, tc.horizontalLastActions, tc.addedAction)

assert.Equal(t, tc.expectedActions, result)

})

}

}

I am not a big fan of array-based testing in this case as the scenario is sequential and the sequence order is important.

pkg/clusteragent/autoscaling/workload/model/pod_autoscaler.go

AliDatadog · 2024-07-05T13:15:49Z

pkg/clusteragent/autoscaling/workload/model/pod_autoscaler.go

+		upscaleRetention := getLongestScalingRulesPeriod(policy.Upscale.Rules)
+		if upscaleRetention > longestRetention {
+			longestRetention = upscaleRetention
+		}


Suggested change

upscaleRetention := getLongestScalingRulesPeriod(policy.Upscale.Rules)

if upscaleRetention > longestRetention {

longestRetention = upscaleRetention

}

longestRetention = getLongestScalingRulesPeriod(policy.Upscale.Rules)

I think upscaleRetention > longestRetention will either be true or equal unless we allow negative values

Indeed, but it keeps symmetry in the upscale/downscale code

AliDatadog · 2024-07-05T13:25:01Z

pkg/clusteragent/autoscaling/workload/model/pod_autoscaler.go

+		actions = actions[:0]
+		actions = append(actions, *action)
+		return actions


Suggested change

actions = actions[:0]

actions = append(actions, *action)

return actions

return []datadoghq.DatadogPodAutoscalerHorizontalAction{*action}

I think it's more concise. If the goal was to update actions in place. I am not sure that it works. Example: https://go.dev/play/p/_iaqW7Zwg3k

It's more concise but it allocates, current code does not. It's true that :0 does not remove memory, so any previous reference to the array will read un-existing elements, but that should not happen here.

pkg/clusteragent/autoscaling/workload/model/pod_autoscaler.go

vboulineau · 2024-07-05T15:47:21Z

/merge

dd-devflow · 2024-07-05T15:47:28Z

🚂 MergeQueue: waiting for PR to be ready

This merge request is not mergeable yet, because of pending checks/missing approvals. It will be added to the queue as soon as checks pass and/or get approvals.
Note: if you pushed new commits since the last approval, you may need additional approval.
You can remove it from the waiting list with /remove command.

Use /merge -c to cancel this operation!

dd-devflow · 2024-07-05T17:27:36Z

🚂 MergeQueue: pull request added to the queue

The median merge time in main is 25m.

Use /merge -c to cancel this operation!

dd-devflow · 2024-07-05T17:30:33Z

🚨 MergeQueue: This merge request is in error

Gitlab pipeline didn't start on its own and we were unable to create it... Please retry.

If you need support, contact us on Slack #devflow with those details!

jennchenn · 2024-07-05T17:50:25Z

/merge

jennchenn · 2024-07-05T17:58:51Z

/merge

dd-devflow · 2024-07-05T18:02:43Z

🚂 MergeQueue: pull request added to the queue

The median merge time in main is 25m.

Use /merge -c to cancel this operation!

dd-devflow · 2024-07-05T18:04:54Z

❌ MergeQueue

PR already in the queue with status in_progress

If you need support, contact us on Slack #devflow with those details!

Make autoscaling model private

60b1de9

vboulineau added team/containers kind/feature changelog/no-changelog qa/done Skip QA week as QA was done before merge and regressions are covered by tests labels Jun 27, 2024

vboulineau added this to the 7.56.0 milestone Jun 27, 2024

vboulineau requested review from a team as code owners June 27, 2024 12:58

Add support of scaling rules for horizontal scaling

2c43fcd

vboulineau force-pushed the vboulineau/autoscaling branch from 8e442c1 to 2c43fcd Compare June 27, 2024 16:18

hush-hush approved these changes Jul 1, 2024

View reviewed changes