feature: only allow autoscaling in a direction once a full metric window has elapsed #125
Conversation
|
Untested, intended for discussion before merging. This is meant to keep from autoscaling up or down repeatedly in a period, before a full metric window has elapsed. Currently we will autoscale again after the grace period TTL has elapsed. |
aaronkvanmeerten
requested review from
bgrozev,
oanaianc,
raluca8x8 and
sawall
| if (desiredCount > group.scalingOptions.maxDesired) { | ||
| desiredCount = group.scalingOptions.maxDesired; | ||
| } | ||
| if (await this.instanceGroupManager.allowAutoscalingByDirection(group.name, 'up')) { |
There was a problem hiding this comment.
We want to be able to react fast especially to scaling up and this might alter this behaviour.
-
For example, let's say scaleUpPeriodsCount = 3, and scaleUpThreshold = 5. At time x, we might see the following metrics per period: 4, 3, 3. Which means we scale up.
-
We wait for gracePeriodTTLSec for the instances to come up and be operational and then we check again the new metrics window. But if the usage is high, the newly instances will be immediately used and the sliding window will become 3, 3, 2. Which means we should scale up again.
-
You change suggests we should not scale up then, but instead wait 2 more scalePeriod and only then take the decision. This might be too late.
If we do want to play with having the grace period larger, I would prefer to leave the choice of the grace period in the hands of the admin of the group. Maybe we should look into adding a configurable grace period per direction per group, and the admin can set it to the desired value, such as scaleUpPeriodsCount * group.scalingOptions.scalePeriod for up.
No description provided.