Vector is a tool that augments your auto-scaling groups. The two features currently offered are Predictive Scaling and Flexible Down Scaling.
Auto Scaling groups do a good job of responding to current load conditions, but if you have a predictable load pattern, it can be nice to scale up your servers a little bit early. Some reasons you might want to do that are:
- If it takes several minutes for an instance to fully boot and ready itself for requests.
- If you have very serious (but predictable) spikes, it's nice to have the capacity in place before the spike starts.
- To give yourself a buffer of time if AWS APIs start throwing errors. If scaling up is going to fail, you'd rather it start failing with a little bit of time before you actually need the capacity so you can begin evasive maneuvers.
Vector examines your existing CloudWatch alarms tied to your Auto Scaling groups, and predicts if they will be triggered in the future based on what happened in the past.
Note: This only works with metrics that are averaged across your group - like CPUUtilization or Load. If you auto-scale based on something like QueueLength, Predictive Scaling will not work right for you.
For each lookback window you specify, Vector will first check the
current value of the metric * the number of nodes, and the past value of
the metric * the past number of nodes. If those numbers are close enough
(within the threshold specified by
--ps-valid-threshold), then it will
Vector will then go back to the lookback window specified, and then
forward in time based on the lookahead window (
It will compute the metric * number of nodes then to get a predicted
aggregate metric value for the current future. It then divides that by
the current number of nodes to get a predicted average value for the
metric. That is then compared against the alarm's threshold.
You have an alarm that checks CPUUtilization of your group, and will trigger the alarm if that goes above 70%. Vector is configured to use a 1 week lookback window, a 1 hour lookahead window, and a valid-threshold of 0.8.
The current value of CPUUtilization is 49%, and there are 2 nodes in the group. CPUUtilization 1 week ago was 53%, and there were 2 nodes in the group. Therefore, total current CPUUtilization is 98%, and 1 week ago was 106%. Those are within 80% of each other (valid-threshold), so we can continue with the prediction.
The value of CPUUtilization 1 week ago, plus 1 hour was 45%, and there were 4 nodes in the group. We calculate total CPUUtilization for that time to be 180%. Assuming no new nodes are launched, the predicted average CPUUtilization for the group 1 hour from now is 180% / 2 = 90%. 90% is above the alarm's 75% threshold, so we trigger the scaleup policy.
If you use Predictive Scaling, you probably also want to use Flexible
Down Scaling (below) so that after scaling up in prediction of load,
your scaledown policy doesn't quickly undo Vector's hard work. You
probably want to set
up-to-down-cooldown to be close to the size of
If you specify a timezone (either explicitly or via the system timezone), Vector will use DST-aware time calculations when evaluating lookback windows. If you don't specify a timezone and your system time is UTC, then 8AM on Monday morning after DST ends will look back 168 hours - which is 7AM on the previous Monday. Predictive scaling would be off by one hour for a whole week in that case.
Flexible Down Scaling
Different Cooldown Periods
Auto Scaling Groups support the concept of "cooldown periods" - a window of time after a scaling activity where no other activities should take place. This is to give the group a chance to settle into the new configuration before deciding whether another action is required.
However, Auto Scaling Groups only support specifying the cooldown period after a certain activity - you can say "After a scale up, wait 5 minutes before doing anything else, and after a scale down, wait 15 minutes." What you can't do is say "After a scale up, wait 5 minutes for another scale up, and 40 minutes for a scale down."
Vector lets you add custom up-to-down and down-to-down cooldown periods. You create your policies and alarms in your Auto Scaling Groups like normal, and then disable the alarms tied to the scale down policy. Then you tell Vector what cooldown periods to use, and he does the rest.
Another benefit to Flexible Down Scaling is the ability to specify multiple alarms for a scaling down policy and require all alarms to trigger before scaling down. With Vector, you can add multiple (disabled) alarms to a policy, and Vector will trigger the policy only when both alarms are in ALARM state. This lets you do something like "only scale down when CPU utilization is < 30% and there is not a backlog of requests on any instances".
Max Sunk Cost
Vector also lets you specify a "max sunk cost" when scaling down a node. Amazon bills on hourly increments, and you pay a full hour for every partial hour used, so you want your instances to terminate as close to their hourly billing renewal (without going past it).
For example, if you specify
--fds-max-sunk-cost 15m and have two nodes
in your group - 47 minutes and 32 minutes away from their hourly billing
renewals - the group will not be scaled down.
(You should make sure to run Vector on an interval smaller than this one, or else it's possible Vector may never find eligible nodes for scaledown and never scaledown.)
When deciding to scale down, a static CPU utilization threshold may be inefficient. For example, if there are 2 nodes running, and you have a minimum of 2, and the average CPU is 75%, removing 1 node would theoretically result in the remaining 2 nodes running at > 100%. However, with 20 nodes running, at an average CPU of 75%, removing 1 node will only result in an average CPU of 79% across the remaining 19 nodes.
When there are more nodes running, you can be more aggressive about removing nodes without overloading the remaining nodes. Variable thresholds allow you to express this.
You can enable variable thresholds with
Integration with Predictive Scaling
Before scaling down, and if Predictive Scaling is in effect, Vector will check to see if the size after scaling down would trigger Predictive Scaling. If it would, the scaling policy will not be executed.
- Auto Scaling groups must have the GroupInServiceInstances metric enabled.
- Auto Scaling groups must have at least one scaling policy with a positive adjustment, and that policy must have at least one CloudWatch alarm with a CPUUtilization metric.
$ gem install vector
Typically vector will be invoked via cron periodically (every 10 minutes is a good choice.)
Usage: vector [options] DURATION can look like 60s, 1m, 5h, 7d, 1w --timezone TIMEZONE Timezone to use for date calculations (like America/Denver) (default: system timezone) --region REGION AWS region to operate in (default: us-east-1) --groups group1,group2 A list of Auto Scaling Groups to evaluate --fleet fleet An AWS ASG Fleet (instead of specifying --groups) -v, --[no-]verbose Run verbosely Predictive Scaling Options --[no-]ps Enable Predictive Scaling --ps-lookback-windows DURATION,DURATION List of lookback windows --ps-lookahead-window DURATION Lookahead window --ps-valid-threshold FLOAT A number from 0.0 - 1.0 specifying how closely previous load must match current load for Predictive Scaling to take effect --ps-valid-period DURATION The period to use when doing the threshold check Flexible Down Scaling Options --[no-]fds Enable Flexible Down Scaling --fds-up-to-down DURATION The cooldown period between up and down scale events --fds-down-to-down DURATION The cooldown period between down and down scale events --fds-max-sunk-cost DURATION Only let a scaledown occur if there is an instance this close to its hourly billing point
Why not just predictively scale based on the past DesiredInstances?
If we don't look at the actual utilization and just look at how many instances we were running in the past, we will end up scaling earlier and earlier, and will never re-adjust and not scale up if load patterns change, and we don't need so much capacity.
What about high availability? What if the box Vector is running on dies?
Luckily Vector is just providing optimizations - the critical component of scaling up based on demand is still provided by the normal Auto Scaling service. If Vector does not run, you just don't get the predictive scaling and down scaling.