Node auto scaling can be enabled by setting docker:auto-scale:enabled to true. It will try to add, remove and rebalance docker nodes used by tsuru.
Node scaling algorithms run in clusters of docker nodes, to specify how clusters will be formed you must tell tsuru how they should be grouped. This is done by setting docker:auto-scale:group-by-metadata configuration entry to the name of a metadata present in your nodes.
There are two different scaling algorithms that will be used, depending on how tsuru is configured: count based scaling, and memory based scaling.
It's chosen if docker:auto-scale:max-container-count is set to a value > 0 in your tsuru configuration.
Having max-container-count value as max, the number of nodes in cluster as nodes, and the total number of containers in all cluster's nodes as total, we get the number of free slots free with:
free = max * nodes − total
If free < 0 then a new node will be added and tsuru will rebalance containers using the new node.
Having docker:auto-scale:scale-down-ratio value ratio. tsuru will try to remove an existing node if:
free > max * ratio
Before removing a node tsuru will move it's containers to other nodes available in the cluster.
To avoid entering loops, removing and adding node, tsuru will require ratio > 1, if this is not true scaling will not run.
It's chosen if docker:auto-scale:max-container-count is not set and your scheduler is configured to use node's memory information, by setting docker:scheduler:total-memory-metadata and docker:scheduler:max-used-memory.
Having the amount of memory necessary by the plan with the largest memory requirement as maxPlanMemory. A new node will be added if for all nodes the amount of unreserved memory (unreserved) satisfies:
unreserved < maxPlanMemory
Considering the amount of memory necessary by the plan with the largest memory requirement as maxPlanMemory and docker:auto-scale:scale-down-ratio value as ratio. A node will be removed if its current containers can be distributed across other nodes in the same pool and at least one node still has unreserved memory (unreserved) satisfying:
unreserved > maxPlanMemory * ratio
Rebalancing containers will be triggered when a new node is added or if rebalancing would decrease the difference of containers in nodes by a number greater than 2, regardless the scaling algorithm.
Also, rebalancing will not run if docker:auto-scale:prevent-rebalance is set to true.
Each time tsuru tries to run an auto scale action (add, remove, or rebalance). It will create an auto scale event. This event will record the result of the auto scale action and possible errors that occurred during its execution.
You can list auto scale events with tsuru-admin docker-autoscale-list
Even if you have docker:auto-scale:enabled set to false, you can make tsuru trigger the execution of the auto scale algorithm by running tsuru-admin docker- autoscale-run.