LimitRange support in Pipeline

LimitRanges, Requests and Limits

Taken from the LimitRange in kubernetes docs.

By default, containers run with unbounded compute resources on a Kubernetes cluster. With resource quotas, cluster administrators can restrict resource consumption and creation on a namespace basis. Within a namespace, a Pod or Container can consume as much CPU and memory as defined by the namespace’s resource quota. There is a concern that one Pod or Container could monopolize all available resources. A LimitRange is a policy to constrain resource allocations (to Pods or Containers) in a namespace.

A LimitRange provides constraints that can:

  • Enforce minimum and maximum compute resources usage per Pod or Container in a namespace.
  • Enforce minimum and maximum storage request per PersistentVolumeClaim in a namespace.
  • Enforce a ratio between request and limit for a resource in a namespace.
  • Set default request/limit for compute resources in a namespace and automatically inject them to Containers at runtime.

LimitRange are validating and mutating Requests and Limits. Let’s look, in a nutshell, on how those work in Kubernetes.

  • Requests are not enforced. If the node has more resource available than the request, the container can use it.
  • Limits on the other hand, are a hard stop. A container going over the limit, will be killed.

Resource types for both are:

  • CPU
  • Memory
  • Ephemeral storage

The next question is : how pods with resource and limits are run/scheduled ? The scheduler computes the amount of CPU and memory requests (using Requests) and tries to find a node to schedule it.

Requests and limits can be applied to both Containers and Init Containers.

  • For init containers, the max of each type is taken
  • For containers, it sums all requests/limits for each containers

This means, if you got the following:

  • initContainer1 : 1 CPU, 100m memory
  • initContainer2 : 2 CPU, 200m memory
  • container1 : 1 CPU, 50m memory
  • container2 : 2 CPU, 250m memory
  • container3 : 3 CPU, 500m memory

The computation will be:

  • CPU : 2 (max init containers) + 6 (sum of containers) = 8 CPU
  • Memory: 200m (max init containers) + 800m (sum of containers) = 1000m (1G)

Tekton support

The way Limits and Requests works in Kubernetes is because it is assumed that all containers run in parallel, and init container run before, each one after the others.

That assumption — containers running in parallel — is not true in Tekton. They do all start together (because there is no way around this) but the entrypoint hack is making sure they actually run in sequence and thus there is always only one container that is actually consuming some resource at the same time.

This means, we need to handle limits, request and LimitRanges in a non-standard way. Since the existing LimitRange mutating webhook won’t take Tekton’s requirements into account, Tekton needs to fully control all the values the LimitRange webhook might set. Let’s try to define that. Tekton needs to take into account all the aspect of the LimitRange : the min/max as well as the default. If there is no default, but there is min/max, Tekton need then to set a default value that is between the min/max. If we set the value too low, the Pod won’t be able to be created, similar if we set the value too high. But those values are set on containers, so we have to do our own computation to know what request to put on each containers.

A LimitRange is in the namespace

We need to get the default (limits), default requests, min and max values (if they are here).

Here are the rules for container’s resources (requests and limits) computations:

  • init containers: they won’t be summed, so the rules are simple
    • a container needs to have request and limits at least at the min and set to the default if any.
    • use the default requests and the default limits (coming from the defaultLimit, or the min, …)
  • containers: those will be summed at the end, so it gets a bit complex
    • a container needs to have request and limits at least at the min
    • the sum of the container request/limits should be as small as possible. This should be ensured by using the “smallest” possible request on it.

One thing to note is that, in the case of a LimitRange being present, we need to not rely on the pod mutation webhook that takes the default into account ; what this means is, we need to specify all request and limits ourselves so that the mutation webhook doesn’t have any work to do.

  • No default value: if there is no default value, we need to treat the min as the default. I think that’s also what k8s does, at least in our computation.
  • Default value: we need to “try” to respect that as much as possible.
    • defaultLimit but no defaultRequest, then we set defaultRequest to be same as min (if present).
    • defaultRequest but no defaultlimit, then we use the max limit as the defaultLimit
    • no defaultLimit, no defaultRequest, then we use the min as defaultRequest and the max as defaultLimit.

Multiple LimitRange are in the namespace

Similar to on LimitRange, except we need to act as if it was one LimitRange (virtual) with the correct value from each of them.

  • Take the maximum of the min values
  • Take the minimum of the max values
  • Take the default request that fits into the previous 2 min/max

Once we have this “virtual” LimitRange, we can act as there was one LimitRange. Note that it is possible to define multiple LimitRange that would go conflict with each other and block any Pod scheduling. Tekton Pipeline will not do anything to try to go around this as it is a behaviour of Kubernetes itself.