Current status: everything works, but it's early days and it might change substantially without notice.
- Write your operator in
- type safe,
null-safe, generics, composable
- type safe,
- Small: <1k LOC, only a dozen classes
- Client-agnostic (works with multiple backends)
- Functional, pure API using cats-effect (v2 only for now)
- First-class object deletion (
SoftDeletedso the types ensure correct handling of deleted resources)
- Best practices baked in (e.g.
ResourceMirroris the equivalent of go's
SharedInformer, except that a
SharedInformerisn't supported by
controller-runtime, you have to write a custom
val foperatorVersion = "TODO" // see releases page for latest version
// core modules
libraryDependencies += "net.gfxmonk" %% "foperator-core" % foperatorVersion
libraryDependencies += "net.gfxmonk" %% "foperator-testkit" % foperatorVersion % Test
// pick a backend
libraryDependencies += "net.gfxmonk" %% "foperator-backend-skuber" % foperatorVersion
libraryDependencies += "net.gfxmonk" %% "foperator-backend-kubernetes-client" % foperatorVersion
Hand-built kubernetes client bindings.
- concurrency: akka-streams + Futures, wrapped with Task / Observable in
- http: akka-http
- serialization: play / jackson
The backend class is
Note: skuber is built on akka, which liberally employs bad practices like
This means you can't use skuber within a test scheduler (e.g. TestControl from cats-effect)
Autogenerated bindings from swagger API definitions.
- concurrency: works with any cats-effect compatible IO implementation
- http: http4s
- serialization: circe
The backend class is
To make implicits available,
This is a stub backend for use in tests. It's backed by a simple in-memory hash, but it's enough to test operations on custom resources. It works with any resource type, so you can use your real Skuber / KubernetesClient resources, or for simplicity you can also use the TestResource type.
See ./sample for a couple of sample operators using the framework.
Models.scala provides a good overview of the implicit values you need to make available for each backend to be able to operate on a custom resource.
If you're going to shun the de-facto standard technologies for creating Kubernetes Operators, you should probably have good justification for doing so. I strongly believe that
foperator (and Scala) lets you write operators with less bugs, less effort and less code then anything built in Go. But this is not just a gut feel, I wrote up some extensive notes on the downsides of
operator-sdk after implementing the same operator with both frameworks.
The flow of information looks a bit like this:
The ResourceMirror is responsible for mirroring the state of watched objects, as well as providing an
ids update stream.
The Dispatcher (an internal component) uses the stream of updated IDs to schedule calls to the reconciler. It handles concurrency, error backoff and periodic retries.
The Reconciler is your job: given the current state of an object, it does whatever it is you want to do, and (typically) updates the status of the reconciled object.
To get a better understanding of how to write foperator code, let's take a tour of the core types you'll encounter:
T is the type of the resource being reconciled, like
IO is e.g. Monix Task or cats-effect IO.
Client is the type of the client -
Skuber if you're using
A reconciler is essentially a function of the type
(Client, T) => IO[ReconcileResult]. Like a function it's inert, you need to provide a client in order to run it.
The reconciler is given the current state of the object it's reconciling, and is responsible for making the world right (which is up to you).
However there are some convenience builders to make this even simpler, e.g. if you don't need to return a custom result or interact with the client, you can just supply a function
T => IO[Unit].
The Client trait is a foperator wrapper around an underlying client (e.g. skuber or kubernetes-client). This is the dependency you'll need in order to actually run anything. Individual backends inherit from this type.
Client paramater is essentially "Self", i.e. the Skuber client implements Client[Task, Skuber]). This is a little ugly and I'd get rid of it if I knew how ;)
Each backend's companion object (e.g. Skuber, KubernetesClient, TestClient) inherits some common companion functionality. Most useful is the
Reconcile builder object, which saves you from having to specify the IO and Client types explicitly.
The Operations type is the convenient interface tying together a Client and a specific resource type. It provides convenience functions for anything you want to do with a client - reading, writing, creating a mirror and running a reconciler.
Operations can only be constructed for a combination of
T where there is an implicit
Engine available, which essentially controls what resources a given backend is capable of operating on (see below).
A ResourceMirror is effectively a local cache of the Kubernetes state that we're interested in. You can set up a ResourceMirror for all objects of a given type, or just for a subset (based on labels). It will initially get a list of all such resources, as well as setting up a watch for any future changes. Using this, it maintains an internal state of every watched object, plus a stream of IDs of updated resources.
We don't send full resources into the dispatcher, since a notification about an updated resource might be queued rather than immediately executed. To ensure we don't waste time reconciling old versions of objects, the dispatcher will grab the current version of the relevant resource from the mirror just before it actually runs the reconcile.
Because a ResourceMirror instance corresponds to an ongoing
watch API query as well as a full local copy of every watched resource, we don't want to make lots of them. In general, an operator controller process should only have one ResourceMirror for each resource type, since they can be safely shared.
This is a simple typeclass to provide resource functionality for a given type. Mostly it relates to getting / setting various kubernetes metadata (version, finalizers, API path, etc).
To build an
Operations, you'll need an implicit Engine[IO, Client, T] available. This is the low-level definition of how a given Client can operate on a given resource type (T).
Various implicit engine definitions are available, with certain constraints on the resource type:
Engine[Task, Skuber, T<:skuber.ObjectResource], which means "a Skuber client can act on any subclass of skuber.ObjectResource".
Engine[IO[_], KubernetesClient, T]for individual T types from the
Engine[IO[_], TestClient, T]for any T which has an ObjectResource instance defined.
The use of typeclasses means foperator can trivially support multiple backends, but can make errors harder to understand.
One common issue is when an implicit is not found, e.g. "No implicits found for e: Engine[Task, TestClient[Task], CustomResourceDefinition]"
Most of the time, an implicit will become available when you import e.g.
foperator.backend.skuber.implicits._ (for the Skuber backend). If not, it often means that there's an implicit missing somewhere in the chain.
In these cases, it's useful to dig into the implicit chain. Firstly, figure out the method which ought to provide this implicit. In this case, it's
TestClient.implicitEngine. Then try to invoke that explicitly, by adding:
val e = TestClient.implicitEngine[Task, CustomResourceDefinition]
In this case you get a new error, "No implicits found for eq: Eq[CustomResourceDefinition]"
In this case an
Eq instance was missing, preventing the implicit Engine from being built. In an earlier version (of foperator) the skuber backend provided an
CustomResource, but not
In general if you're operating on builtin kubernetes types, you may need to provide some of your own instances (like
ObjectEditor), whenever there's not a blanket implementation available in
Most "hello world" operators (and even many real-world operators) involve a single Kubernetes resource. When a resource spec changes, you do something to make sure "the world" is in sync, then update the resource's status to match. If you're managing resources outside the Kubernetes cluster, this is really all you need. But what about when you're juggling multiple different types of Kubernetes resources?
Consider the ReplicaSet controller, which typically manages pods for a deployment. The ReplicaSet defines "I want 3 of these pods running", and the controller is responsible for making sure there are 3 pods running. Pods can come and go of their own accord, so only watching the ReplicaSet and relying on periodic reconciliation to notice dead Pods isn't very responsive.
It turns out we can use multiple ResourceMirrors for this: The
Reconciler[ReplicaSet] reconcile function can hold a reference to a secondary
ResourceMirror[Pod]. Since a ResourceMirror contains a local copy of the state of a whole set of resources, it's cheap for the ReplicaSet reconcile function to check every time that the right number of Pods exist in this cache, without having to make API calls to Kubernetes all the time.
That makes reconciliation cheap, but it doesn't make it responsive. For this, my first thought was that "we also need to reconcile Pods" - i.e. when a Pod comes or goes, we should reconcile it to make sure each ReplicaSet's desired state is upheld. But that's a bad idea. In particular, the framework guarantees that only one reconcile will be running for a given ReplicaSet at any time, to prevent conflicts. But if we also do some actions per Pod, we may be reconciling a Pod and its corresponding ReplicaSet concurrently, which is hard to reason about. It also makes the code more complex since we have two reconcile loops to reason about.
Instead, we watch pods without reconciling them, and just use them as an additional trigger for ReplicaSet reconciliation. That is, a ReplicaSet will be reconciled whenever:
- the ReplicaSet itself is updated, OR
- a Pod is updated which "relates" to this ReplicaSet
The relationship is simple enough to implement, and is also facilitated by the ResourceMirror. For each Pod update:
- get all ReplicaSets which should relate to this pod but don't (the selector matches but the ReplicaSet doesn't own the pod)
- get all ReplicaSets which do reference this pod but shouldn't (the ReplicaSet owns the pod but the selector doesn't match)
The ReplicaSet IDs obtained from these relationships are merged with the IDs of updated resources, and either action will cause a reconcile of the ReplicaSet. In this way there is one piece of code responsible for taking action, but we augment it with additional triggers due to changes outside the resource itself. The
ReconcileSource trait is provided by ResourceMirror, and has functions for merging updates from multiple ResourceMirrors in this way.