controller-runtime uses a kind of logging called structured logging. If you've used a library like Zap or logrus before, you'll be familiar with the concepts we use. If you've only used a logging library like the "log" package (in the Go standard library) or "glog" (in Kubernetes), you'll need to adjust how you think about logging a bit.
With structured logging, we associate a constant log message with some variable key-value pairs. For instance, suppose we wanted to log that we were starting reconciliation on a pod. In the Go standard library logger, we might write:
log.Printf("starting reconciliation for pod %s/%s", podNamespace, podName)
In controller-runtime, we'd instead write:
logger.Info("starting reconciliation", "pod", req.NamespacedNamed)
or even write
func (r *Reconciler) Reconcile(req reconcile.Request) (reconcile.Response, error) {
logger := logger.WithValues("pod", req.NamespacedName)
// do some stuff
logger.Info("starting reconciliation")
}
Notice how we've broken out the information that we want to convey into
a constant message ("starting reconciliation"
) and some key-value pairs
that convey variable information ("pod", req.NamespacedName
). We've
there-by added "structure" to our logs, which makes them easier to save
and search later, as well as correlate with metrics and events.
All of controller-runtime's logging is done via logr, a generic interface for structured logging. You can use whichever logging library you want to implement the actual mechanics of the logging. controller-runtime provides some helpers to make it easy to use Zap as the implementation.
You can configure the logging implementation using
"sigs.k8s.io/controller-runtime/pkg/log".SetLogger
. That
package also contains the convenience functions for setting up Zap.
You can get a handle to the the "root" logger using
"sigs.k8s.io/controller-runtime/pkg/log".Log
, and can then call
WithName
to create individual named loggers. You can call WithName
repeatedly to chain names together:
logger := log.Log.WithName("controller").WithName("replicaset")
// in reconcile...
logger = logger.WithValues("replicaset", req.NamespacedName)
// later on in reconcile...
logger.Info("doing things with pods", "pod", newPod)
As seen above, you can also call WithValue
to create a new sub-logger
that always attaches some key-value pairs to a logger.
Finally, you can use V(1)
to mark a particular log line as "debug" logs:
logger.V(1).Info("this is particularly verbose!", "state of the world",
allKubernetesObjectsEverywhere)
While it's possible to use higher log levels, it's recommended that you
stick with V(1)
or V(0)(which is equivalent to not specifying
V), and then filter later based on key-value pairs or messages; different numbers tend to lose meaning easily over time, and you'll be left wondering why particular logs lines are at
V(5)instead of
V(7)`.
Errors should always be logged with log.Error
, which allows logr
implementations to provide special handling of errors (for instance,
providing stack traces in debug mode).
It's acceptable to log call log.Error
with a nil error object. This
conveys that an error occurred in some capacity, but that no actual
error
object was involved.
Errors returned by the Reconcile
implementation of the Reconciler
interface are commonly logged as a Reconciler error
.
It's a developer choice to create an additional error log in the Reconcile
implementation so a more specific file name and line for the error are returned.
-
Don't put variable content in your messages -- use key-value pairs for that. Never use
fmt.Sprintf
in your message. -
Try to match the terminology in your messages with your key-value pairs -- for instance, if you have a key-value pairs
api version
, use the termAPIVersion
instead ofGroupVersion
in your message.
Kubernetes objects should be logged directly, like log.Info("this is a Kubernetes object", "pod", somePod)
. controller-runtime provides
a special encoder for Zap that will transform Kubernetes objects into
name, namespace, apiVersion, kind
objects, when available and not in
development mode. Other logr implementations should implement similar
logic.
-
Use lower-case, space separated keys. For example
object
for objects,api version
forAPIVersion
-
Be consistent across your application, and with controller-runtime when possible.
-
Try to be brief but descriptive.
-
Match terminology in keys with terminology in the message.
-
Be careful logging non-Kubernetes objects verbatim if they're very large.
-
Kinds should not be logged alone (they're meaningless alone). Use a
GroupKind
object to log them instead, or aGroupVersionKind
when version is relevant. -
If you need to log an API version string, use
api version
as the key (formatted as with aGroupVersion
, or as received directly from API discovery).
-
If code works with a generic Kubernetes
runtime.Object
, use theobject
key. For specific objects, prefer the resource name as the key (e.g.pod
forv1.Pod
objects). -
For non-Kubernetes objects, the
object
key may also be used, if you accept a generic interface. -
When logging a raw type, log it using the
type
key, with a value offmt.Sprintf("%T", typ)
-
If there's specific context around a type, the key may be more specific, but should end with
type
-- for instance,OwnerType
should be logged asowner
in the context oflog.Error(err, "Could not get ObjectKinds for OwnerType",
owner type, fmt.Sprintf("%T"))
. When possible, favor communicating kind instead.
- When logging multiple things, simply pluralize the key.
-
Reconcile requests should be logged as
request
, although normal code should favor logging the key. -
Reconcile keys should be logged as with the same key as if you were logging the object directly (e.g.
log.Info("reconciling pod", "pod", req.NamespacedName)
). This ends up having a similar effect to logging the object directly.