• crassirostris
  • piosz title: Logging Using Stackdriver

Before reading this page, it's highly recommended to familiarize yourself with the overview of logging in Kubernetes.

Note: By default, Stackdriver logging collects only your container's standard output and standard error streams. To collect any logs your application writes to a file (for example), see the sidecar approach in the Kubernetes logging overview.


To ingest logs, you must deploy the Stackdriver Logging agent to each node in your cluster. The agent is a configured fluentd instance, where the configuration is stored in a ConfigMap and the instances are managed using a Kubernetes DaemonSet. The actual deployment of the ConfigMap and DaemonSet for your cluster depends on your individual cluster setup.

Deploying to a new cluster

Google Container Engine

Stackdriver is the default logging solution for clusters deployed on Google Container Engine. Stackdriver Logging is deployed to a new cluster by default unless you explicitly opt-out.

Other platforms

To deploy Stackdriver Logging on a new cluster that you're creating using kube-up.sh, do the following:

  1. Set the KUBE_LOGGING_DESTINATION environment variable to gcp.
  2. If not running on GCE, include the beta.kubernetes.io/fluentd-ds-ready=true in the KUBE_NODE_LABELS variable.

Once your cluster has started, each node should be running the Stackdriver Logging agent. The DaemonSet and ConfigMap are configured as addons. If you're not using kube-up.sh, consider starting a cluster without a pre-configured logging solution and then deploying Stackdriver Logging agents to the running cluster.

Deploying to an existing cluster

  1. Apply a label on each node, if not already present.

    The Stackdriver Logging agent deployment uses node labels to determine to which nodes it should be allocated. These labels were introduced to distinguish nodes with the Kubernetes version 1.6 or higher. If the cluster was created with Stackdriver Logging configured and node has version 1.5.X or lower, it will have fluentd as static pod. Node cannot have more than one instance of fluentd, therefore only apply labels to the nodes that don't have fluentd pod allocated already. You can ensure that your node is labelled properly by running kubectl describe as follows:

    shell kubectl describe node $NODE_NAME

    The output should be similar to this:

    Name: NODE_NAME Role: Labels: beta.kubernetes.io/fluentd-ds-ready=true ...

    Ensure that the output contains the label beta.kubernetes.io/fluentd-ds-ready=true. If it is not present, you can add it using the kubectl label command as follows:

    shell kubectl label node $NODE_NAME beta.kubernetes.io/fluentd-ds-ready=true

    Note: If a node fails and has to be recreated, you must re-apply the label to the recreated node. To make this easier, you can use Kubelet's command-line parameter for applying node labels in your node startup script.

  2. Deploy a ConfigMap with the logging agent configuration by running the following command:

    shell kubectl create -f https://k8s.io/docs/tasks/debug-application-cluster/fluentd-gcp-configmap.yaml

    The command creates the ConfigMap in the default namespace. You can download the file manually and change it before creating the ConfigMap object.

  3. Deploy the logging agent DaemonSet by running the following command:

    shell kubectl create -f https://k8s.io/docs/tasks/debug-application-cluster/fluentd-gcp-ds.yaml

    You can download and edit this file before using it as well.

Verifying your Logging Agent Deployment

After Stackdriver DaemonSet is deployed, you can discover logging agent deployment status by running the following command:

kubectl get ds --all-namespaces

If you have 3 nodes in the cluster, the output should looks similar to this:

NAMESPACE     NAME               DESIRED   CURRENT   READY     NODE-SELECTOR                              AGE
kube-system   fluentd-gcp-v2.0   3         3         3         beta.kubernetes.io/fluentd-ds-ready=true   6d

To understand how logging with Stackdriver works, consider the following synthetic log generator pod specification counter-pod.yaml:

{% include code.html language="yaml" file="counter-pod.yaml" ghlink="/docs/tasks/debug-application-cluster/counter-pod.yaml" %}

This pod specification has one container that runs a bash script that writes out the value of a counter and the date once per second, and runs indefinitely. Let's create this pod in the default namespace.

kubectl create -f http://k8s.io/docs/user-guide/logging/examples/counter-pod.yaml

You can observe the running pod:

$ kubectl get pods
NAME                                           READY     STATUS    RESTARTS   AGE
counter                                        1/1       Running   0          5m

For a short period of time you can observe the 'Pending' pod status, because the kubelet has to download the container image first. When the pod status changes to Running you can use the kubectl logs command to view the output of this counter pod.

$ kubectl logs counter
0: Mon Jan  1 00:00:00 UTC 2001
1: Mon Jan  1 00:00:01 UTC 2001
2: Mon Jan  1 00:00:02 UTC 2001

As described in the logging overview, this command fetches log entries from the container log file. If the container is killed and then restarted by Kubernetes, you can still access logs from the previous container. However, if the pod is evicted from the node, log files are lost. Let's demonstrate this by deleting the currently running counter container:

$ kubectl delete pod counter
pod "counter" deleted

and then recreating it:

$ kubectl create -f http://k8s.io/docs/user-guide/logging/examples/counter-pod.yaml
pod "counter" created

After some time, you can access logs from the counter pod again:

$ kubectl logs counter
0: Mon Jan  1 00:01:00 UTC 2001
1: Mon Jan  1 00:01:01 UTC 2001
2: Mon Jan  1 00:01:02 UTC 2001

As expected, only recent log lines are present. However, for a real-world application you will likely want to be able to access logs from all containers, especially for the debug purposes. This is exactly when the previously enabled Stackdriver Logging can help.

Viewing logs

Stackdriver Logging agent attaches metadata to each log entry, for you to use later in queries to select only the messages you're interested in: for example, the messages from a particular pod.

The most important pieces of metadata are the resource type and log name. The resource type of a container log is container, which is named GKE Containers in the UI (even if the Kubernetes cluster is not on GKE). The log name is the name of the container, so that if you have a pod with two containers, named container_1 and container_2 in the spec, their logs will have log names container_1 and container_2 respectively.

System components have resource type compute, which is named GCE VM Instance in the interface. Log names for system components are fixed. For a GKE node, every log entry from a system component has one the following log names:

  • docker
  • kubelet
  • kube-proxy

You can learn more about viewing logs on the dedicated Stackdriver page.

One of the possible ways to view logs is using the gcloud logging command line interface from the Google Cloud SDK. It uses Stackdriver Logging filtering syntax to query specific logs. For example, you can run the following command:

$ gcloud beta logging read 'logName="projects/$YOUR_PROJECT_ID/logs/count"' --format json | jq '.[].textPayload'
"2: Mon Jan  1 00:01:02 UTC 2001\n"
"1: Mon Jan  1 00:01:01 UTC 2001\n"
"0: Mon Jan  1 00:01:00 UTC 2001\n"
"2: Mon Jan  1 00:00:02 UTC 2001\n"
"1: Mon Jan  1 00:00:01 UTC 2001\n"
"0: Mon Jan  1 00:00:00 UTC 2001\n"

As you can see, it outputs messages for the count container from both the first and second runs, despite the fact that the kubelet already deleted the logs for the first container.

Exporting logs

You can export logs to Google Cloud Storage or to BigQuery to run further analysis. Stackdriver Logging offers the concept of sinks, where you can specify the destination of log entries. More information is available on the Stackdriver Exporting Logs page.