Services or capabilities described in Amazon Web Services documentation might vary by Region. To see the differences applicable to the China Regions, see Getting Started with Amazon Web Services in China (PDF).

(Optional) Set up Fluentd as a DaemonSet to send logs to CloudWatch Logs

To set up Fluentd to collect logs from your containers, you can follow the steps in Quick Start setup for Container Insights on Amazon EKS and Kubernetes or you can follow the steps in this section. In the following steps, you set up Fluentd as a DaemonSet to send logs to CloudWatch Logs. When you complete this step, Fluentd creates the following log groups if they don't already exist.

Step 1: Create a namespace for CloudWatch

Use the following step to create a Kubernetes namespace called amazon-cloudwatch for CloudWatch. You can skip this step if you have already created this namespace.

Step 2: Install Fluentd

Start this process by downloading Fluentd. When you finish these steps, the deployment creates the following resources on the cluster:

Step 3: Verify the Fluentd setup

To verify your Fluentd setup, use the following steps.

Multi-line log support

On August 19 2019, we added multi-line log support for the logs collected by Fluentd.

By default, the multi-line log entry starter is any character with no white space. This means that all log lines that start with a character that does not have white space are considered as a new multi-line log entry.

If your own application logs use a different multi-line starter, you can support them by making two changes in the fluentd.yaml file.

First, exclude them from the default multi-line support by adding the pathnames of your log files to an exclude_path field in the containers section of fluentd.yaml. The following is an example.

<source>
  @type tail
  @id in_tail_container_logs
  @label @containers
  path /var/log/containers/*.log
  exclude_path ["full_pathname_of_log_file*", "full_pathname_of_log_file2*"]

Next, add a block for your log files to the fluentd.yaml file. The example below is used for the CloudWatch agent's log file, which uses a timestamp regular expression as the multi-line starter. You can copy this block and add it to fluentd.yaml. Change the indicated lines to reflect your application log file name and the multi-line starter that you want to use.

<source>
  @type tail
  @id in_tail_cwagent_logs
  @label @cwagentlogs
  path /var/log/containers/cloudwatch-agent*
  pos_file /var/log/cloudwatch-agent.log.pos
  tag *
  read_from_head true
 <parse>
    @type json
    time_format %Y-%m-%dT%H:%M:%S.%NZ
 </parse>
</source> 

<label @cwagentlogs>
  <filter **>
    @type kubernetes_metadata
    @id filter_kube_metadata_cwagent
  </filter>

  <filter **>
    @type record_transformer
    @id filter_cwagent_stream_transformer
    <record>
      stream_name ${tag_parts[3]}
    </record>
 </filter>

  <filter **>
    @type concat
    key log
    multiline_start_regexp /^\d{4}[-/]\d{1,2}[-/]\d{1,2}/
    separator ""
    flush_interval 5
    timeout_label @NORMAL
 </filter>

 <match **>
    @type relabel
    @label @NORMAL
 </match>
</label>

(Optional) Reducing the log volume from Fluentd

By default, we send Fluentd application logs and Kubernetes metadata to CloudWatch. If you want to reduce the volume of data being sent to CloudWatch, you can stop one or both of these data sources from being sent to CloudWatch.

To stop Fluentd application logs, remove the following section from the fluentd.yaml file.

<source>
  @type tail
  @id in_tail_fluentd_logs
  @label @fluentdlogs
  path /var/log/containers/fluentd*
  pos_file /var/log/fluentd.log.pos
  tag *
  read_from_head true
   <parse>
   @type json
   time_format %Y-%m-%dT%H:%M:%S.%NZ
  </parse>
</source>  

<label @fluentdlogs>
  <filter **>
    @type kubernetes_metadata
    @id filter_kube_metadata_fluentd
  </filter>

  <filter **>
    @type record_transformer
    @id filter_fluentd_stream_transformer
    <record>
      stream_name ${tag_parts[3]}
    </record>
   </filter>

   <match **>
     @type relabel
     @label @NORMAL
   </match>
</label>

To remove Kubernetes metadata from being appended to log events that are sent to CloudWatch, add one line to the record_transformer section in the fluentd.yaml file. In the log source where you want to remove this metadata, add the following line.

remove_keys $.kubernetes.pod_id, $.kubernetes.master_url, $.kubernetes.container_image_id, $.kubernetes.namespace_id

For example:

<filter **>
  @type record_transformer
  @id filter_containers_stream_transformer
  <record>
    stream_name ${tag_parts[3]}
  </record>
  remove_keys $.kubernetes.pod_id, $.kubernetes.master_url, $.kubernetes.container_image_id, $.kubernetes.namespace_id
</filter>

Troubleshooting

If you don't see these log groups and are looking in the correct Region, check the logs for the Fluentd DaemonSet pods to look for the error.

Run the following command and make sure that the status is Running.

kubectl get pods -n amazon-cloudwatch

In the results of the previous command, note the pod name that starts with fluentd-cloudwatch. Use this pod name in the following command.

kubectl logs pod_name -n amazon-cloudwatch

If the logs have errors related to IAM permissions, check the IAM role attached to the cluster nodes. For more information about the permissions required to run an Amazon EKS cluster, see Amazon EKS IAM Policies, Roles, and Permissions in the Amazon EKS User Guide.

If the pod status is CreateContainerConfigError, get the exact error by running the following command.

kubectl describe pod pod_name -n amazon-cloudwatch

If the pod status is CrashLoopBackOff, make sure that the architecture of the Fluentd container image is the same as the node when you installed Fluentd. If your cluster has both x86 and ARM64 nodes, you can use a kubernetes.io/arch label to place the images on the correct node. For more information, see kuberntes.io/arch.