Overview

The Elasticsearch, Fluentd, and Kibana (EFK) stack aggregates logs from nodes and applications running inside your OpenShift Container Platform installation. Once deployed it uses Fluentd to aggregate event logs from all nodes, projects, and pods into Elasticsearch (ES). It also provides a centralized Kibana web UI where users and administrators can create rich visualizations and dashboards with the aggregated data.

Fluentd bulk uploads logs to an index, in JSON format, then Elasticsearch routes your search requests to the appropriate shards.

Installation

The general procedure for installing an aggregate logging stack in OpenShift Container Platform is described in Aggregating Container Logs. There are some important things to keep in mind while going through the installation guide:

In order for the logging pods to spread evenly across your cluster, an empty node selector should be used.

$ oc adm new-project logging --node-selector=""

In conjunction with node labeling, which is done later, this controls pod placement across the logging project. You can now create the logging project.

$ oc project logging

Elasticsearch (ES) should be deployed with a cluster size of at least three for resiliency to node failures. This is specified by setting the openshift_logging_es_cluster_size parameter in the inventory host file.

Refer to Ansible Variables for a full list of parameters.

If you do not have an existing Kibana installation, you can use kibana.example.com as a value to openshift_logging_kibana_hostname.

Installation can take some time depending on whether the images were already retrieved from the registry or not, and on the size of your cluster.

Inside the logging namespace, you can check your deployment with oc get all.

$ oc get all

NAME                          REVISION                 REPLICAS      TRIGGERED BY
logging-curator               1                        1
logging-es-6cvk237t           1                        1
logging-es-e5x4t4ai           1                        1
logging-es-xmwvnorv           1                        1
logging-kibana                1                        1
NAME                          DESIRED                  CURRENT       AGE
logging-curator-1             1                        1             3d
logging-es-6cvk237t-1         1                        1             3d
logging-es-e5x4t4ai-1         1                        1             3d
logging-es-xmwvnorv-1         1                        1             3d
logging-kibana-1              1                        1             3d
NAME                          HOST/PORT                PATH          SERVICE              TERMINATION   LABELS
logging-kibana                kibana.example.com                     logging-kibana       reencrypt     component=support,logging-infra=support,provider=openshift
logging-kibana-ops            kibana-ops.example.com                 logging-kibana-ops   reencrypt     component=support,logging-infra=support,provider=openshift
NAME                          CLUSTER-IP               EXTERNAL-IP   PORT(S)              AGE
logging-es                    172.24.155.177           <none>        9200/TCP             3d
logging-es-cluster            None                     <none>        9300/TCP             3d
logging-es-ops                172.27.197.57            <none>        9200/TCP             3d
logging-es-ops-cluster        None                     <none>        9300/TCP             3d
logging-kibana                172.27.224.55            <none>        443/TCP              3d
logging-kibana-ops            172.25.117.77            <none>        443/TCP              3d
NAME                          READY                    STATUS        RESTARTS             AGE
logging-curator-1-6s7wy       1/1                      Running       0                    3d
logging-deployer-un6ut        0/1                      Completed     0                    3d
logging-es-6cvk237t-1-cnpw3   1/1                      Running       0                    3d
logging-es-e5x4t4ai-1-v933h   1/1                      Running       0                    3d
logging-es-xmwvnorv-1-adr5x   1/1                      Running       0                    3d
logging-fluentd-156xn         1/1                      Running       0                    3d
logging-fluentd-40biz         1/1                      Running       0                    3d
logging-fluentd-8k847         1/1                      Running       0                    3d

You should end up with a similar setup to the following.

$ oc get pods -o wide

NAME                          READY     STATUS      RESTARTS   AGE       NODE
logging-curator-1-6s7wy       1/1       Running     0          3d        ip-172-31-24-239.us-west-2.compute.internal
logging-deployer-un6ut        0/1       Completed   0          3d        ip-172-31-6-152.us-west-2.compute.internal
logging-es-6cvk237t-1-cnpw3   1/1       Running     0          3d        ip-172-31-24-238.us-west-2.compute.internal
logging-es-e5x4t4ai-1-v933h   1/1       Running     0          3d        ip-172-31-24-235.us-west-2.compute.internal
logging-es-xmwvnorv-1-adr5x   1/1       Running     0          3d        ip-172-31-24-233.us-west-2.compute.internal
logging-fluentd-156xn         1/1       Running     0          3d        ip-172-31-24-241.us-west-2.compute.internal
logging-fluentd-40biz         1/1       Running     0          3d        ip-172-31-24-236.us-west-2.compute.internal
logging-fluentd-8k847         1/1       Running     0          3d        ip-172-31-24-237.us-west-2.compute.internal
logging-fluentd-9a3qx         1/1       Running     0          3d        ip-172-31-24-231.us-west-2.compute.internal
logging-fluentd-abvgj         1/1       Running     0          3d        ip-172-31-24-228.us-west-2.compute.internal
logging-fluentd-bh74n         1/1       Running     0          3d        ip-172-31-24-238.us-west-2.compute.internal
...
...

By default the amount of RAM allocated to each ES instance is 8GB. openshift_logging_es_memory_limit is the parameter used in the openshift-ansible host inventory file. Keep in mind that half of this value will be passed to the individual elasticsearch pods java processes heap size.

Large Clusters

At 100 nodes or more, it is recommended to first pre-pull the logging images from docker pull registry.access.redhat.com/openshift3/logging-fluentd:v3.7. After deploying the logging infrastructure pods (Elasticsearch, Kibana, and Curator), node labeling should be done in steps of 20 nodes at a time. For example:

Using a simple loop:

$ while read node; do oc label nodes $node logging-infra-fluentd=true; done < 20_fluentd.lst

The following also works:

$ oc label nodes 10.10.0.{100..119} logging-infra-fluentd=true

Labeling nodes in groups paces the DaemonSets used by OpenShift logging, helping to avoid contention on shared resources such as the image registry.

Check for the occurence of any "CrashLoopBackOff | ImagePullFailed | Error" issues. oc logs <pod>, oc describe pod <pod> and oc get event are helpful diagnostic commands.

Systemd-journald and rsyslog

Rate-limiting

In Red Hat Enterprise Linux (RHEL) 7 the systemd-journald.socket unit creates /dev/log during the boot process, and then passes input to systemd-journald.service. Every syslog() call goes to the journal.

Rsyslog uses the imjournal module as a default input mode for journal files. Refer to Interaction of rsyslog and journal for detailed information about this topic.

A simple test harness was developed, which uses logger across the cluster nodes to make entries of different sizes at different rates in the system log. During testing simulations under a default Red Hat Enterprise Linux (RHEL) 7 installation with systemd-219-19.el7.x86_64 at certain logging rates (approximately 40 log lines per second), we encountered the default rate limit of rsyslogd. After adjusting these limits, entries stopped being written to journald due to local journal file corruption. This issue is resolved in later versions of systemd.

Scaling up

As you scale up your project, the default logging environment might need some adjustments. After updating to systemd-219-22.el7.x86_64, we added:

$IMUXSockRateLimitInterval 0
$IMJournalRatelimitInterval 0

to /etc/rsyslog.conf and:

# Disable rate limiting
RateLimitInterval=1s
RateLimitBurst=10000
Storage=volatile
Compress=no
MaxRetentionSec=30s

to /etc/systemd/journald.conf.

Now, restart the services.

$ systemctl restart systemd-journald.service
$ systemctl restart rsyslog.service

These settings account for the bursty nature of uploading in bulk.

After removing the rate limit, you may see increased CPU utilization on the system logging daemons as it processes any messages that would have previously been throttled.

Rsyslog is configured (see ratelimit.interval, ratelimit.burst) to rate-limit entries read from the journal at 10,000 messages in 300 seconds. A good rule of thumb is to ensure that the rsyslog rate-limits account for the systemd-journald rate-limits.

Scaling up EFK Logging

If you do not indicate the desired scale at first deployment, the least disruptive way of adjusting your cluster is by re-running the Ansible logging playbook after updating the inventory file with an updated openshift_logging_es_cluster_size value. parameter. Refer to the Performing Administrative Elasticsearch Operations section for more in-depth information.

Storage Considerations

An Elasticsearch index is a collection of shards and its corresponding replica shards. This is how ES implements high availability internally, therefore there is little need to use hardware based mirroring RAID variants. RAID 0 can still be used to increase overall disk performance.

Every search request needs to hit a copy of every shard in the index. Each ES instance requires its own individual storage, but an OpenShift Container Platform deployment can only provide volumes shared by all of its pods, which again means that Elasticsearch shouldn’t be implemented with a single node.

A persistent volume should be added to each Elasticsearch deployment configuration so that we have one volume per replica shard. On OpenShift Container Platform this is often achieved through Persistent Volume Claims

  • 1 volume per shard

  • 1 volume per replica shard

The PVCs must be named based on the openshift_logging_es_pvc_prefix setting. Refer to Persistent Elasticsearch Storage for more details.

Below are capacity planning guidelines for OpenShift Container Platform aggregate logging. Example scenario

Assumptions:

  1. Which application: Apache

  2. Bytes per line: 256

  3. Lines per second load on application: 1

  4. Raw text data → JSON

Baseline (256 characters per minute → 15KB/min)

Logging Infra Pods Storage Throughput

3 es 1 kibana 1 curator 1 fluentd

6 pods total: 90000 x 86400 = 7,7 GB/day

3 es 1 kibana 1 curator 11 fluentd

16 pods total: 225000 x 86400 = 24,0 GB/day

3 es 1 kibana 1 curator 20 fluentd

25 pods total: 225000 x 86400 = 32,4 GB/day

Calculating total logging throughput and disk space required for your logging environment requires knowledge of your application. For example, if one of your applications on average logs 10 lines-per-second, each 256 bytes-per-line, calculate per-application throughput and disk space as follows:

 (bytes-per-line * (lines-per-second) = 2560 bytes per app per second
 (2560) * (number-of-pods-per-node,100) = 256,000 bytes per second per node
 256k * (number-of-nodes) = total logging throughput per cluster

Fluentd ships any logs from e systemd journal and /var/lib/docker/containers/ to Elasticsearch. Learn more.

Local SSD drives are recommended in order to achieve the best performance. In Red Hat Enterprise Linux (RHEL) 7, the deadline IO scheduler is the default for all block devices except SATA disks. For SATA disks, the default IO scheduler is cfq.

Sizing storage for ES is greatly dependent on how you optimize your indices. Therefore, consider how much data you need in advance and that you are aggregating application log data. Some Elasticsearch users have found that it is necessary to keep absolute storage consumption around 50% and below 70% at all times. This helps to avoid Elasticsearch becoming unresponsive during large merge operations.