×

The guidance in this section is only relevant for installations with cloud provider integration.

Apply the following best practices to scale the number of worker machines in your OKD cluster. You scale the worker machines by increasing or decreasing the number of replicas that are defined in the worker machine set.

When scaling up the cluster to higher node counts:

  • Spread nodes across all of the available zones for higher availability.

  • Scale up by no more than 25 to 50 machines at once.

  • Consider creating new compute machine sets in each available zone with alternative instance types of similar size to help mitigate any periodic provider capacity constraints. For example, on AWS, use m5.large and m5d.large.

Cloud providers might implement a quota for API services. Therefore, gradually scale the cluster.

The controller might not be able to create the machines if the replicas in the compute machine sets are set to higher numbers all at one time. The number of requests the cloud platform, which OKD is deployed on top of, is able to handle impacts the process. The controller will start to query more while trying to create, check, and update the machines with the status. The cloud platform on which OKD is deployed has API request limits; excessive queries might lead to machine creation failures due to cloud platform limitations.

Enable machine health checks when scaling to large node counts. In case of failures, the health checks monitor the condition and automatically repair unhealthy machines.

When scaling large and dense clusters to lower node counts, it might take large amounts of time because the process involves draining or evicting the objects running on the nodes being terminated in parallel. Also, the client might start to throttle the requests if there are too many objects to evict. The default client queries per second (QPS) and burst rates are currently set to 5 and 10 respectively. These values cannot be modified in OKD.

The control plane node resource requirements depend on the number and type of nodes and objects in the cluster. The following control plane node size recommendations are based on the results of a control plane density focused testing, or Cluster-density. This test creates the following objects across a given number of namespaces:

  • 1 image stream

  • 1 build

  • 5 deployments, with 2 pod replicas in a sleep state, mounting 4 secrets, 4 config maps, and 1 downward API volume each

  • 5 services, each one pointing to the TCP/8080 and TCP/8443 ports of one of the previous deployments

  • 1 route pointing to the first of the previous services

  • 10 secrets containing 2048 random string characters

  • 10 config maps containing 2048 random string characters

Number of worker nodes Cluster-density (namespaces) CPU cores Memory (GB)

27

500

4

16

120

1000

8

32

252

4000

16, but 24 if using the OVN-Kubernetes network plug-in

64, but 128 if using the OVN-Kubernetes network plug-in

501, but untested with the OVN-Kubernetes network plug-in

4000

16

96

The data from the table above is based on an OKD running on top of AWS, using r5.4xlarge instances as control-plane nodes and m5.2xlarge instances as worker nodes.

On a large and dense cluster with three control plane nodes, the CPU and memory usage will spike up when one of the nodes is stopped, rebooted, or fails. The failures can be due to unexpected issues with power, network, underlying infrastructure, or intentional cases where the cluster is restarted after shutting it down to save costs. The remaining two control plane nodes must handle the load in order to be highly available, which leads to increase in the resource usage. This is also expected during upgrades because the control plane nodes are cordoned, drained, and rebooted serially to apply the operating system updates, as well as the control plane Operators update. To avoid cascading failures, keep the overall CPU and memory resource usage on the control plane nodes to at most 60% of all available capacity to handle the resource usage spikes. Increase the CPU and memory on the control plane nodes accordingly to avoid potential downtime due to lack of resources.

The node sizing varies depending on the number of nodes and object counts in the cluster. It also depends on whether the objects are actively being created on the cluster. During object creation, the control plane is more active in terms of resource usage compared to when the objects are in the running phase.

Operator Lifecycle Manager (OLM ) runs on the control plane nodes and its memory footprint depends on the number of namespaces and user installed operators that OLM needs to manage on the cluster. Control plane nodes need to be sized accordingly to avoid OOM kills. Following data points are based on the results from cluster maximums testing.

Number of namespaces OLM memory at idle state (GB) OLM memory with 5 user operators installed (GB)

500

0.823

1.7

1000

1.2

2.5

1500

1.7

3.2

2000

2

4.4

3000

2.7

5.6

4000

3.8

7.6

5000

4.2

9.02

6000

5.8

11.3

7000

6.6

12.9

8000

6.9

14.8

9000

8

17.7

10,000

9.9

21.6

You can modify the control plane node size in a running OKD 4 cluster for the following configurations only:

  • Clusters installed with a user-provisioned installation method.

  • AWS clusters installed with an installer-provisioned infrastructure installation method.

  • Clusters that use a control plane machine set to manage control plane machines.

For all other configurations, you must estimate your total node count and use the suggested control plane node size during installation.

The recommendations are based on the data points captured on OKD clusters with OpenShift SDN as the network plugin.

In OKD 4, half of a CPU core (500 millicore) is now reserved by the system by default compared to OKD 3.11 and previous versions. The sizes are determined taking that into consideration.

If the control plane machines in an Amazon Web Services (AWS) cluster require more resources, you can select a larger AWS instance type for the control plane machines to use.

The procedure for clusters that use a control plane machine set is different than the procedure for clusters that do not use a control plane machine set.

If you are uncertain about the state of the ControlPlaneMachineSet CR in your cluster, you can verify the CR status.

You can change the Amazon Web Services (AWS) instance type that your control plane machines use by updating the specification in the control plane machine set custom resource (CR).

Prerequisites
  • Your AWS cluster uses a control plane machine set.

Procedure
  1. Edit your control plane machine set CR by running the following command:

    $ oc --namespace openshift-machine-api edit controlplanemachineset.machine.openshift.io cluster
  2. Edit the following line under the providerSpec field:

    providerSpec:
      value:
        ...
        instanceType: <compatible_aws_instance_type> (1)
    1 Specify a larger AWS instance type with the same base as the previous selection. For example, you can change m6i.xlarge to m6i.2xlarge or m6i.4xlarge.
  3. Save your changes.

    • For clusters that use the default RollingUpdate update strategy, the Operator automatically propagates the changes to your control plane configuration.

    • For clusters that are configured to use the OnDelete update strategy, you must replace your control plane machines manually.

You can change the Amazon Web Services (AWS) instance type that your control plane machines use by updating the instance type in the AWS console.

Prerequisites
  • You have access to the AWS console with the permissions required to modify the EC2 Instance for your cluster.

  • You have access to the OKD cluster as a user with the cluster-admin role.

Procedure
  1. Open the AWS console and fetch the instances for the control plane machines.

  2. Choose one control plane machine instance.

    1. For the selected control plane machine, back up the etcd data by creating an etcd snapshot. For more information, see "Backing up etcd".

    2. In the AWS console, stop the control plane machine instance.

    3. Select the stopped instance, and click ActionsInstance SettingsChange instance type.

    4. Change the instance to a larger type, ensuring that the type is the same base as the previous selection, and apply changes. For example, you can change m6i.xlarge to m6i.2xlarge or m6i.4xlarge.

    5. Start the instance.

    6. If your OKD cluster has a corresponding Machine object for the instance, update the instance type of the object to match the instance type set in the AWS console.

  3. Repeat this process for each control plane machine.

Because etcd writes data to disk and persists proposals on disk, its performance depends on disk performance. Although etcd is not particularly I/O intensive, it requires a low latency block device for optimal performance and stability. Because etcd’s consensus protocol depends on persistently storing metadata to a log (WAL), etcd is sensitive to disk-write latency. Slow disks and disk activity from other processes can cause long fsync latencies.

Those latencies can cause etcd to miss heartbeats, not commit new proposals to the disk on time, and ultimately experience request timeouts and temporary leader loss. High write latencies also lead to an OpenShift API slowness, which affects cluster performance. Because of these reasons, avoid colocating other workloads on the control-plane nodes.

In terms of latency, run etcd on top of a block device that can write at least 50 IOPS of 8000 bytes long sequentially. That is, with a latency of 20ms, keep in mind that uses fdatasync to synchronize each write in the WAL. For heavy loaded clusters, sequential 500 IOPS of 8000 bytes (2 ms) are recommended. To measure those numbers, you can use a benchmarking tool, such as fio.

To achieve such performance, run etcd on machines that are backed by SSD or NVMe disks with low latency and high throughput. Consider single-level cell (SLC) solid-state drives (SSDs), which provide 1 bit per memory cell, are durable and reliable, and are ideal for write-intensive workloads.

The following hard disk features provide optimal etcd performance:

  • Low latency to support fast read operation.

  • High-bandwidth writes for faster compactions and defragmentation.

  • High-bandwidth reads for faster recovery from failures.

  • Solid state drives as a minimum selection, however NVMe drives are preferred.

  • Server-grade hardware from various manufacturers for increased reliability.

  • RAID 0 technology for increased performance.

  • Dedicated etcd drives. Do not place log files or other heavy workloads on etcd drives.

Avoid NAS or SAN setups and spinning drives. Always benchmark by using utilities such as fio. Continuously monitor the cluster performance as it increases.

Avoid using the Network File System (NFS) protocol or other network based file systems.

Some key metrics to monitor on a deployed OKD cluster are p99 of etcd disk write ahead log duration and the number of etcd leader changes. Use Prometheus to track these metrics.

The etcd member database sizes can vary in a cluster during normal operations. This difference does not affect cluster upgrades, even if the leader size is different from the other members.

To validate the hardware for etcd before or after you create the OKD cluster, you can use fio.

Prerequisites
  • Container runtimes such as Podman or Docker are installed on the machine that you’re testing.

  • Data is written to the /var/lib/etcd path.

Procedure
  • Run fio and analyze the results:

    • If you use Podman, run this command:

      $ sudo podman run --volume /var/lib/etcd:/var/lib/etcd:Z quay.io/openshift-scale/etcd-perf
    • If you use Docker, run this command:

      $ sudo docker run --volume /var/lib/etcd:/var/lib/etcd:Z quay.io/openshift-scale/etcd-perf

The output reports whether the disk is fast enough to host etcd by comparing the 99th percentile of the fsync metric captured from the run to see if it is less than 20 ms. A few of the most important etcd metrics that might affected by I/O performance are as follow:

  • etcd_disk_wal_fsync_duration_seconds_bucket metric reports the etcd’s WAL fsync duration

  • etcd_disk_backend_commit_duration_seconds_bucket metric reports the etcd backend commit latency duration

  • etcd_server_leader_changes_seen_total metric reports the leader changes

Because etcd replicates the requests among all the members, its performance strongly depends on network input/output (I/O) latency. High network latencies result in etcd heartbeats taking longer than the election timeout, which results in leader elections that are disruptive to the cluster. A key metric to monitor on a deployed OKD cluster is the 99th percentile of etcd network peer latency on each etcd cluster member. Use Prometheus to track the metric.

The histogram_quantile(0.99, rate(etcd_network_peer_round_trip_time_seconds_bucket[2m])) metric reports the round trip time for etcd to finish replicating the client requests between the members. Ensure that it is less than 50 ms.

For large and dense clusters, etcd can suffer from poor performance if the keyspace grows too large and exceeds the space quota. Periodically maintain and defragment etcd to free up space in the data store. Monitor Prometheus for etcd metrics and defragment it when required; otherwise, etcd can raise a cluster-wide alarm that puts the cluster into a maintenance mode that accepts only key reads and deletes.

Monitor these key metrics:

  • etcd_server_quota_backend_bytes, which is the current quota limit

  • etcd_mvcc_db_total_size_in_use_in_bytes, which indicates the actual database usage after a history compaction

  • etcd_debugging_mvcc_db_total_size_in_bytes, which shows the database size, including free space waiting for defragmentation

Defragment etcd data to reclaim disk space after events that cause disk fragmentation, such as etcd history compaction.

History compaction is performed automatically every five minutes and leaves gaps in the back-end database. This fragmented space is available for use by etcd, but is not available to the host file system. You must defragment etcd to make this space available to the host file system.

Defragmentation occurs automatically, but you can also trigger it manually.

Automatic defragmentation is good for most cases, because the etcd operator uses cluster information to determine the most efficient operation for the user.

The etcd Operator automatically defragments disks. No manual intervention is needed.

Verify that the defragmentation process is successful by viewing one of these logs:

  • etcd logs

  • cluster-etcd-operator pod

  • operator status error log

Automatic defragmentation can cause leader election failure in various OpenShift core components, such as the Kubernetes controller manager, which triggers a restart of the failing component. The restart is harmless and either triggers failover to the next running instance or the component resumes work again after the restart.

Example log output for successful defragmentation
etcd member has been defragmented: <member_name>, memberID: <member_id>
Example log output for unsuccessful defragmentation
failed defrag on member: <member_name>, memberID: <member_id>: <error_message>

A Prometheus alert indicates when you need to use manual defragmentation. The alert is displayed in two cases:

  • When etcd uses more than 50% of its available space for more than 10 minutes

  • When etcd is actively using less than 50% of its total database size for more than 10 minutes

You can also determine whether defragmentation is needed by checking the etcd database size in MB that will be freed by defragmentation with the PromQL expression: (etcd_mvcc_db_total_size_in_bytes - etcd_mvcc_db_total_size_in_use_in_bytes)/1024/1024

Defragmenting etcd is a blocking action. The etcd member will not respond until defragmentation is complete. For this reason, wait at least one minute between defragmentation actions on each of the pods to allow the cluster to recover.

Follow this procedure to defragment etcd data on each etcd member.

Prerequisites
  • You have access to the cluster as a user with the cluster-admin role.

Procedure
  1. Determine which etcd member is the leader, because the leader should be defragmented last.

    1. Get the list of etcd pods:

      $ oc get pods -n openshift-etcd -o wide | grep -v guard | grep etcd
      Example output
      etcd-ip-10-0-159-225.example.redhat.com                3/3     Running     0          175m   10.0.159.225   ip-10-0-159-225.example.redhat.com   <none>           <none>
      etcd-ip-10-0-191-37.example.redhat.com                 3/3     Running     0          173m   10.0.191.37    ip-10-0-191-37.example.redhat.com    <none>           <none>
      etcd-ip-10-0-199-170.example.redhat.com                3/3     Running     0          176m   10.0.199.170   ip-10-0-199-170.example.redhat.com   <none>           <none>
    2. Choose a pod and run the following command to determine which etcd member is the leader:

      $ oc rsh -n openshift-etcd etcd-ip-10-0-159-225.example.redhat.com etcdctl endpoint status --cluster -w table
      Example output
      Defaulting container name to etcdctl.
      Use 'oc describe pod/etcd-ip-10-0-159-225.example.redhat.com -n openshift-etcd' to see all of the containers in this pod.
      +---------------------------+------------------+---------+---------+-----------+------------+-----------+------------+--------------------+--------+
      |         ENDPOINT          |        ID        | VERSION | DB SIZE | IS LEADER | IS LEARNER | RAFT TERM | RAFT INDEX | RAFT APPLIED INDEX | ERRORS |
      +---------------------------+------------------+---------+---------+-----------+------------+-----------+------------+--------------------+--------+
      |  https://10.0.191.37:2379 | 251cd44483d811c3 |   3.4.9 |  104 MB |     false |      false |         7 |      91624 |              91624 |        |
      | https://10.0.159.225:2379 | 264c7c58ecbdabee |   3.4.9 |  104 MB |     false |      false |         7 |      91624 |              91624 |        |
      | https://10.0.199.170:2379 | 9ac311f93915cc79 |   3.4.9 |  104 MB |      true |      false |         7 |      91624 |              91624 |        |
      +---------------------------+------------------+---------+---------+-----------+------------+-----------+------------+--------------------+--------+

      Based on the IS LEADER column of this output, the https://10.0.199.170:2379 endpoint is the leader. Matching this endpoint with the output of the previous step, the pod name of the leader is etcd-ip-10-0-199-170.example.redhat.com.

  2. Defragment an etcd member.

    1. Connect to the running etcd container, passing in the name of a pod that is not the leader:

      $ oc rsh -n openshift-etcd etcd-ip-10-0-159-225.example.redhat.com
    2. Unset the ETCDCTL_ENDPOINTS environment variable:

      sh-4.4# unset ETCDCTL_ENDPOINTS
    3. Defragment the etcd member:

      sh-4.4# etcdctl --command-timeout=30s --endpoints=https://localhost:2379 defrag
      Example output
      Finished defragmenting etcd member[https://localhost:2379]

      If a timeout error occurs, increase the value for --command-timeout until the command succeeds.

    4. Verify that the database size was reduced:

      sh-4.4# etcdctl endpoint status -w table --cluster
      Example output
      +---------------------------+------------------+---------+---------+-----------+------------+-----------+------------+--------------------+--------+
      |         ENDPOINT          |        ID        | VERSION | DB SIZE | IS LEADER | IS LEARNER | RAFT TERM | RAFT INDEX | RAFT APPLIED INDEX | ERRORS |
      +---------------------------+------------------+---------+---------+-----------+------------+-----------+------------+--------------------+--------+
      |  https://10.0.191.37:2379 | 251cd44483d811c3 |   3.4.9 |  104 MB |     false |      false |         7 |      91624 |              91624 |        |
      | https://10.0.159.225:2379 | 264c7c58ecbdabee |   3.4.9 |   41 MB |     false |      false |         7 |      91624 |              91624 |        | (1)
      | https://10.0.199.170:2379 | 9ac311f93915cc79 |   3.4.9 |  104 MB |      true |      false |         7 |      91624 |              91624 |        |
      +---------------------------+------------------+---------+---------+-----------+------------+-----------+------------+--------------------+--------+

      This example shows that the database size for this etcd member is now 41 MB as opposed to the starting size of 104 MB.

    5. Repeat these steps to connect to each of the other etcd members and defragment them. Always defragment the leader last.

      Wait at least one minute between defragmentation actions to allow the etcd pod to recover. Until the etcd pod recovers, the etcd member will not respond.

  3. If any NOSPACE alarms were triggered due to the space quota being exceeded, clear them.

    1. Check if there are any NOSPACE alarms:

      sh-4.4# etcdctl alarm list
      Example output
      memberID:12345678912345678912 alarm:NOSPACE
    2. Clear the alarms:

      sh-4.4# etcdctl alarm disarm
Next steps

After defragmentation, if etcd still uses more than 50% of its available space, consider increasing the disk quota for etcd.

Infrastructure nodes are nodes that are labeled to run pieces of the OKD environment. The infrastructure node resource requirements depend on the cluster age, nodes, and objects in the cluster, as these factors can lead to an increase in the number of metrics or time series in Prometheus. The following infrastructure node size recommendations are based on the results observed in cluster-density testing detailed in the Control plane node sizing section, where the monitoring stack and the default ingress-controller were moved to these nodes.

Number of worker nodes Cluster density, or number of namespaces CPU cores Memory (GB)

25

500

4

48

100

1000

8

96

252

4000

16

128

501

4000

32

128

In general, three infrastructure nodes are recommended per cluster.

These sizing recommendations are based on scale tests, which create a large number of objects across the cluster. These tests include reaching some of the cluster maximums. In the case of 250 and 500 node counts on an OKD 4 cluster, these maximums are 10000 namespaces with 61000 pods, 10000 deployments, 181000 secrets, 400 config maps, and so on. Prometheus is a highly memory intensive application; the resource usage depends on various factors including the number of nodes, objects, the Prometheus metrics scraping interval, metrics or time series, and the age of the cluster. The disk size also depends on the retention period. You must take these factors into consideration and size them accordingly.

These sizing recommendations are only applicable for the Prometheus, Router, and Registry infrastructure components, which are installed during cluster installation. Logging is a day-two operation and is not included in these recommendations.

In OKD 4, half of a CPU core (500 millicore) is now reserved by the system by default compared to OKD 3.11 and previous versions. This influences the stated sizing recommendations.