Kay Yan
b7d99749b8
* preinstall: generalize minimum-value semantics for managed sysctls Move the hardcoded sysctl floor map out of 0080-system-configurations.yml into a default variable, and introduce additional_sysctl_min so users can declare their own sysctl floors. Floor semantics only apply a managed value when the node's current value is lower; pre-tuned images (Bottlerocket, EKS AMIs, OpenShift RHCOS, hardened distros) and newer kernels that already set higher values are left alone. Default floors are aligned with prior art from the OpenShift Node Tuning Operator (TuneD '=>' floor syntax in the openshift / openshift-node profiles) and kops. None of the floor keys are part of the kubelet protect-kernel-defaults validation list, so the floor mechanism cannot interfere with kubelet startup. Keys not present on the running kernel/distro are skipped without failing. additional_sysctl (exact-value semantics) is unchanged for backward compatibility. Signed-off-by: Kay Yan <kay.yan@daocloud.io> * preinstall: refine sysctl floor handling Keep built-in floor values internal while preserving user overrides through additional_sysctl_min. Use sysctl --ignore for unknown keys and filter empty results without masking other read errors. Signed-off-by: Kay Yan <kay.yan@daocloud.io> --------- Signed-off-by: Kay Yan <kay.yan@daocloud.io>
3.0 KiB
Large deployments of K8s
For a large scaled deployments, consider the following configuration changes:
-
Tune ansible settings for
forksandtimeoutvars to fit large numbers of nodes being deployed. -
Override containers'
foo_image_repovars to point to intranet registry. -
Override the
download_run_once: trueand/ordownload_localhost: true. See Downloading binaries and containers for details. -
Adjust the
retry_staggerglobal var as appropriate. It should provide sane load on a delegate (the first K8s control plane node) then retrying failed push or download operations. -
Tune parameters for DNS related applications Those are
dns_replicas,dns_cpu_limit,dns_cpu_requests,dns_memory_limit,dns_memory_requests. Please note that limits must always be greater than or equal to requests. -
Tune CPU/memory limits and requests. Those are located in roles' defaults and named like
foo_memory_limit,foo_memory_requestsandfoo_cpu_limit,foo_cpu_requests. Note that 'Mi' memory units for K8s will be submitted as 'M', if applied fordocker run, and cpu K8s units will end up with the 'm' skipped for docker as well. This is required as docker does not understand k8s units well. -
Tune
kubelet_status_update_frequencyto increase reliability of kubelet.kube_controller_node_monitor_grace_period,kube_controller_node_monitor_period,kube_apiserver_pod_eviction_not_ready_timeout_seconds&kube_apiserver_pod_eviction_unreachable_timeout_secondsfor better Kubernetes reliability. Check out Kubernetes Reliability -
Tune network prefix sizes. Those are
kube_network_node_prefix,kube_service_addressesandkube_pods_subnet. -
Add calico_rr nodes if you are deploying with Calico or Canal. Nodes recover from host/network interruption much quicker with calico_rr.
-
Check out the Inventory section of the Getting started guide for tips on creating a large scale Ansible inventory.
-
Override the
etcd_events_cluster_setup: truestore events in a separate dedicated etcd instance.
For example, when deploying 200 nodes, you may want to run ansible with
--forks=50, --timeout=600 and define the retry_stagger: 60.
Sysctl tuning with floor semantics
Built-in sysctl floors are only applied when the node's current value is
lower, so pre-tuned images and newer kernels are left alone. Use
additional_sysctl_min to add or override floors (set a key to 0 to
disable a default); use additional_sysctl for exact-value semantics.
additional_sysctl_min:
net.core.somaxconn: 65535
kernel.pid_max: 0
See roles/kubespray_defaults/vars/main/main.yml for the built-in floor
set.