From 7d816aecf10932ace7609b7684dd3637d660e1c8 Mon Sep 17 00:00:00 2001 From: Bogdan Dobrelya Date: Fri, 22 Jul 2016 14:42:44 +0200 Subject: [PATCH] Add ha docs Signed-off-by: Bogdan Dobrelya --- docs/ha-mode.md | 34 ++++++++++++++++++++++++++++++++++ 1 file changed, 34 insertions(+) create mode 100644 docs/ha-mode.md diff --git a/docs/ha-mode.md b/docs/ha-mode.md new file mode 100644 index 000000000..b3d752cc9 --- /dev/null +++ b/docs/ha-mode.md @@ -0,0 +1,34 @@ +HA endpoints for K8s +==================== + +The following components require a highly available endpoints: +* etcd cluster, +* kube-apiserver service instances. + +The former provides the +[etcd-proxy](https://coreos.com/etcd/docs/latest/proxy.html) service to access +the cluster members in HA fashion. + +The latter relies on a 3rd side reverse proxies, like Nginx or HAProxy, to +achieve the same goal. + +Etcd +---- + +Etcd proxies are deployed on each node in the `k8s-cluster` group. A proxy is +a separate etcd process. It has a `localhost:2379` frontend and all of the etcd +cluster members as backends. Note that the `access_ip` is used as the backend +IP, if specified. Frontend endpoints cannot be accessed externally as they are +bound to a localhost only. + +The `etcd_access_endpoint` fact provides an access pattern for clients. And the +`etcd_multiaccess` (defaults to `false`) group var controlls that behavior. +When enabled, it makes deployed components to access the etcd cluster members +directly: `http://ip1:2379, http://ip2:2379,...`. This mode assumes the clients +do a loadbalancing and handle HA for connections. Note, a pod definition of a +flannel networking plugin always uses a single `--etcd-server` endpoint! + + +Kube-apiserver +-------------- +TODO(bogdando) TBD