Full cluster restart upgrade

When you shut down a node, the allocation process will immediately try to replicate the shards that were on that node to other nodes in the cluster, causing a lot of wasted I/O. This can be avoided by disabling allocation before shutting down a node:

PUT /_cluster/settings
{
  "persistent": {
    "cluster.routing.allocation.enable": "none"
  }
}

If upgrading from 0.90.x to 1.x, then use these settings instead:

PUT /_cluster/settings
{
  "persistent": {
    "cluster.routing.allocation.disable_allocation": true,
    "cluster.routing.allocation.enable": "none"
  }
}

Shard recovery will be much faster if you stop indexing and issue a synced-flush request:

POST /_flush/synced

A synced flush request is a “best effort” operation. It will fail if there are any pending indexing operations, but it is safe to reissue the request multiple times if necessary.

Stop all Elasticsearch services on all nodes in the cluster. Each node can be upgraded following the same procedure described in Stop and upgrade a single node.

If you have dedicated master nodes — nodes with node.master set to true(the default) and node.data set to false —  then it is a good idea to start them first. Wait for them to form a cluster and to elect a master before proceeding with the data nodes. You can check progress by looking at the logs.

As soon as the minimum number of master-eligible nodes have discovered each other, they will form a cluster and elect a master. From that point on, the _cat/health and _cat/nodes APIs can be used to monitor nodes joining the cluster:

GET _cat/health

GET _cat/nodes

Use these APIs to check that all nodes have successfully joined the cluster.

As soon as each node has joined the cluster, it will start to recover any primary shards that are stored locally. Initially, the _cat/health request will report a status of red, meaning that not all primary shards have been allocated.

Once each node has recovered its local shards, the status will become yellow, meaning all primary shards have been recovered, but not all replica shards are allocated. This is to be expected because allocation is still disabled.

Delaying the allocation of replicas until all nodes have joined the cluster allows the master to allocate replicas to nodes which already have local shard copies. At this point, with all the nodes in the cluster, it is safe to reenable shard allocation:

PUT /_cluster/settings
{
  "persistent": {
    "cluster.routing.allocation.enable": "all"
  }
}

If upgrading from 0.90.x to 1.x, then use these settings instead:

PUT /_cluster/settings
{
  "persistent": {
    "cluster.routing.allocation.disable_allocation": false,
    "cluster.routing.allocation.enable": "all"
  }
}

The cluster will now start allocating replica shards to all data nodes. At this point it is safe to resume indexing and searching, but your cluster will recover more quickly if you can delay indexing and searching until all shards have recovered.

You can monitor progress with the _cat/health and _cat/recovery APIs:

GET _cat/health

GET _cat/recovery

Once the status column in the _cat/health output has reached green, all primary and replica shards have been successfully allocated.