Full-cluster restart and rolling restartedit

There may be situations where you want to perform a full-cluster restart or a rolling restart. In the case of full-cluster restart, you shut down and restart all the nodes in the cluster while in the case of rolling restart, you shut down only one node at a time, so the service remains uninterrupted.

Full-cluster restartedit

  1. Disable shard allocation.

    When you shut down a data node, the allocation process waits for index.unassigned.node_left.delayed_timeout (by default, one minute) before starting to replicate the shards on that node to other nodes in the cluster, which can involve a lot of I/O. Since the node is shortly going to be restarted, this I/O is unnecessary. You can avoid racing the clock by disabling allocation of replicas before shutting down data nodes:

    PUT _cluster/settings
    {
      "persistent": {
        "cluster.routing.allocation.enable": "primaries"
      }
    }
  2. Stop indexing and perform a synced flush.

    Performing a synced-flush speeds up shard recovery.

    POST _flush/synced

    When you perform a synced flush, check the response to make sure there are no failures. Synced flush operations that fail due to pending indexing operations are listed in the response body, although the request itself still returns a 200 OK status. If there are failures, reissue the request.

    Note that synced flush is deprecated and will be removed in 8.0. A flush has the same effect as a synced flush on Elasticsearch 7.6 or later.

  3. Temporarily stop the tasks associated with active machine learning jobs and datafeeds. (Optional)

    Machine learning features require specific subscriptions.

    You have two options to handle machine learning jobs and datafeeds when you shut down a cluster:

    • Temporarily halt the tasks associated with your machine learning jobs and datafeeds and prevent new jobs from opening by using the set upgrade mode API:

      POST _ml/set_upgrade_mode?enabled=true

      When you disable upgrade mode, the jobs resume using the last model state that was automatically saved. This option avoids the overhead of managing active jobs during the shutdown and is faster than explicitly stopping datafeeds and closing jobs.

    • Stop all datafeeds and close all jobs. This option saves the model state at the time of closure. When you reopen the jobs after the cluster restart, they use the exact same model. However, saving the latest model state takes longer than using upgrade mode, especially if you have a lot of jobs or jobs with large model states.
  4. Shut down all nodes.

    • If you are running Elasticsearch with systemd:

      sudo systemctl stop elasticsearch.service
    • If you are running Elasticsearch with SysV init:

      sudo -i service elasticsearch stop
    • If you are running Elasticsearch as a daemon:

      kill $(cat pid)
  5. Perform any needed changes.
  6. Restart nodes.

    If you have dedicated master nodes, start them first and wait for them to form a cluster and elect a master before proceeding with your data nodes. You can check progress by looking at the logs.

    As soon as enough master-eligible nodes have discovered each other, they form a cluster and elect a master. At that point, you can use the cat health and cat nodes APIs to monitor nodes joining the cluster:

    GET _cat/health
    
    GET _cat/nodes

    The status column returned by _cat/health shows the health of each node in the cluster: red, yellow, or green.

  7. Wait for all nodes to join the cluster and report a status of yellow.

    When a node joins the cluster, it begins to recover any primary shards that are stored locally. The _cat/health API initially reports a status of red, indicating that not all primary shards have been allocated.

    Once a node recovers its local shards, the cluster status switches to yellow, indicating that all primary shards have been recovered, but not all replica shards are allocated. This is to be expected because you have not yet re-enabled allocation. Delaying the allocation of replicas until all nodes are yellow allows the master to allocate replicas to nodes that already have local shard copies.

  8. Re-enable allocation.

    When all nodes have joined the cluster and recovered their primary shards, re-enable allocation by restoring cluster.routing.allocation.enable to its default:

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

    Once allocation is re-enabled, the cluster starts allocating replica shards to the data nodes. At this point it is safe to resume indexing and searching, but your cluster will recover more quickly if you can wait until all primary and replica shards have been successfully allocated and the status of all nodes is green.

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

    GET _cat/health
    
    GET _cat/recovery
  9. Restart machine learning jobs. (Optional)

    If you temporarily halted the tasks associated with your machine learning jobs, use the set upgrade mode API to return them to active states:

    POST _ml/set_upgrade_mode?enabled=false

    If you closed all machine learning jobs before stopping the nodes, open the jobs and start the datafeeds from Kibana or with the open jobs and start datafeed APIs.

Rolling restartedit

  1. Disable shard allocation.

    When you shut down a data node, the allocation process waits for index.unassigned.node_left.delayed_timeout (by default, one minute) before starting to replicate the shards on that node to other nodes in the cluster, which can involve a lot of I/O. Since the node is shortly going to be restarted, this I/O is unnecessary. You can avoid racing the clock by disabling allocation of replicas before shutting down data nodes:

    PUT _cluster/settings
    {
      "persistent": {
        "cluster.routing.allocation.enable": "primaries"
      }
    }
  2. Stop indexing and perform a synced flush.

    Performing a synced-flush speeds up shard recovery.

    POST _flush/synced

    When you perform a synced flush, check the response to make sure there are no failures. Synced flush operations that fail due to pending indexing operations are listed in the response body, although the request itself still returns a 200 OK status. If there are failures, reissue the request.

    Note that synced flush is deprecated and will be removed in 8.0. A flush has the same effect as a synced flush on Elasticsearch 7.6 or later.

  3. Temporarily stop the tasks associated with active machine learning jobs and datafeeds. (Optional)

    Machine learning features require specific subscriptions.

    You have two options to handle machine learning jobs and datafeeds when you shut down a cluster:

    • Temporarily halt the tasks associated with your machine learning jobs and datafeeds and prevent new jobs from opening by using the set upgrade mode API:

      POST _ml/set_upgrade_mode?enabled=true

      When you disable upgrade mode, the jobs resume using the last model state that was automatically saved. This option avoids the overhead of managing active jobs during the shutdown and is faster than explicitly stopping datafeeds and closing jobs.

    • Stop all datafeeds and close all jobs. This option saves the model state at the time of closure. When you reopen the jobs after the cluster restart, they use the exact same model. However, saving the latest model state takes longer than using upgrade mode, especially if you have a lot of jobs or jobs with large model states.
    • If you perform a rolling restart, you can also leave your machine learning jobs running. When you shut down a machine learning node, its jobs automatically move to another node and restore the model states. This option enables your jobs to continue running during the shutdown but it puts increased load on the cluster.
  4. Shut down a single node in case of rolling restart.

    • If you are running Elasticsearch with systemd:

      sudo systemctl stop elasticsearch.service
    • If you are running Elasticsearch with SysV init:

      sudo -i service elasticsearch stop
    • If you are running Elasticsearch as a daemon:

      kill $(cat pid)
  5. Perform any needed changes.
  6. Restart the node you changed.

    Start the node and confirm that it joins the cluster by checking the log file or by submitting a _cat/nodes request:

    GET _cat/nodes
  7. Reenable shard allocation.

    For data nodes, once the node has joined the cluster, remove the cluster.routing.allocation.enable setting to enable shard allocation and start using the node:

    PUT _cluster/settings
    {
      "persistent": {
        "cluster.routing.allocation.enable": null
      }
    }
  8. Repeat in case of rolling restart.

    When the node has recovered and the cluster is stable, repeat these steps for each node that needs to be changed.

  9. Restart machine learning jobs. (Optional)

    If you temporarily halted the tasks associated with your machine learning jobs, use the set upgrade mode API to return them to active states:

    POST _ml/set_upgrade_mode?enabled=false

    If you closed all machine learning jobs before stopping the nodes, open the jobs and start the datafeeds from Kibana or with the open jobs and start datafeed APIs.