Cross-cluster replication is done on an index-by-index basis. Replication is configured at the index level. For each configured replication there is a replication source index called the leader index and a replication target index called the follower index.

Replication is active-passive. This means that while the leader index can directly be written into, the follower index can not directly receive writes.

Replication is pull-based. This means that replication is driven by the follower index. This simplifies state management on the leader index and means that cross-cluster replication does not interfere with indexing on the leader index.

Cross-cluster replication requires remote clusters.

Configuring replicationedit

Replication can be configured in two ways:

For more information about managing cross-cluster replication in Kibana, see Working with remote clusters.

When you initiate replication either manually or through an auto-follow pattern, the follower index is created on the local cluster. Once the follower index is created, the remote recovery process copies all of the Lucene segment files from the remote cluster to the local cluster.

By default, if you initiate following manually (by using Kibana or the create follower API), the recovery process is asynchronous in relationship to the create follower request. The request returns before the remote recovery process completes. If you would like to wait on the process to complete, you can use the wait_for_active_shards parameter.

The mechanics of replicationedit

While replication is managed at the index level, replication is performed at the shard level. When a follower index is created, it is automatically configured to have an identical number of shards as the leader index. A follower shard task in the follower index pulls from the corresponding leader shard in the leader index by sending read requests for new operations. These read requests can be served from any copy of the leader shard (primary or replicas).

For each read request sent by the follower shard task, if there are new operations available on the leader shard, the leader shard responds with operations limited by the read parameters that you established when you configured the follower index. If there are no new operations available on the leader shard, the leader shard waits up to a configured timeout for new operations. If new operations occur within that timeout, the leader shard immediately responds with those new operations. Otherwise, if the timeout elapses, the follower shard replies that there are no new operations. The follower shard task updates some statistics and immediately sends another read request to the leader shard. This ensures that the network connections between the remote cluster and the local cluster are continually being used so as to avoid forceful termination by an external source (such as a firewall).

If a read request fails, the cause of the failure is inspected. If the cause of the failure is deemed to be a failure that can be recovered from (for example, a network failure), the follower shard task enters into a retry loop. Otherwise, the follower shard task is paused and requires user intervention before it can be resumed with the resume follower API.

When operations are received by the follower shard task, they are placed in a write buffer. The follower shard task manages this write buffer and submits bulk write requests from this write buffer to the follower shard. The write buffer and these write requests are managed by the write parameters that you established when you configured the follower index. The write buffer serves as back-pressure against read requests. If the write buffer exceeds its configured limits, no additional read requests are sent by the follower shard task. The follower shard task resumes sending read requests when the write buffer no longer exceeds its configured limits.

The intricacies of how operations are replicated from the leader are governed by settings that you can configure when you create the follower index in Kibana or by using the create follower API.

Mapping updates applied to the leader index are automatically retrieved as-needed by the follower index.

Settings updates applied to the leader index that are needed by the follower index are automatically retried as-needed by the follower index. Not all settings updates are needed by the follower index. For example, changing the number of replicas on the leader index is not replicated by the follower index.

If you apply a non-dynamic settings change to the leader index that is needed by the follower index, the follower index will go through a cycle of closing itself, applying the settings update, and then re-opening itself. The follower index will be unavailable for reads and not replicating writes during this cycle.

Inspecting the progress of replicationedit

You can inspect the progress of replication at the shard level with the get follower stats API. This API gives you insight into the read and writes managed by the follower shard task. It also reports read exceptions that can be retried and fatal exceptions that require user intervention.

Pausing and resuming replicationedit

You can pause replication with the pause follower API and then later resume replication with the resume follower API. Using these APIs in tandem enables you to adjust the read and write parameters on the follower shard task if your initial configuration is not suitable for your use case.

Leader index retaining operations for replicationedit

If the follower is unable to replicate operations from a leader for a period of time, the following process can fail due to the leader lacking a complete history of operations necessary for replication.

Operations replicated to the follower are identified using a sequence number generated when the operation was initially performed. Lucene segment files are occasionally merged in order to optimize searches and save space. When these merges occur, it is possible for operations associated with deleted or updated documents to be pruned during the merge. When the follower requests the sequence number for a pruned operation, the process will fail due to the operation missing on the leader.

This scenario is not possible in an append-only workflow. As documents are never deleted or updated, the underlying operation will not be pruned.

Elasticsearch attempts to mitigate this potential issue for update workflows using a Lucene feature called soft deletes. When a document is updated or deleted, the underlying operation is retained in the Lucene index for a period of time. This period of time is governed by the index.soft_deletes.retention_lease.period setting which can be configured on the leader index.

When a follower initiates the index following, it acquires a retention lease from the leader. This informs the leader that it should not allow a soft delete to be pruned until either the follower indicates that it has received the operation or the lease expires. It is valuable to have monitoring in place to detect a follower replication issue prior to the lease expiring so that the problem can be remedied before the follower falls fatally behind.

Remedying a follower that has fallen behindedit

If a follower falls sufficiently behind a leader that it can no longer replicate operations this can be detected in Kibana or by using the get follow stats API. It will be reported as a indices[].fatal_exception.

In order to restart the follower, you must pause the following process, close the index, and the create follower index again. For example:

POST /follower_index/_ccr/pause_follow

POST /follower_index/_close

PUT /follower_index/_ccr/follow?wait_for_active_shards=1
  "remote_cluster" : "remote_cluster",
  "leader_index" : "leader_index"

Re-creating the follower index is a destructive action. All of the existing Lucene segment files are deleted on the follower cluster. The remote recovery process copies the Lucene segment files from the leader again. After the follower index initializes, the following process starts again.

Terminating replicationedit

You can terminate replication with the unfollow API. This API converts a follower index to a regular (non-follower) index.