Shadow replica indices

edit

Deprecated in 5.2.0.

Shadow replicas don’t see much usage and we are planning to remove them

If you would like to use a shared filesystem, you can use the shadow replicas settings to choose where on disk the data for an index should be kept, as well as how Elasticsearch should replay operations on all the replica shards of an index.

In order to fully utilize the index.data_path and index.shadow_replicas settings, you need to allow Elasticsearch to use the same data directory for multiple instances by setting node.add_lock_id_to_custom_path to false in elasticsearch.yml:

node.add_lock_id_to_custom_path: false

You will also need to indicate to the security manager where the custom indices will be, so that the correct permissions can be applied. You can do this by setting the path.shared_data setting in elasticsearch.yml:

path.shared_data: /opt/data

This means that Elasticsearch can read and write to files in any subdirectory of the path.shared_data setting.

You can then create an index with a custom data path, where each node will use this path for the data:

Because shadow replicas do not index the document on replica shards, it’s possible for the replica’s known mapping to be behind the index’s known mapping if the latest cluster state has not yet been processed on the node containing the replica. Because of this, it is highly recommended to use pre-defined mappings when using shadow replicas.

curl -XPUT 'localhost:9200/my_index' -H 'Content-Type: application/json' -d '
{
    "index" : {
        "number_of_shards" : 1,
        "number_of_replicas" : 4,
        "data_path": "/opt/data/my_index",
        "shadow_replicas": true
    }
}'

In the above example, the "/opt/data/my_index" path is a shared filesystem that must be available on every node in the Elasticsearch cluster. You must also ensure that the Elasticsearch process has the correct permissions to read from and write to the directory used in the index.data_path setting.

The data_path does not have to contain the index name, in this case, "my_index" was used but it could easily also have been "/opt/data/"

An index that has been created with the index.shadow_replicas setting set to "true" will not replicate document operations to any of the replica shards, instead, it will only continually refresh. Once segments are available on the filesystem where the shadow replica resides (after an Elasticsearch "flush"), a regular refresh (governed by the index.refresh_interval) can be used to make the new data searchable.

Since documents are only indexed on the primary shard, realtime GET requests could fail to return a document if executed on the replica shard, therefore, GET API requests automatically have the ?preference=_primary flag set if there is no preference flag already set.

In order to ensure the data is being synchronized in a fast enough manner, you may need to tune the flush threshold for the index to a desired number. A flush is needed to fsync segment files to disk, so they will be visible to all other replica nodes. Users should test what flush threshold levels they are comfortable with, as increased flushing can impact indexing performance.

The Elasticsearch cluster will still detect the loss of a primary shard, and transform the replica into a primary in this situation. This transformation will take slightly longer, since no IndexWriter is maintained for each shadow replica.

Below is the list of settings that can be changed using the update settings API:

index.data_path (string)
Path to use for the index’s data. Note that by default Elasticsearch will append the node ordinal by default to the path to ensure multiple instances of Elasticsearch on the same machine do not share a data directory.
index.shadow_replicas
Boolean value indicating this index should use shadow replicas. Defaults to false.
index.shared_filesystem
Boolean value indicating this index uses a shared filesystem. Defaults to the true if index.shadow_replicas is set to true, false otherwise.
index.shared_filesystem.recover_on_any_node
Boolean value indicating whether the primary shards for the index should be allowed to recover on any node in the cluster. If a node holding a copy of the shard is found, recovery prefers that node. Defaults to false.