A split operation:

Going from N shards to N+1 shards, aka. incremental resharding, is indeed a feature that is supported by many key-value stores. Adding a new shard and pushing new data to this new shard only is not an option: this would likely be an indexing bottleneck, and figuring out which shard a document belongs to given its _id, which is necessary for get, delete and update requests, would become quite complex. This means that we need to rebalance existing data using a different hashing scheme.

The most common way that key-value stores do this efficiently is by using consistent hashing. Consistent hashing only requires 1/N-th of the keys to be relocated when growing the number of shards from N to N+1. However Elasticsearch’s unit of storage, shards, are Lucene indices. Because of their search-oriented data structure, taking a significant portion of a Lucene index, be it only 5% of documents, deleting them and indexing them on another shard typically comes with a much higher cost than with a key-value store. This cost is kept reasonable when growing the number of shards by a multiplicative factor as described in the above section: this allows Elasticsearch to perform the split locally, which in-turn allows to perform the split at the index level rather than reindexing documents that need to move, as well as using hard links for efficient file copying.

In the case of append-only data, it is possible to get more flexibility by creating a new index and pushing new data to it, while adding an alias that covers both the old and the new index for read operations. Assuming that the old and new indices have respectively M and N shards, this has no overhead compared to searching an index that would have M+N shards.

To split my_source_index into a new index called my_target_index, issue the following request:

POST /my_source_index/_split/my_target_index
{
  "settings": {
    "index.number_of_shards": 2
  }
}

The above request returns immediately once the target index has been added to the cluster state — it doesn’t wait for the split operation to start.

The _split API is similar to the create index API and accepts settings and aliases parameters for the target index:

POST /my_source_index/_split/my_target_index
{
  "settings": {
    "index.number_of_shards": 5 
  },
  "aliases": {
    "my_search_indices": {}
  }
}

The split process can be monitored with the _cat recovery API, or the cluster health API can be used to wait until all primary shards have been allocated by setting the wait_for_status parameter to yellow.

The _split API returns as soon as the target index has been added to the cluster state, before any shards have been allocated. At this point, all shards are in the state unassigned. If, for any reason, the target index can’t be allocated, its primary shard will remain unassigned until it can be allocated on that node.

Once the primary shard is allocated, it moves to state initializing, and the split process begins. When the split operation completes, the shard will become active. At that point, Elasticsearch will try to allocate any replicas and may decide to relocate the primary shard to another node.

Because the split operation creates a new index to split the shards to, the wait for active shards setting on index creation applies to the split index action as well.