Search across clustersedit

Cross-cluster search lets you run a single search request against one or more remote clusters. For example, you can use a cross-cluster search to filter and analyze log data stored on clusters in different data centers.

Supported APIsedit

The following APIs support cross-cluster search:

  • Search
  • Async search
  • Multi search
  • Search template
  • Multi search template
  • Field capabilities
  • Painless execute API
  • [preview] This functionality is in technical preview and may be changed or removed in a future release. Elastic will work to fix any issues, but features in technical preview are not subject to the support SLA of official GA features. EQL search
  • [preview] This functionality is in technical preview and may be changed or removed in a future release. Elastic will work to fix any issues, but features in technical preview are not subject to the support SLA of official GA features. SQL search
  • [preview] This functionality is in technical preview and may be changed or removed in a future release. Elastic will work to fix any issues, but features in technical preview are not subject to the support SLA of official GA features. Vector tile search

Prerequisitesedit

  • If you use sniff mode, the local coordinating node must be able to connect to seed and gateway nodes on the remote cluster.

    We recommend using gateway nodes capable of serving as coordinating nodes. The seed nodes can be a subset of these gateway nodes.

  • If you use proxy mode, the local coordinating node must be able to connect to the configured proxy_address. The proxy at this address must be able to route connections to gateway and coordinating nodes on the remote cluster.
  • Cross-cluster search requires different security privileges on the local cluster and remote cluster. See Configure privileges for cross-cluster search and Remote clusters.

Cross-cluster search examplesedit

Remote cluster setupedit

The following cluster update settings API request adds three remote clusters: cluster_one, cluster_two, and cluster_three.

response = client.cluster.put_settings(
  body: {
    persistent: {
      cluster: {
        remote: {
          cluster_one: {
            seeds: [
              '35.238.149.1:9300'
            ],
            skip_unavailable: true
          },
          cluster_two: {
            seeds: [
              '35.238.149.2:9300'
            ],
            skip_unavailable: false
          },
          cluster_three: {
            seeds: [
              '35.238.149.3:9300'
            ]
          }
        }
      }
    }
  }
)
puts response
PUT _cluster/settings
{
  "persistent": {
    "cluster": {
      "remote": {
        "cluster_one": {
          "seeds": [
            "35.238.149.1:9300"
          ],
          "skip_unavailable": true
        },
        "cluster_two": {
          "seeds": [
            "35.238.149.2:9300"
          ],
          "skip_unavailable": false
        },
        "cluster_three": {  
          "seeds": [
            "35.238.149.3:9300"
          ]
        }
      }
    }
  }
}

Since skip_unavailable was not set on cluster_three, it uses the default of false. See the Optional remote clusters section for details.

Search a single remote clusteredit

In the search request, you specify data streams and indices on a remote cluster as <remote_cluster_name>:<target>.

The following search API request searches the my-index-000001 index on a single remote cluster, cluster_one.

response = client.search(
  index: 'cluster_one:my-index-000001',
  body: {
    size: 1,
    query: {
      match: {
        'user.id' => 'kimchy'
      }
    },
    _source: [
      'user.id',
      'message',
      'http.response.status_code'
    ]
  }
)
puts response
GET /cluster_one:my-index-000001/_search
{
  "size": 1,
  "query": {
    "match": {
      "user.id": "kimchy"
    }
  },
  "_source": ["user.id", "message", "http.response.status_code"]
}

The API returns the following response. Note that when you search one or more remote clusters, a _clusters section is included to provide information about the search on each cluster.

{
  "took": 150,
  "timed_out": false,
  "_shards": {
    "total": 12,
    "successful": 12,
    "failed": 0,
    "skipped": 0
  },
  "_clusters": {
    "total": 1,  
    "successful": 1,
    "skipped": 0,
    "running": 0,
    "partial": 0,
    "failed": 0,
    "details": {
      "cluster_one": {  
        "status": "successful",
        "indices": "my-index-000001", 
        "took": 148,  
        "timed_out": false,
        "_shards": {  
          "total": 12,
          "successful": 12,
          "skipped": 0,
          "failed": 0
        }
      }
    }
  },
  "hits": {
    "total" : {
        "value": 1,
        "relation": "eq"
    },
    "max_score": 1,
    "hits": [
      {
        "_index": "cluster_one:my-index-000001", 
        "_id": "0",
        "_score": 1,
        "_source": {
          "user": {
            "id": "kimchy"
          },
          "message": "GET /search HTTP/1.1 200 1070000",
          "http": {
            "response":
              {
                "status_code": 200
              }
          }
        }
      }
    ]
  }
}

This section of counters shows all possible cluster search states and how many cluster searches are currently in that state. The clusters can be one of the following statuses: running, successful (searches on all shards were successful), partial (searches on at least one shard of the cluster was successful and at least one failed), skipped (the search failed on a cluster marked with skip_unavailable=true) or failed (the search failed on a cluster marked with skip_unavailable=false).

The _clusters/details section shows metadata about the search on each cluster.

The index expression supplied by the user. If you provide a wildcard such as logs-*, this section will show the value with the wildcard, not the concrete indices being searched.

How long (in milliseconds) the sub-search took on that cluster.

The shard details for the sub-search on that cluster.

The search response body includes the name of the remote cluster in the _index parameter.

Search multiple remote clustersedit

The following search API request searches the my-index-000001 index on three clusters:

  • The local ("querying") cluster, with 10 shards
  • Two remote clusters, cluster_one, with 12 shards and cluster_two with 6 shards.
response = client.search(
  index: 'my-index-000001,cluster_one:my-index-000001,cluster_two:my-index-000001',
  body: {
    query: {
      match: {
        'user.id' => 'kimchy'
      }
    },
    _source: [
      'user.id',
      'message',
      'http.response.status_code'
    ]
  }
)
puts response
GET /my-index-000001,cluster_one:my-index-000001,cluster_two:my-index-000001/_search
{
  "query": {
    "match": {
      "user.id": "kimchy"
    }
  },
  "_source": ["user.id", "message", "http.response.status_code"]
}

The API returns the following response:

{
  "took": 150,
  "timed_out": false,
  "num_reduce_phases": 4,
  "_shards": {
    "total": 28,
    "successful": 28,
    "failed": 0,
    "skipped": 0
  },
  "_clusters": {
    "total": 3,
    "successful": 3,
    "skipped": 0,
    "running": 0,
    "partial": 0,
    "failed": 0,
    "details": {
      "(local)": {            
        "status": "successful",
        "indices": "my-index-000001",
        "took": 21,
        "timed_out": false,
        "_shards": {
          "total": 10,
          "successful": 10,
          "skipped": 0,
          "failed": 0
        }
      },
      "cluster_one": {
        "status": "successful",
        "indices": "my-index-000001",
        "took": 48,
        "timed_out": false,
        "_shards": {
          "total": 12,
          "successful": 12,
          "skipped": 0,
          "failed": 0
        }
      },
      "cluster_two": {
        "status": "successful",
        "indices": "my-index-000001",
        "took": 141,
        "timed_out": false,
        "_shards": {
          "total" : 6,
          "successful" : 6,
          "skipped": 0,
          "failed": 0
        }
      }
    }
  },
  "hits": {
    "total" : {
        "value": 3,
        "relation": "eq"
    },
    "max_score": 1,
    "hits": [
      {
        "_index": "my-index-000001", 
        "_id": "0",
        "_score": 2,
        "_source": {
          "user": {
            "id": "kimchy"
          },
          "message": "GET /search HTTP/1.1 200 1070000",
          "http": {
            "response":
              {
                "status_code": 200
              }
          }
        }
      },
      {
        "_index": "cluster_one:my-index-000001", 
        "_id": "0",
        "_score": 1,
        "_source": {
          "user": {
            "id": "kimchy"
          },
          "message": "GET /search HTTP/1.1 200 1070000",
          "http": {
            "response":
              {
                "status_code": 200
              }
          }
        }
      },
      {
        "_index": "cluster_two:my-index-000001", 
        "_id": "0",
        "_score": 1,
        "_source": {
          "user": {
            "id": "kimchy"
          },
          "message": "GET /search HTTP/1.1 200 1070000",
          "http": {
            "response":
              {
                "status_code": 200
              }
          }
        }
      }
    ]
  }
}

The local (querying) cluster is identified as "(local)".

This document’s _index parameter doesn’t include a cluster name. This means the document came from the local cluster.

This document came from cluster_one.

This document came from cluster_two.

Using async search for cross-cluster search with ccs_minimize_roundtrips=trueedit

Remote clusters can be queried asynchronously using the async search API. A cross-cluster search accepts a ccs_minimize_roundtrips parameter. For asynchronous searches it defaults to false. (Note: for synchronous searches it defaults to true.) See Considerations for choosing whether to minimize roundtrips in a cross-cluster search to learn more about this option.

The following request does an asynchronous search of the my-index-000001 index using ccs_minimize_roundtrips=true against three clusters (same ones as the previous example).

response = client.async_search.submit(
  index: 'my-index-000001,cluster_one:my-index-000001,cluster_two:my-index-000001',
  ccs_minimize_roundtrips: true,
  body: {
    query: {
      match: {
        'user.id' => 'kimchy'
      }
    },
    _source: [
      'user.id',
      'message',
      'http.response.status_code'
    ]
  }
)
puts response
POST /my-index-000001,cluster_one:my-index-000001,cluster_two:my-index-000001/_async_search?ccs_minimize_roundtrips=true
{
  "query": {
    "match": {
      "user.id": "kimchy"
    }
  },
  "_source": ["user.id", "message", "http.response.status_code"]
}

The API returns the following response:

{
  "id": "FklQYndoTDJ2VEFlMEVBTzFJMGhJVFEaLVlKYndBWWZSMUdicUc4WVlEaFl4ZzoxNTU=", 
  "is_partial": true,
  "is_running": true,
  "start_time_in_millis": 1685563581380,
  "expiration_time_in_millis": 1685995581380,
  "response": {
    "took": 1020,
    "timed_out": false,
    "num_reduce_phases": 0,
    "_shards": {
      "total": 10,     
      "successful": 0,
      "failed": 0,
      "skipped": 0
    },
    "_clusters": {    
      "total" : 3,
      "successful" : 0,
      "skipped": 0,
      "running": 3,
      "partial": 0,
      "failed": 0,
      "details": {
        "(local)": {
          "status": "running",
          "indices": "my-index-000001",
          "timed_out": false
        },
        "cluster_one": {
          "status": "running",
          "indices": "my-index-000001",
          "timed_out": false
        },
        "cluster_one": {
          "status": "running",
          "indices": "my-index-000001",
          "timed_out": false
        }
      }
    },
    "hits": {
      "total" : {
          "value": 0,
          "relation": "eq"
      },
      "max_score": null,
      "hits": []
    }
  }
}

The async search id.

When ccs_minimize_roundtrips = true and searches on the remote clusters are still running, this section indicates the number of shards in scope for the local cluster only. This will be updated to include the total number of shards across all clusters only when the search is completed. When ccs_minimize_roundtrips= false, the total shard count is known up front and will be correct.

The _clusters section indicates that 3 clusters are in scope for the search and all are currently in the "running" state.

If you query the get async search endpoint while the query is still running, you will see an update in the _clusters and _shards section of the response when the local search has finished.

response = client.async_search.get(
  id: 'FklQYndoTDJ2VEFlMEVBTzFJMGhJVFEaLVlKYndBWWZSMUdicUc4WVlEaFl4ZzoxNTU='
)
puts response
GET /_async_search/FklQYndoTDJ2VEFlMEVBTzFJMGhJVFEaLVlKYndBWWZSMUdicUc4WVlEaFl4ZzoxNTU=

Response:

{
  "id": "FklQYndoTDJ2VEFlMEVBTzFJMGhJVFEaLVlKYndBWWZSMUdicUc4WVlEaFl4ZzoxNTU=",
  "is_partial": true,
  "is_running": true,
  "start_time_in_millis": 1685564911108,
  "expiration_time_in_millis": 1685996911108,
  "response": {
    "took": 11164,
    "timed_out": false,
    "terminated_early": false,
    "_shards": {
      "total": 10,
      "successful": 10,  
      "skipped": 0,
      "failed": 0
    },
    "_clusters": {
      "total": 3,
      "successful": 1,  
      "skipped": 0,
      "running": 2,
      "partial": 0,
      "failed": 0,
      "details": {
        "(local)": {
          "status": "successful",
          "indices": "my-index-000001",
          "took": 2034,
          "timed_out": false,
          "_shards": {
            "total": 10,
            "successful": 10,
            "skipped": 0,
            "failed": 0
          }
        },
        "cluster_one": {
          "status": "running",
          "indices": "my-index-000001",
          "timed_out": false
        },
        "cluster_two": {
          "status": "running",
          "indices": "my-index-000001",
          "timed_out": false
        }
      }
    },
    "hits": {
      "total": {
        "value": 167,  
        "relation": "eq"
      },
      "max_score": null,
      "hits": []
    }
  }
}

All the local cluster shards have completed.

The local cluster search has completed, so the "successful" clusters entry is set to 1 and "running" clusters entry reduced to 2. The _clusters response metadata will be updated as each cluster finishes.

Number of hits from the local cluster search. Final hits are not shown until searches on all clusters have been completed and merged.

After searches on all the clusters have completed, querying the get async search endpoint will show the final status of the _clusters and _shards section as well as the hits and any aggregation results.

response = client.async_search.get(
  id: 'FklQYndoTDJ2VEFlMEVBTzFJMGhJVFEaLVlKYndBWWZSMUdicUc4WVlEaFl4ZzoxNTU='
)
puts response
GET /_async_search/FklQYndoTDJ2VEFlMEVBTzFJMGhJVFEaLVlKYndBWWZSMUdicUc4WVlEaFl4ZzoxNTU=

Response:

{
  "id": "FklQYndoTDJ2VEFlMEVBTzFJMGhJVFEaLVlKYndBWWZSMUdicUc4WVlEaFl4ZzoxNTU=",
  "is_partial": false,
  "is_running": false,
  "start_time_in_millis": 1685564911108,
  "expiration_time_in_millis": 1685996911108,
  "completion_time_in_millis": 1685564938727,  
  "response": {
    "took": 27619,
    "timed_out": false,
    "num_reduce_phases": 4,
    "_shards": {
      "total": 28,
      "successful": 28,  
      "skipped": 0,
      "failed": 0
    },
    "_clusters": {
      "total": 3,
      "successful": 3,   
      "skipped": 0,
      "running": 0,
      "partial": 0,
      "failed": 0,
      "details": {
        "(local)": {
          "status": "successful",
          "indices": "my-index-000001",
          "took": 14382,
          "timed_out": false,
          "_shards": {
            "total": 10,
            "successful": 10,
            "skipped": 0,
            "failed": 0
          }
        },
        "cluster_one": {
          "status": "successful",
          "indices": "my-index-000001",
          "took": 22193,
          "timed_out": false,
          "_shards": {
            "total": 12,
            "successful": 12,
            "skipped": 0,
            "failed": 0
          }
        },
        "cluster_two": {
          "status": "successful",
          "indices": "my-index-000001",
          "took": 27550,
          "timed_out": false,
          "_shards": {
            "total": 6,
            "successful": 6,
            "skipped": 0,
            "failed": 0
          }
        }
      }
    },
    "hits": {
      "total": {
        "value": 1067,
        "relation": "eq"
      },
      "max_score": 1.8293576,
      "hits": [...list of hits here...]
    }
  }
}

Once the search has finished, the completion_time is present.

The _shards section is now updated to show that 28 total shards were searched across all clusters and that all were successful.

The _clusters section shows that searches on all 3 clusters were successful.

Cross-cluster search failuresedit

Failures during a cross-cluster search can result in one of two conditions:

  1. partial results (2xx HTTP status code)
  2. a failed search (4xx or 5xx HTTP status code)

Failure details will be present in the search response in both cases.

A search will be failed if a cluster marked with skip_unavailable=false is unavailable, disconnects during the search, or has search failures on all shards. In all other cases, failures will result in partial results.

Search failures on individual shards will be present in both the _shards section and the _clusters section of the response.

A failed search will have an additional top-level errors entry in the response.

Here is an example of a search with partial results due to a failure on one shard of one cluster. The search would be similar to ones shown previously. The _async_search/status endpoint is used here to show the completion status and not show the hits.

response = client.async_search.status(
  id: 'FmpwbThueVB4UkRDeUxqb1l4akIza3cbWEJyeVBPQldTV3FGZGdIeUVabXBldzoyMDIw'
)
puts response
GET /_async_search/status/FmpwbThueVB4UkRDeUxqb1l4akIza3cbWEJyeVBPQldTV3FGZGdIeUVabXBldzoyMDIw

Response:

{
  "id": "FmpwbThueVB4UkRDeUxqb1l4akIza3cbWEJyeVBPQldTV3FGZGdIeUVabXBldzoyMDIw",
  "is_partial": true,  
  "is_running": false,
  "start_time_in_millis": 1692106901478,
  "expiration_time_in_millis": 1692538901478,
  "completion_time_in_millis": 1692106903547,
  "response": {
    "took": 2069,
    "timed_out": false,
    "num_reduce_phases": 4,
    "_shards": {
      "total": 28,
      "successful": 27,
      "skipped": 0,
      "failed": 1,
      "failures": [   
        {
          "shard": 1,
          "index": "cluster_two:my-index-000001",
          "node": "LMpUnAu0QEeCUMfg_56sAg",
          "reason": {
            "type": "query_shard_exception",
            "reason": "failed to create query: [my-index-000001][1] exception message here",
            "index_uuid": "4F2VWx8RQSeIhUE-nksvCQ",
            "index": "cluster_two:my-index-000001",
            "caused_by": {
              "type": "runtime_exception",
              "reason": "runtime_exception: [my-index-000001][1] exception message here"
            }
          }
        }
      ]
    },
    "_clusters": {
      "total": 3,
      "successful": 2,
      "skipped": 0,
      "running": 0,
      "partial": 1,   
      "failed": 0,
      "details": {
        "(local)": {
          "status": "successful",
          "indices": "my-index-000001",
          "took": 1753,
          "timed_out": false,
          "_shards": {
            "total": 10,
            "successful": 10,
            "skipped": 0,
            "failed": 0
          }
        },
        "cluster_one": {
          "status": "successful",
          "indices": "my-index-000001",
          "took": 2054,
          "timed_out": false,
          "_shards": {
            "total": 12,
            "successful": 12,
            "skipped": 0,
            "failed": 0
          }
        },
        "cluster_two": {
          "status": "partial",   
          "indices": "my-index-000001",
          "took": 2039,
          "timed_out": false,
          "_shards": {
            "total": 6,
            "successful": 5,
            "skipped": 0,
            "failed": 1   
          },
          "failures": [  
            {
              "shard": 1,
              "index": "cluster_two:my-index-000001",
              "node": "LMpUnAu0QEeCUMfg_56sAg",
              "reason": {
                "type": "query_shard_exception",
                "reason": "failed to create query: [my-index-000001][1] exception message here",
                "index_uuid": "4F2VWx8RQSeIhUE-nksvCQ",
                "index": "cluster_two:my-index-000001",
                "caused_by": {
                  "type": "runtime_exception",
                  "reason": "runtime_exception: [my-index-000001][1] exception message here"
                }
              }
            }
          ]
        }
      }
    },
    "hits": {
    }
  }
}

The search results are marked as partial, since at least one shard search failed.

The _shards section includes shard failure info.

Clusters that have partial results are still marked as "partial". They are marked with status "skipped" (or "failed") only if no data was returned from the search.

The partial status has been applied to the cluster with partial results.

The failed shard count is shown.

The shard failures are listed under the cluster/details entry also.

Here is an example where both cluster_one and cluster_two lost connectivity during a cross-cluster search. Since cluster_one is marked as skip_unavailable=true, its status is skipped and since cluster_two is marked as skip_unavailable=false, its status is failed. Since there was a failed cluster, a top level error is also present and this returns an HTTP status of 500 (not shown).

If you want the search to still return results even when a cluster is unavailable, set skip_unavailable=true for all the remote clusters.

response = client.async_search.get(
  id: 'FjktRGJ1Y2w1U0phLTRhZnVyeUZ2MVEbWEJyeVBPQldTV3FGZGdIeUVabXBldzo5NzA4'
)
puts response
GET /_async_search/FjktRGJ1Y2w1U0phLTRhZnVyeUZ2MVEbWEJyeVBPQldTV3FGZGdIeUVabXBldzo5NzA4

Response:

{
  "id": "FjktRGJ1Y2w1U0phLTRhZnVyeUZ2MVEbWEJyeVBPQldTV3FGZGdIeUVabXBldzo5NzA4",
  "is_partial": true,
  "is_running": false,
  "start_time_in_millis": 1692112102650,
  "expiration_time_in_millis": 1692544102650,
  "completion_time_in_millis": 1692112106177,
  "response": {
    "took": 3527,
    "timed_out": false,
    "terminated_early": false,
    "_shards": {
      "total": 10,   
      "successful": 10,
      "skipped": 0,
      "failed": 0
    },
    "_clusters": {
      "total": 3,
      "successful": 1,
      "skipped": 1,
      "running": 0,
      "partial": 0,
      "failed": 1,
      "details": {
        "(local)": {
          "status": "successful",
          "indices": "my-index-000001",
          "took": 1473,
          "timed_out": false,
          "_shards": {
            "total": 10,
            "successful": 10,
            "skipped": 0,
            "failed": 0
          }
        },
        "cluster_one": {
          "status": "skipped",   
          "indices": "my-index-000001",
          "timed_out": false,
          "failures": [
            {
              "shard": -1,
              "index": null,
              "reason": {
                "type": "node_disconnected_exception",   
                "reason": "[myhostname1][35.238.149.1:9300][indices:data/read/search] disconnected"
              }
            }
          ]
        },
        "cluster_two": {
          "status": "failed",   
          "indices": "my-index-000001",
          "timed_out": false,
          "failures": [
            {
              "shard": -1,
              "index": null,
              "reason": {
                "type": "node_disconnected_exception",
                "reason": "[myhostname2][35.238.149.2:9300][indices:data/read/search] disconnected"
              }
            }
          ]
        }
      }
    },
    "hits": {
    },
  }
  "error": {  
    "type": "status_exception",
    "reason": "error while executing search",
    "caused_by": {
      "type": "node_disconnected_exception",
      "reason": "[myhostname2][35.238.149.2:9300][indices:data/read/search] disconnected"
    }
  }
}

The shard accounting will often be only partial when errors like this occur, since we need to be able to get shard info from remote clusters on each search.

cluster_one disconnected during the search and it returned no results. Since it is marked in the remote cluster configuration as skip_unavailable=true, its status is "skipped", which will not fail the entire search.

The failures list shows that the remote cluster node disconnected from the querying cluster.

cluster_two status is "failed", since it is marked in the remote cluster configuration as skip_unavailable=false.

A top level error entry is included when there is a "failed" cluster.

Excluding clusters or indices from a cross-cluster searchedit

If you use a wildcard to include a large list of clusters and/or indices, you can explicitly exclude one or more clusters or indices with a - minus sign in front of the cluster or index.

To exclude an entire cluster, you would put the minus sign in front of the cluster alias, such as: -mycluster:*. When excluding a cluster, you must use * in the index position or an error will be returned.

To exclude a specific remote index, you would put the minus sign in front of the index, such as mycluster:-myindex.

Exclude a remote cluster

Here’s how you would exclude cluster_three from a cross-cluster search that uses a wildcard to specify a list of clusters:

POST /my-index-000001,cluster*:my-index-000001,-cluster_three:*/_async_search  
{
  "query": {
    "match": {
      "user.id": "kimchy"
    }
  },
  "_source": ["user.id", "message", "http.response.status_code"]
}

The cluster* notation would naturally include cluster_one, cluster_two and cluster_three. To exclude cluster_three use a - before the cluster name along with a simple wildcard * in the index position. This indicates that you do not want the search to make any contact with cluster_three.

Exclude a remote index

Suppose you want to search all indices matching my-index-* but you want to exclude my-index-000001 on cluster_three. Here’s how you could do that:

POST /my-index-000001,cluster*:my-index-*,cluster_three:-my-index-000001/_async_search  
{
  "query": {
    "match": {
      "user.id": "kimchy"
    }
  },
  "_source": ["user.id", "message", "http.response.status_code"]
}

This will not exclude cluster_three from the search. It will still be contacted and told to search any indexes matching my-index-* except for my-index-000001.

Using async search for cross-cluster search with ccs_minimize_roundtrips=falseedit

The _shards and _clusters section of the response behave differently when ccs_minimize_roundtrips is false.

Key differences are:

  1. The _shards section total count will be accurate immediately as the total number of shards is gathered from all clusters before the search starts.
  2. The _shards section will be incrementally updated as searches on individual shards complete, so you will get a more accurate accounting of progress during a long-running search compared to when minimize roundtrips is used.
  3. The _cluster section starts off listing all of its shard counts, since they are also obtained before the query phase begins.

Example using the same setup as in the previous section (ccs_minimize_roundtrips=true):

POST /my-index-000001,cluster_one:my-index-000001,cluster_two:my-index-000001/_async_search?ccs_minimize_roundtrips=false
{
  "query": {
    "match": {
      "user.id": "kimchy"
    }
  },
  "_source": ["user.id", "message", "http.response.status_code"]
}

The API returns the following response if the query takes longer than the wait_for_completion_timeout duration (see Async search).

{
  "id": "FklQYndoTDJ2VEFlMEVBTzFJMGhJVFEaLVlKYndBWWZSMUdicUc4WVlEaFl4ZzoxNTU=",
  "is_partial": true,
  "is_running": true,
  "start_time_in_millis": 1685563581380,
  "expiration_time_in_millis": 1685995581380,
  "response": {
    "took": 1020,
    "timed_out": false,
    "_shards": {
      "total": 28,     
      "successful": 0,
      "failed": 0,
      "skipped": 0
    },
    "_clusters": {
      "total" : 3,
      "successful": 0,
      "skipped": 0,
      "running": 3,    
      "partial": 0,
      "failed": 0,
      "details": {    
        "(local)": {
          "status": "running",
          "indices": "my-index-000001",
          "timed_out": false,
          "_shards": {
            "total": 10,
            "successful": 0,
            "skipped": 0,
            "failed": 0
          }
        },
        "cluster_one": {
          "status": "running",
          "indices": "my-index-000001",
          "timed_out": false,
          "_shards": {
            "total": 12,
            "successful": 0,
            "skipped": 0,
            "failed": 0
          }
        },
        "cluster_two": {
          "status": "running",
          "indices": "my-index-000001",
          "timed_out": false,
          "_shards": {
            "total": 6,
            "successful": 0,
            "skipped": 0,
            "failed": 0
          }
        }
      }
    },
    "hits": {
      "total" : {
          "value": 0,
          "relation": "eq"
      },
      "max_score": null,
      "hits": []
    }
  }
}

All shards from all clusters in scope for the search are listed here. Watch this section and/or the _clusters section for updates to monitor search progress.

From the _clusters section we can see that all the clusters are in "running" state.

The _clusters section shows that shard information was successfully gathered from all 3 clusters and the total shard count on each cluster is listed.

Optional remote clustersedit

By default, a cross-cluster search fails if a remote cluster in the request is unavailable or returns an error where the search on all shards failed. Use the skip_unavailable cluster setting to mark a specific remote cluster as optional for cross-cluster search.

If skip_unavailable is true, a cross-cluster search:

  • Skips the remote cluster if its nodes are unavailable during the search. The response’s _clusters.skipped value contains a count of any skipped clusters and the _clusters.details section of the response will show a skipped status.
  • Ignores errors returned by the remote cluster, such as errors related to unavailable shards or indices. This can include errors related to search parameters such as allow_no_indices and ignore_unavailable.
  • Ignores the allow_partial_search_results parameter and the related search.default_allow_partial_results cluster setting when searching the remote cluster. This means searches on the remote cluster may return partial results.

The following cluster update settings API request changes skip_unavailable setting to true for cluster_two.

response = client.cluster.put_settings(
  body: {
    persistent: {
      'cluster.remote.cluster_two.skip_unavailable' => true
    }
  }
)
puts response
PUT _cluster/settings
{
  "persistent": {
    "cluster.remote.cluster_two.skip_unavailable": true
  }
}

If cluster_two is disconnected or unavailable during a cross-cluster search, Elasticsearch won’t include matching documents from that cluster in the final results. If at least one shard provides results, those results will be used and the search will return partial data. (If doing cross-cluster search using async search, the is_partial field will be set to true to indicate partial results.)

How cross-cluster search handles network delaysedit

Because cross-cluster search involves sending requests to remote clusters, any network delays can impact search speed. To avoid slow searches, cross-cluster search offers two options for handling network delays:

Minimize network roundtrips

By default, Elasticsearch reduces the number of network roundtrips between remote clusters. This reduces the impact of network delays on search speed. However, Elasticsearch can’t reduce network roundtrips for large search requests, such as those including a scroll or inner hits.

See Considerations for choosing whether to minimize roundtrips in a cross-cluster search to learn how this option works.

Don’t minimize network roundtrips

For search requests that include a scroll or inner hits, Elasticsearch sends multiple outgoing and ingoing requests to each remote cluster. You can also choose this option by setting the ccs_minimize_roundtrips parameter to false. While typically slower, this approach may work well for networks with low latency.

See Don’t minimize network roundtrips to learn how this option works.

The vector tile search API always minimizes network roundtrips and doesn’t include the ccs_minimize_roundtrips parameter.

The Approximate kNN search doesn’t support minimizing network roundtrips, and sets the parameter ccs_minimize_roundtrips to false.

Considerations for choosing whether to minimize roundtrips in a cross-cluster searchedit

For cross-cluster searches that query a large number of shards, the minimize roundtrips option typically provides much better performance. This is especially true if the clusters being searched have high network latency (e.g., distant geographic regions).

However, not minimizing roundtrips allows you to get back incremental results of any aggregations in your query when using async-search while the search is still running.

By default, synchronous searches minimize roundtrips, while asynchronous searches do not. You can override the default by using the ccs_minimize_roundtrips parameter, setting it to either true or false, as shown in several examples earlier in this document.

Minimize network roundtripsedit

Here’s how cross-cluster search works when you minimize network roundtrips.

  1. You send a cross-cluster search request to your local cluster. A coordinating node in that cluster receives and parses the request.

    ccs min roundtrip client request

  2. The coordinating node sends a single search request to each cluster, including the local cluster. Each cluster performs the search request independently, applying its own cluster-level settings to the request.

    ccs min roundtrip cluster search

  3. Each remote cluster sends its search results back to the coordinating node.

    ccs min roundtrip cluster results

  4. After collecting results from each cluster, the coordinating node returns the final results in the cross-cluster search response.

    ccs min roundtrip client response

Don’t minimize network roundtripsedit

Here’s how cross-cluster search works when you don’t minimize network roundtrips.

  1. You send a cross-cluster search request to your local cluster. A coordinating node in that cluster receives and parses the request.

    ccs min roundtrip client request

  2. The coordinating node sends a "search shards" transport layer request to each remote cluster to have them to perform a "can match" search to determine which shards on each cluster should be searched.

    ccs min roundtrip cluster search

  3. Each remote cluster sends its response back to the coordinating node. This response contains information about the indices and shards the cross-cluster search request will be executed on.

    ccs min roundtrip cluster results

  4. The coordinating node sends a search request to each shard, including those in its own cluster. Each shard performs the search request independently.

    When network roundtrips aren’t minimized, the search is executed as if all data were in the coordinating node’s cluster. We recommend updating cluster-level settings that limit searches, such as action.search.shard_count.limit, pre_filter_shard_size, and max_concurrent_shard_requests, to account for this. If these limits are too low, the search may be rejected.

    ccs dont min roundtrip shard search

  5. Each shard sends its search results back to the coordinating node.

    ccs dont min roundtrip shard results

  6. After collecting results from each cluster, the coordinating node returns the final results in the cross-cluster search response.

    ccs min roundtrip client response

Supported cross-cluster search configurationsedit

In 8.0+, Elastic supports searches from a local cluster to a remote cluster running:

  • The previous minor version.
  • The same version.
  • A newer minor version in the same major version.

Elastic also supports searches from a local cluster running the last minor version of a major version to a remote cluster running any minor version in the following major version. For example, a local 7.17 cluster can search any remote 8.x cluster.

Remote cluster version

Local cluster version

6.8

7.1–7.16

7.17

8.0

8.1

8.2

8.3

8.4

8.5

8.6

8.7

8.8

8.9

8.10

8.11

8.12

6.8

Yes

Yes

Yes

No

No

No

No

No

No

No

No

No

No

No

No

No

7.1–7.16

Yes

Yes

Yes

No

No

No

No

No

No

No

No

No

No

No

No

No

7.17

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

8.0

No

No

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

8.1

No

No

No

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

8.2

No

No

No

No

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

8.3

No

No

No

No

No

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

8.4

No

No

No

No

No

No

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

8.5

No

No

No

No

No

No

No

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

8.6

No

No

No

No

No

No

No

No

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

8.7

No

No

No

No

No

No

No

No

No

Yes

Yes

Yes

Yes

Yes

Yes

Yes

8.8

No

No

No

No

No

No

No

No

No

No

Yes

Yes

Yes

Yes

Yes

Yes

8.9

No

No

No

No

No

No

No

No

No

No

No

Yes

Yes

Yes

Yes

Yes

8.10

No

No

No

No

No

No

No

No

No

No

No

No

Yes

Yes

Yes

Yes

8.11

No

No

No

No

No

No

No

No

No

No

No

No

No

Yes

Yes

Yes

8.12

No

No

No

No

No

No

No

No

No

No

No

No

No

No

Yes

Yes

For the EQL search API, the local and remote clusters must use the same Elasticsearch version if they have versions prior to 7.17.7 (included) or prior to 8.5.1 (included).

For example, a local 8.0 cluster can search a remote 7.17 or any remote 8.x cluster. However, a search from a local 8.0 cluster to a remote 7.16 or 6.8 cluster is not supported.

Only features that exist across all searched clusters are supported. Using a feature with a remote cluster where the feature is not supported will result in undefined behavior.

A cross-cluster search using an unsupported configuration may still work. However, such searches aren’t tested by Elastic, and their behavior isn’t guaranteed.

Ensure cross-cluster search supportedit

The simplest way to ensure your clusters support cross-cluster search is to keep each cluster on the same version of Elasticsearch. If you need to maintain clusters with different versions, you can:

  • Maintain a dedicated cluster for cross-cluster search. Keep this cluster on the earliest version needed to search the other clusters. For example, if you have 7.17 and 8.x clusters, you can maintain a dedicated 7.17 cluster to use as the local cluster for cross-cluster search.
  • Keep each cluster no more than one minor version apart. This lets you use any cluster as the local cluster when running a cross-cluster search.

Cross-cluster search during an upgradeedit

You can still search a remote cluster while performing a rolling upgrade on the local cluster. However, the local coordinating node’s "upgrade from" and "upgrade to" version must be compatible with the remote cluster’s gateway node.

Running multiple versions of Elasticsearch in the same cluster beyond the duration of an upgrade is not supported.

For more information about upgrades, see Upgrading Elasticsearch.