Removal of mapping types

Since the first release of Elasticsearch, each document has been stored in a single index and assigned a single mapping type. A mapping type was used to represent the type of document or entity being indexed, for instance a twitter index might have a user type and a tweet type.

Each mapping type could have its own fields, so the user type might have a full_name field, a user_name field, and an email field, while the tweet type could have a content field, a tweeted_at field and, like the user type, a user_name field.

Each document had a _type metadata field containing the type name, and searches could be limited to one or more types by specifying the type name(s) in the URL:

GET twitter/user,tweet/_search
{
  "query": {
    "match": {
      "user_name": "kimchy"
    }
  }
}

The _type field was combined with the document’s _id to generate a _uid field, so documents of different types with the same _id could exist in a single index.

Mapping types were also used to establish a parent-child relationship between documents, so documents of type question could be parents to documents of type answer.

Initially, we spoke about an “index” being similar to a “database” in an SQL database, and a “type” being equivalent to a “table”.

This was a bad analogy that led to incorrect assumptions. In an SQL database, tables are independent of each other. The columns in one table have no bearing on columns with the same name in another table. This is not the case for fields in a mapping type.

In an Elasticsearch index, fields that have the same name in different mapping types are backed by the same Lucene field internally. In other words, using the example above, the user_name field in the user type is stored in exactly the same field as the user_name field in the tweet type, and both user_name fields must have the same mapping (definition) in both types.

This can lead to frustration when, for example, you want deleted to be a date field in one type and a boolean field in another type in the same index.

On top of that, storing different entities that have few or no fields in common in the same index leads to sparse data and interferes with Lucene’s ability to compress documents efficiently.

For these reasons, we have decided to remove the concept of mapping types from Elasticsearch.

The first alternative is to have an index per document type. Instead of storing tweets and users in a single twitter index, you could store tweets in the tweets index and users in the user index. Indices are completely independent of each other and so there will be no conflict of field types between indices.

This approach has two benefits:

Each index can be sized appropriately for the number of documents it will contain: you can use a smaller number of primary shards for users and a larger number of primary shards for tweets.

Of course, there is a limit to how many primary shards can exist in a cluster so you may not want to waste an entire shard for a collection of only a few thousand documents. In this case, you can implement your own custom type field which will work in a similar way to the old _type.

Let’s take the user/tweet example above. Originally, the workflow would have looked something like this:

PUT twitter
{
  "mappings": {
    "user": {
      "properties": {
        "name": { "type": "text" },
        "user_name": { "type": "keyword" },
        "email": { "type": "keyword" }
      }
    },
    "tweet": {
      "properties": {
        "content": { "type": "text" },
        "user_name": { "type": "keyword" },
        "tweeted_at": { "type": "date" }
      }
    }
  }
}

PUT twitter/user/kimchy
{
  "name": "Shay Banon",
  "user_name": "kimchy",
  "email": "shay@kimchy.com"
}

PUT twitter/tweet/1
{
  "user_name": "kimchy",
  "tweeted_at": "2017-10-24T09:00:00Z",
  "content": "Types are going away"
}

GET twitter/tweet/_search
{
  "query": {
    "match": {
      "user_name": "kimchy"
    }
  }
}

You can achieve the same thing by adding a custom type field as follows:

PUT twitter
{
  "mappings": {
    "_doc": {
      "properties": {
        "type": { "type": "keyword" }, 
        "name": { "type": "text" },
        "user_name": { "type": "keyword" },
        "email": { "type": "keyword" },
        "content": { "type": "text" },
        "tweeted_at": { "type": "date" }
      }
    }
  }
}

PUT twitter/_doc/user-kimchy
{
  "type": "user", 
  "name": "Shay Banon",
  "user_name": "kimchy",
  "email": "shay@kimchy.com"
}

PUT twitter/_doc/tweet-1
{
  "type": "tweet", 
  "user_name": "kimchy",
  "tweeted_at": "2017-10-24T09:00:00Z",
  "content": "Types are going away"
}

GET twitter/_search
{
  "query": {
    "bool": {
      "must": {
        "match": {
          "user_name": "kimchy"
        }
      },
      "filter": {
        "match": {
          "type": "tweet" 
        }
      }
    }
  }
}

Previously, a parent-child relationship was represented by making one mapping type the parent, and one or more other mapping types the children. Without types, we can no longer use this syntax. The parent-child feature will continue to function as before, except that the way of expressing the relationship between documents has been changed to use the new join field.

This is a big change for our users, so we have tried to make it as painless as possible. The change will roll out as follows:

The Reindex API can be used to convert multi-type indices to single-type indices. The following examples can be used in Elasticsearch 5.6 or Elasticsearch 6.x. In 6.x, there is no need to specify index.mapping.single_type as that is the default.

This first example splits our twitter index into a tweets index and a users index:

PUT users
{
  "settings": {
    "index.mapping.single_type": true
  },
  "mappings": {
    "_doc": {
      "properties": {
        "name": {
          "type": "text"
        },
        "user_name": {
          "type": "keyword"
        },
        "email": {
          "type": "keyword"
        }
      }
    }
  }
}

PUT tweets
{
  "settings": {
    "index.mapping.single_type": true
  },
  "mappings": {
    "_doc": {
      "properties": {
        "content": {
          "type": "text"
        },
        "user_name": {
          "type": "keyword"
        },
        "tweeted_at": {
          "type": "date"
        }
      }
    }
  }
}

POST _reindex
{
  "source": {
    "index": "twitter",
    "type": "user"
  },
  "dest": {
    "index": "users",
    "type": "_doc"
  }
}

POST _reindex
{
  "source": {
    "index": "twitter",
    "type": "tweet"
  },
  "dest": {
    "index": "tweets",
    "type": "_doc"
  }
}

This next example adds a custom type field and sets it to the value of the original _type. It also adds the type to the _id in case there are any documents of different types which have conflicting IDs:

PUT new_twitter
{
  "mappings": {
    "_doc": {
      "properties": {
        "type": {
          "type": "keyword"
        },
        "name": {
          "type": "text"
        },
        "user_name": {
          "type": "keyword"
        },
        "email": {
          "type": "keyword"
        },
        "content": {
          "type": "text"
        },
        "tweeted_at": {
          "type": "date"
        }
      }
    }
  }
}


POST _reindex
{
  "source": {
    "index": "twitter"
  },
  "dest": {
    "index": "new_twitter"
  },
  "script": {
    "source": """
      ctx._source.type = ctx._type;
      ctx._id = ctx._type + '-' + ctx._id;
      ctx._type = '_doc';
    """
  }
}

In Elasticsearch 7.0, each API will support typeless requests, and specifying a type will produce a deprecation warning.

Index creation, index template, and mapping APIs support a new include_type_name URL parameter that specifies whether mapping definitions in requests and responses should contain the type name. The parameter defaults to true in version 6.8 to match the pre-7.0 behavior of using type names in mappings. It defaults to false in version 7.0 and will be removed in version 8.0.

It should be set explicitly in 6.8 to prepare to upgrade to 7.0. To avoid deprecation warnings in 6.8, the parameter can be set to either true or false. In 7.0, setting include_type_name at all will result in a deprecation warning.

See some examples of interactions with Elasticsearch with this option set to false:

PUT index?include_type_name=false
{
  "mappings": {
    "properties": { 
      "foo": {
        "type": "keyword"
      }
    }
  }
}

PUT index/_mappings?include_type_name=false
{
  "properties": { 
    "bar": {
      "type": "text"
    }
  }
}

GET index/_mappings?include_type_name=false

The above call returns

{
  "index": {
    "mappings": {
      "properties": { 
        "foo": {
          "type": "keyword"
        },
        "bar": {
          "type": "text"
        }
      }
    }
  }
}

In 7.0, index APIs must be called with the {index}/_doc path for automatic generation of the _id and {index}/_doc/{id} with explicit ids.

PUT index/_doc/1
{
  "foo": "baz"
}

{
  "_index": "index",
  "_id": "1",
  "_type": "_doc",
  "_version": 1,
  "result": "created",
  "_shards": {
    "total": 2,
    "successful": 1,
    "failed": 0
  },
  "_seq_no": 0,
  "_primary_term": 1
}

Similarly, the get and delete APIs use the path {index}/_doc/{id}:

GET index/_doc/1

For API paths that contain both a type and endpoint name like _update, in 7.0 the endpoint will immediately follow the index name:

POST index/_update/1
{
  "doc" : {
    "foo" : "qux"
  }
}

GET /index/_source/1

Types should also no longer appear in the body of requests. The following example of bulk indexing omits the type both in the URL, and in the individual bulk commands:

POST _bulk
{ "index" : { "_index" : "index", "_id" : "3" } }
{ "foo" : "baz" }
{ "index" : { "_index" : "index", "_id" : "4" } }
{ "foo" : "qux" }

When calling a search API such _search, _msearch, or _explain, types should not be included in the URL. Additionally, the _type field should not be used in queries, aggregations, or scripts.

The document and search APIs will continue to return a _type key in responses, to avoid breaks to response parsing. However, the key is considered deprecated and should no longer be referenced. Types will be completely removed from responses in 8.0.

Note that when a deprecated typed API is used, the index’s mapping type will be returned as normal, but that typeless APIs will return the dummy type _doc in the response. For example, the following typeless get call will always return _doc as the type, even if the mapping has a custom type name like my_type:

PUT index/my_type/1
{
  "foo": "baz"
}

GET index/_doc/1

{
  "_index" : "index",
  "_type" : "_doc",
  "_id" : "1",
  "_version" : 1,
  "_seq_no" : 0,
  "_primary_term" : 1,
  "found": true,
  "_source" : {
    "foo" : "baz"
  }
}

It is recommended to make index templates typeless by re-adding them with include_type_name set to false. Under the hood, typeless templates will use the dummy type _doc when creating indices.

In case typeless templates are used with typed index creation calls or typed templates are used with typeless index creation calls, the template will still be applied but the index creation call decides whether there should be a type or not. For instance in the below example, index-1-01 will have a type in spite of the fact that it matches a template that is typeless, and index-2-01 will be typeless in spite of the fact that it matches a template that defines a type. Both index-1-01 and index-2-01 will inherit the foo field from the template that they match.

PUT _template/template1
{
  "index_patterns":[ "index-1-*" ],
  "mappings": {
    "properties": {
      "foo": {
        "type": "keyword"
      }
    }
  }
}

PUT _template/template2?include_type_name=true
{
  "index_patterns":[ "index-2-*" ],
  "mappings": {
    "type": {
      "properties": {
        "foo": {
          "type": "keyword"
        }
      }
    }
  }
}

PUT index-1-01?include_type_name=true
{
  "mappings": {
    "type": {
      "properties": {
        "bar": {
          "type": "long"
        }
      }
    }
  }
}

PUT index-2-01
{
  "mappings": {
    "properties": {
      "bar": {
        "type": "long"
      }
    }
  }
}

In case of implicit index creation, because of documents that get indexed in an index that doesn’t exist yet, the template is always honored. This is usually not a problem due to the fact that typeless index calls work on typed indices.

In a cluster composed of both 6.8 and 7.0 nodes, the parameter include_type_name should be specified in index APIs like index creation. This is because the parameter has a different default between 6.8 and 7.0, so the same mapping definition will not be valid for both node versions.

Typeless document APIs such as bulk and update are only available as of 7.0, and will not work with 6.8 nodes. This also holds true for the typeless versions of queries that perform document lookups, such as terms.