Join field type

The join data type is a special field that creates parent/child relation within documents of the same index. The relations section defines a set of possible relations within the documents, each relation being a parent name and a child name. A parent/child relation can be defined as follows:

  1. PUT my-index-000001
  2. {
  3.   "mappings": {
  4.     "properties": {
  5.       "my_id": {
  6.         "type": "keyword"
  7.       },
  8.       "my_join_field": { 
  9.         "type": "join",
  10.         "relations": {
  11.           "question": "answer" 
  12.         }
  13.       }
  14.     }
  15.   }
  16. }

The name for the field

Defines a single relation where question is parent of answer.

To index a document with a join, the name of the relation and the optional parent of the document must be provided in the source. For instance the following example creates two parent documents in the question context:

  1. PUT my-index-000001/_doc/1?refresh
  2. {
  3.   "my_id": "1",
  4.   "text": "This is a question",
  5.   "my_join_field": {
  6.     "name": "question" 
  7.   }
  8. }
  10. PUT my-index-000001/_doc/2?refresh
  11. {
  12.   "my_id": "2",
  13.   "text": "This is another question",
  14.   "my_join_field": {
  15.     "name": "question"
  16.   }
  17. }

This document is a question document.

When indexing parent documents, you can choose to specify just the name of the relation as a shortcut instead of encapsulating it in the normal object notation:

  1. PUT my-index-000001/_doc/1?refresh
  2. {
  3.   "my_id": "1",
  4.   "text": "This is a question",
  5.   "my_join_field": "question" 
  6. }
  8. PUT my-index-000001/_doc/2?refresh
  9. {
  10.   "my_id": "2",
  11.   "text": "This is another question",
  12.   "my_join_field": "question"
  13. }

Simpler notation for a parent document just uses the relation name.

When indexing a child, the name of the relation as well as the parent id of the document must be added in the _source.

It is required to index the lineage of a parent in the same shard so you must always route child documents using their greater parent id.

For instance the following example shows how to index two child documents:

  1. PUT my-index-000001/_doc/3?routing=1&refresh 
  2. {
  3.   "my_id": "3",
  4.   "text": "This is an answer",
  5.   "my_join_field": {
  6.     "name": "answer", 
  7.     "parent": "1" 
  8.   }
  9. }
  11. PUT my-index-000001/_doc/4?routing=1&refresh
  12. {
  13.   "my_id": "4",
  14.   "text": "This is another answer",
  15.   "my_join_field": {
  16.     "name": "answer",
  17.     "parent": "1"
  18.   }
  19. }

The routing value is mandatory because parent and child documents must be indexed on the same shard

answer is the name of the join for this document

The parent id of this child document

Parent-join and performance.

The join field shouldn’t be used like joins in a relation database. In Elasticsearch the key to good performance is to de-normalize your data into documents. Each join field, has_child or has_parent query adds a significant tax to your query performance.

The only case where the join field makes sense is if your data contains a one-to-many relationship where one entity significantly outnumbers the other entity. An example of such case is a use case with products and offers for these products. In the case that offers significantly outnumbers the number of products then it makes sense to model the product as parent document and the offer as child document.

Parent-join restrictions

  • Only one join field mapping is allowed per index.
  • Parent and child documents must be indexed on the same shard. This means that the same routing value needs to be provided when getting, deleting, or updating a child document.
  • An element can have multiple children but only one parent.
  • It is possible to add a new relation to an existing join field.
  • It is also possible to add a child to an existing element but only if the element is already a parent.

Searching with parent-join

The parent-join creates one field to index the name of the relation within the document (my_parent, my_child, …​).

It also creates one field per parent/child relation. The name of this field is the name of the join field followed by # and the name of the parent in the relation. So for instance for the my_parent → [my_child, another_child] relation, the join field creates an additional field named my_join_field#my_parent.

This field contains the parent _id that the document links to if the document is a child (my_child or another_child) and the _id of document if it’s a parent (my_parent).

When searching an index that contains a join field, these two fields are always returned in the search response:

  1. GET my-index-000001/_search
  2. {
  3.   "query": {
  4.     "match_all": {}
  5.   },
  6.   "sort": ["my_id"]
  7. }

Will return:

  1. {
  2.   ...,
  3.   "hits": {
  4.     "total": {
  5.       "value": 4,
  6.       "relation": "eq"
  7.     },
  8.     "max_score": null,
  9.     "hits": [
  10.       {
  11.         "_index": "my-index-000001",
  12.         "_type": "_doc",
  13.         "_id": "1",
  14.         "_score": null,
  15.         "_source": {
  16.           "my_id": "1",
  17.           "text": "This is a question",
  18.           "my_join_field": "question"         
  19.         },
  20.         "sort": [
  21.           "1"
  22.         ]
  23.       },
  24.       {
  25.         "_index": "my-index-000001",
  26.         "_type": "_doc",
  27.         "_id": "2",
  28.         "_score": null,
  29.         "_source": {
  30.           "my_id": "2",
  31.           "text": "This is another question",
  32.           "my_join_field": "question"          
  33.         },
  34.         "sort": [
  35.           "2"
  36.         ]
  37.       },
  38.       {
  39.         "_index": "my-index-000001",
  40.         "_type": "_doc",
  41.         "_id": "3",
  42.         "_score": null,
  43.         "_routing": "1",
  44.         "_source": {
  45.           "my_id": "3",
  46.           "text": "This is an answer",
  47.           "my_join_field": {
  48.             "name": "answer",                 
  49.             "parent": "1"                     
  50.           }
  51.         },
  52.         "sort": [
  53.           "3"
  54.         ]
  55.       },
  56.       {
  57.         "_index": "my-index-000001",
  58.         "_type": "_doc",
  59.         "_id": "4",
  60.         "_score": null,
  61.         "_routing": "1",
  62.         "_source": {
  63.           "my_id": "4",
  64.           "text": "This is another answer",
  65.           "my_join_field": {
  66.             "name": "answer",
  67.             "parent": "1"
  68.           }
  69.         },
  70.         "sort": [
  71.           "4"
  72.         ]
  73.       }
  74.     ]
  75.   }
  76. }

This document belongs to the question join

This document belongs to the question join

This document belongs to the answer join

The linked parent id for the child document

Parent-join queries and aggregations

See the has_child and has_parent queries, the children aggregation, and inner hits for more information.

The value of the join field is accessible in aggregations and scripts, and may be queried with the parent_id query:

  1. GET my-index-000001/_search
  2. {
  3.   "query": {
  4.     "parent_id": { 
  5.       "type": "answer",
  6.       "id": "1"
  7.     }
  8.   },
  9.   "aggs": {
  10.     "parents": {
  11.       "terms": {
  12.         "field": "my_join_field#question", 
  13.         "size": 10
  14.       }
  15.     }
  16.   },
  17.   "script_fields": {
  18.     "parent": {
  19.       "script": {
  20.          "source": "doc['my_join_field#question']" 
  21.       }
  22.     }
  23.   }
  24. }

Querying the parent id field (also see the has_parent query and the has_child query)

Aggregating on the parent id field (also see the children aggregation)

Accessing the parent id` field in scripts

Global ordinals

The join field uses global ordinals to speed up joins. Global ordinals need to be rebuilt after any change to a shard. The more parent id values are stored in a shard, the longer it takes to rebuild the global ordinals for the join field.

Global ordinals, by default, are built eagerly: if the index has changed, global ordinals for the join field will be rebuilt as part of the refresh. This can add significant time to the refresh. However most of the times this is the right trade-off, otherwise global ordinals are rebuilt when the first parent-join query or aggregation is used. This can introduce a significant latency spike for your users and usually this is worse as multiple global ordinals for the join field may be attempt rebuilt within a single refresh interval when many writes are occurring.

When the join field is used infrequently and writes occur frequently it may make sense to disable eager loading:

  1. PUT my-index-000001
  2. {
  3.   "mappings": {
  4.     "properties": {
  5.       "my_join_field": {
  6.         "type": "join",
  7.         "relations": {
  8.            "question": "answer"
  9.         },
  10.         "eager_global_ordinals": false
  11.       }
  12.     }
  13.   }
  14. }

The amount of heap used by global ordinals can be checked per parent relation as follows:

  1. # Per-index
  2. GET _stats/fielddata?human&fields=my_join_field#question
  4. # Per-node per-index
  5. GET _nodes/stats/indices/fielddata?human&fields=my_join_field#question

Multiple children per parent

It is also possible to define multiple children for a single parent:

  1. PUT my-index-000001
  2. {
  3.   "mappings": {
  4.     "properties": {
  5.       "my_join_field": {
  6.         "type": "join",
  7.         "relations": {
  8.           "question": ["answer", "comment"]  
  9.         }
  10.       }
  11.     }
  12.   }
  13. }

question is parent of answer and comment.

Multiple levels of parent join

Using multiple levels of relations to replicate a relational model is not recommended. Each level of relation adds an overhead at query time in terms of memory and computation. You should de-normalize your data if you care about performance.

Multiple levels of parent/child:

  1. PUT my-index-000001
  2. {
  3.   "mappings": {
  4.     "properties": {
  5.       "my_join_field": {
  6.         "type": "join",
  7.         "relations": {
  8.           "question": ["answer", "comment"],  
  9.           "answer": "vote" 
  10.         }
  11.       }
  12.     }
  13.   }
  14. }

question is parent of answer and comment

answer is parent of vote

The mapping above represents the following tree:

  1.    question
  2.     /    \
  3.    /      \
  4. comment  answer
  5.            |
  6.            |
  7.           vote

Indexing a grandchild document requires a routing value equals to the grand-parent (the greater parent of the lineage):

  1. PUT my-index-000001/_doc/3?routing=1&refresh 
  2. {
  3.   "text": "This is a vote",
  4.   "my_join_field": {
  5.     "name": "vote",
  6.     "parent": "2" 
  7.   }
  8. }

This child document must be on the same shard than its grand-parent and parent

The parent id of this document (must points to an answer document)