SIREn Join Plugin for Elasticsearch

This plugin extends Elasticsearch with new search actions and a filter query parser that enables to perform a "Filter Join" between two set of documents (in the same index or in different indexes).

The Filter Join is basically a (left) semi-join between two set of documents based on a common attribute, where the result only contains the attributes of one of the joined set of documents. This join is used to filter one document set based on a second document set, hence its name. It is equivalent to the EXISTS() operator in SQL.

Compatibility

The following table shows the compatibility between releases of Elasticsearch and the SIREn Join plugin:

Elasticsearch	SIREn Join
2.3.4	2.3.4
2.3.3	2.3.3-1
2.2.0	2.2.0-1
2.1.2	2.1.2
2.1.1	2.1.1
1.7.x	1.0

Installing the Plugin

Online Download

You can use the following command to download the plugin from the online repository:

$ bin/plugin install solutions.siren/siren-join/2.3.4

Offline Download

• Get the ZIPball from maven.org

• Install with the downloaded file

  $ bin/plugin install file:/path/to/folder/with/siren-join-2.3.4.zip

Manual

Alternatively, you can assemble it via Maven (you must build it as a non-root user):

$ git clone [email protected]:sirensolutions/siren-join.git
$ cd siren-join
$ mvn package

This creates a single Zip file that can be installed using the Elasticsearch plugin command:

$ bin/plugin install file:/PATH-TO-SIRENJOIN-PROJECT/target/releases/siren-join-2.3.4.zip

Interacting with the Plugin

You can now start Elasticsearch and see that our plugin gets loaded:

$ bin/elasticsearch
...
[2013-09-04 17:33:27,443][INFO ][node    ] [Andrew Chord] initializing ...
[2013-09-04 17:33:27,455][INFO ][plugins ] [Andrew Chord] loaded [siren-join], sites []
...

To uninstall the plugin:

$ bin/plugin remove siren-join

Usage

Coordinate Search API

This plugin introduces two new search actions, _coordinate_search that replaces the _search action, and _coordinate_msearch that replaces the _msearch action. Both actions are wrappers around the original elasticsearch actions and therefore supports the same API. One must use these actions with the filterjoin filter, as the filterjoin filter is not supported by the original elaticsearch actions.

Parameters

• filterjoin: the filter name
• indices: the index names to lookup the terms from (optional, default to all indices).
• types: the index types to lookup the terms from (optional, default to all types).
• path: the path within the document to lookup the terms from.
• query: the query used to lookup terms with.
• orderBy: the ordering to use to lookup the maximum number of terms: default, doc_score (optional, default to default ordering).
• maxTermsPerShard: the maximum number of terms per shard to lookup (optional, default to all terms).
• termsEncoding: the encoding to use when transferring terms across the network: long, integer, bloom, bytes (optional, default to long).

Example

In this example, we will join all the documents from index1 with the documents of index2. The query first filters documents from index2 and of type type with the query { "terms" : { "tag" : [ "aaa" ] } }. It then retrieves the ids of the documents from the field id specified by the parameter path. The list of ids is then used as filter and applied on the field foreign_key of the documents from index1.

    {
      "bool" : {
        "filter" : {
          "filterjoin" : {
            "foreign_key" : {
              "indices" : ["index2"],
              "types" : ["type"],
              "path" : "id",
              "query" : {
                "terms" : {
                  "tag" : [ "aaa" ]
                }
              }
            }
          }
        }
      }
    }

Response Format

The response returned by the coordinate search API is identical to the response returned by Elasticsearch's search API, but augmented with additional information about the execution of the relational query planning. This additional information is stored within the field named coordinate_search at the root of the response, see example below. The object contains the following parameters:

• actions: a list of actions that has been executed - an action represents the execution of one single join.
• relations: the definition of the relations of the join - it contains two nested objects, from and to, one for each relation.
• size: the size of the filter used to compute the join, i.e., the number of terms across all shards used by the filterjoin.
• size_in_bytes: the size in bytes of the filter used to compute the join.
• is_pruned: a flag to indicate if the join computation has been pruned based on the maxTermsPerShard limit.
• cache_hit: a flag to indicate if the join was already computed and cached.
• terms_encoding: the terms encoding used to transfer terms across the network.
• took: the time it took to construct the filter.

    {
      "coordinate_search": {
        "actions": [
          {
            "relations": {
              "from": {
                "indices": ["index2"],
                "types": ["type"],
                "field": "id"
              },
              "to": {
                "indices": null,
                "types": null,
                "field": "foreign_key"
              }
            },
            "size": 2,
            "size_in_bytes": 20,
            "is_pruned": false,
            "cache_hit": false,
            "terms_encoding" : "long",
            "took": 313
          }
        ]
      },
    ...
    }

Performance Considerations

• We recommend to activate caching for all queries via the setting index.queries.cache.everything: true. The new caching policy of Elasticsearch will not cache a filterjoin query on small segments which can lead to a significant drop of performance. See issue 16529 for more information.
• Joining numeric attributes is more efficient than joining string attributes.
• The bloom filter is the most efficient and the default encoding method for terms. It can encode 40M unique values in ~30MB. However, this trades precision for space, i.e., the bloom filter can lead to false-positive results. If precision is critical, then it is recommended to switch to the terms encoding to long.
• If the joined attributes of your documents contain incremental integers, switch the terms encoding to integer.
• The filterjoin includes a circuit breaker to prevent OOME when joining a field with a large number of unique values. As a rule of thumb, the maximum amount of unique values transferred across the shards should be around 50 to 100M when using bloom encoding, 5 to 10M when using long or integer encoding. It is recommended to configure a maxTermsPerShard limit if the attribute defined by the path parameter contains a larger number of values.
• The bytes terms encoding will likely provide better performance for highly selective queries over large indices, as it will perform the filtering based on a dictionary lookup instead of a doc value scan.

Acknowledgement

Part of this plugin is inspired and based on the pull request 3278 submitted by Matt Weber to the Elasticsearch project.

---

Copyright (c) 2016, SIREn Solutions. All Rights Reserved.