Terms Aggregation

The size parameter can be set to define how many term buckets should be returned out of the overall terms list. By default, the node coordinating the search process will request each shard to provide its own top size term buckets and once all shards respond, it will reduce the results to the final list that will then be returned to the client. This means that if the number of unique terms is greater than size, the returned list is slightly off and not accurate (it could be that the term counts are slightly off and it could even be that a term that should have been in the top size buckets was not returned). If set to 0, the size will be set to Integer.MAX_VALUE.

As described above, the document counts (and the results of any sub aggregations) in the terms aggregation are not always accurate. This is because each shard provides its own view of what the ordered list of terms should be and these are combined to give a final view. Consider the following scenario:

A request is made to obtain the top 5 terms in the field product, ordered by descending document count from an index with 3 shards. In this case each shard is asked to give its top 5 terms.

{
    "aggs" : {
        "products" : {
            "terms" : {
                "field" : "product",
                "size" : 5
            }
        }
    }
}

The terms for each of the three shards are shown below with their respective document counts in brackets:

The shards will return their top 5 terms so the results from the shards will be:

Taking the top 5 results from each of the shards (as requested) and combining them to make a final top 5 list produces the following:

Because Product A was returned from all shards we know that its document count value is accurate. Product C was only returned by shards A and C so its document count is shown as 50 but this is not an accurate count. Product C exists on shard B, but its count of 4 was not high enough to put Product C into the top 5 list for that shard. Product Z was also returned only by 2 shards but the third shard does not contain the term. There is no way of knowing, at the point of combining the results to produce the final list of terms, that there is an error in the document count for Product C and not for Product Z. Product H has a document count of 44 across all 3 shards but was not included in the final list of terms because it did not make it into the top five terms on any of the shards.

The higher the requested size is, the more accurate the results will be, but also, the more expensive it will be to compute the final results (both due to bigger priority queues that are managed on a shard level and due to bigger data transfers between the nodes and the client).

The shard_size parameter can be used to minimize the extra work that comes with bigger requested size. When defined, it will determine how many terms the coordinating node will request from each shard. Once all the shards responded, the coordinating node will then reduce them to a final result which will be based on the size parameter - this way, one can increase the accuracy of the returned terms and avoid the overhead of streaming a big list of buckets back to the client. If set to 0, the shard_size will be set to Integer.MAX_VALUE.

It is possible to not limit the number of terms that are returned by setting size to 0. Don’t use this on high-cardinality fields as this will kill both your CPU since terms need to be return sorted, and your network.

The default shard_size is a multiple of the size parameter which is dependant on the number of shards.

There are two error values which can be shown on the terms aggregation. The first gives a value for the aggregation as a whole which represents the maximum potential document count for a term which did not make it into the final list of terms. This is calculated as the sum of the document count from the last term returned from each shard .For the example given above the value would be 46 (2 + 15 + 29). This means that in the worst case scenario a term which was not returned could have the 4th highest document count.

{
    ...

    "aggregations" : {
        "products" : {
            "doc_count_error_upper_bound" : 46,
            "buckets" : [
                {
                    "key" : "Product A",
                    "doc_count" : 100
                },
                {
                    "key" : "Product Z",
                    "doc_count" : 52
                },
                ...
            ]
        }
    }
}

The second error value can be enabled by setting the show_term_doc_count_error parameter to true. This shows an error value for each term returned by the aggregation which represents the worst case error in the document count and can be useful when deciding on a value for the shard_size parameter. This is calculated by summing the document counts for the last term returned by all shards which did not return the term. In the example above the error in the document count for Product C would be 15 as Shard B was the only shard not to return the term and the document count of the last termit did return was 15. The actual document count of Product C was 54 so the document count was only actually off by 4 even though the worst case was that it would be off by 15. Product A, however has an error of 0 for its document count, since every shard returned it we can be confident that the count returned is accurate.

{
    ...

    "aggregations" : {
        "products" : {
            "doc_count_error_upper_bound" : 46,
            "buckets" : [
                {
                    "key" : "Product A",
                    "doc_count" : 100,
                    "doc_count_error_upper_bound" : 0
                },
                {
                    "key" : "Product Z",
                    "doc_count" : 52,
                    "doc_count_error_upper_bound" : 2
                },
                ...
            ]
        }
    }
}

These errors can only be calculated in this way when the terms are ordered by descending document count. When the aggregation is ordered by the terms values themselves (either ascending or descending) there is no error in the document count since if a shard does not return a particular term which appears in the results from another shard, it must not have that term in its index. When the aggregation is either sorted by a sub aggregation or in order of ascending document count, the error in the document counts cannot be determined and is given a value of -1 to indicate this.

The order of the buckets can be customized by setting the order parameter. By default, the buckets are ordered by their doc_count descending. It is also possible to change this behaviour as follows:

Ordering the buckets by their doc_count in an ascending manner:

{
    "aggs" : {
        "genders" : {
            "terms" : {
                "field" : "gender",
                "order" : { "_count" : "asc" }
            }
        }
    }
}

Ordering the buckets alphabetically by their terms in an ascending manner:

{
    "aggs" : {
        "genders" : {
            "terms" : {
                "field" : "gender",
                "order" : { "_term" : "asc" }
            }
        }
    }
}

Ordering the buckets by single value metrics sub-aggregation (identified by the aggregation name):

{
    "aggs" : {
        "genders" : {
            "terms" : {
                "field" : "gender",
                "order" : { "avg_height" : "desc" }
            },
            "aggs" : {
                "avg_height" : { "avg" : { "field" : "height" } }
            }
        }
    }
}

Ordering the buckets by multi value metrics sub-aggregation (identified by the aggregation name):

{
    "aggs" : {
        "genders" : {
            "terms" : {
                "field" : "gender",
                "order" : { "height_stats.avg" : "desc" }
            },
            "aggs" : {
                "height_stats" : { "stats" : { "field" : "height" } }
            }
        }
    }
}

It is also possible to order the buckets based on a "deeper" aggregation in the hierarchy. This is supported as long as the aggregations path are of a single-bucket type, where the last aggregation in the path may either be a single-bucket one or a metrics one. If it’s a single-bucket type, the order will be defined by the number of docs in the bucket (i.e. doc_count), in case it’s a metrics one, the same rules as above apply (where the path must indicate the metric name to sort by in case of a multi-value metrics aggregation, and in case of a single-value metrics aggregation the sort will be applied on that value).

The path must be defined in the following form:

AGG_SEPARATOR       :=  '>'
METRIC_SEPARATOR    :=  '.'
AGG_NAME            :=  <the name of the aggregation>
METRIC              :=  <the name of the metric (in case of multi-value metrics aggregation)>
PATH                :=  <AGG_NAME>[<AGG_SEPARATOR><AGG_NAME>]*[<METRIC_SEPARATOR><METRIC>]

{
    "aggs" : {
        "countries" : {
            "terms" : {
                "field" : "address.country",
                "order" : { "females>height_stats.avg" : "desc" }
            },
            "aggs" : {
                "females" : {
                    "filter" : { "term" : { "gender" :  "female" }},
                    "aggs" : {
                        "height_stats" : { "stats" : { "field" : "height" }}
                    }
                }
            }
        }
    }
}

The above will sort the countries buckets based on the average height among the female population.

It is possible to only return terms that match more than a configured number of hits using the min_doc_count option:

{
    "aggs" : {
        "tags" : {
            "terms" : {
                "field" : "tags",
                "min_doc_count": 10
            }
        }
    }
}

The above aggregation would only return tags which have been found in 10 hits or more. Default value is 1.

Terms are collected and ordered on a shard level and merged with the terms collected from other shards in a second step. However, the shard does not have the information about the global document count available. The decision if a term is added to a candidate list depends only on the order computed on the shard using local shard frequencies. The min_doc_count criterion is only applied after merging local terms statistics of all shards. In a way the decision to add the term as a candidate is made without being very certain about if the term will actually reach the required min_doc_count. This might cause many (globally) high frequent terms to be missing in the final result if low frequent terms populated the candidate lists. To avoid this, the shard_size parameter can be increased to allow more candidate terms on the shards. However, this increases memory consumption and network traffic.

shard_min_doc_count parameter

The parameter shard_min_doc_count regulates the certainty a shard has if the term should actually be added to the candidate list or not with respect to the min_doc_count. Terms will only be considered if their local shard frequency within the set is higher than the shard_min_doc_count. If your dictionary contains many low frequent terms and you are not interested in those (for example misspellings), then you can set the shard_min_doc_count parameter to filter out candidate terms on a shard level that will with a reasonable certainty not reach the required min_doc_count even after merging the local counts. shard_min_doc_count is set to 0 per default and has no effect unless you explicitly set it.

Generating the terms using a script:

{
    "aggs" : {
        "genders" : {
            "terms" : {
                "script" : "doc['gender'].value"
            }
        }
    }
}

It is possible to filter the values for which buckets will be created. This can be done using the include and exclude parameters which are based on regular expressions.

{
    "aggs" : {
        "tags" : {
            "terms" : {
                "field" : "tags",
                "include" : ".*sport.*",
                "exclude" : "water_.*"
            }
        }
    }
}

In the above example, buckets will be created for all the tags that has the word sport in them, except those starting with water_ (so the tag water_sports will no be aggregated). The include regular expression will determine what values are "allowed" to be aggregated, while the exclude determines the values that should not be aggregated. When both are defined, the exclude has precedence, meaning, the include is evaluated first and only then the exclude.

The regular expression are based on the Java™ Pattern, and as such, they it is also possible to pass in flags that will determine how the compiled regular expression will work:

{
    "aggs" : {
        "tags" : {
             "terms" : {
                 "field" : "tags",
                 "include" : {
                     "pattern" : ".*sport.*",
                     "flags" : "CANON_EQ|CASE_INSENSITIVE" 
                 },
                 "exclude" : {
                     "pattern" : "water_.*",
                     "flags" : "CANON_EQ|CASE_INSENSITIVE"
                 }
             }
         }
    }
}

The possible flags that can be used are: CANON_EQ, CASE_INSENSITIVE, COMMENTS, DOTALL, LITERAL, MULTILINE, UNICODE_CASE, UNICODE_CHARACTER_CLASS and UNIX_LINES

The terms aggregation does not support collecting terms from multiple fields in the same document. The reason is that the terms agg doesn’t collect the string term values themselves, but rather uses global ordinals to produce a list of all of the unique values in the field. Global ordinals results in an important performance boost which would not be possible across multiple fields.

There are two approaches that you can use to perform a terms agg across multiple fields:

Deferring calculation of child aggregations

For fields with many unique terms and a small number of required results it can be more efficient to delay the calculation of child aggregations until the top parent-level aggs have been pruned. Ordinarily, all branches of the aggregation tree are expanded in one depth-first pass and only then any pruning occurs. In some rare scenarios this can be very wasteful and can hit memory constraints. An example problem scenario is querying a movie database for the 10 most popular actors and their 5 most common co-stars:

{
    "aggs" : {
        "actors" : {
             "terms" : {
                 "field" : "actors",
                 "size" : 10
             },
            "aggs" : {
                "costars" : {
                     "terms" : {
                         "field" : "actors",
                         "size" : 5
                     }
                 }
            }
         }
    }
}

Even though the number of movies may be comparatively small and we want only 50 result buckets there is a combinatorial explosion of buckets during calculation - a single movie will produce n² buckets where n is the number of actors. The sane option would be to first determine the 10 most popular actors and only then examine the top co-stars for these 10 actors. This alternative strategy is what we call the breadth_first collection mode as opposed to the default depth_first mode:

{
    "aggs" : {
        "actors" : {
             "terms" : {
                 "field" : "actors",
                 "size" : 10,
                 "collect_mode" : "breadth_first"
             },
            "aggs" : {
                "costars" : {
                     "terms" : {
                         "field" : "actors",
                         "size" : 5
                     }
                 }
            }
         }
    }
}

When using breadth_first mode the set of documents that fall into the uppermost buckets are cached for subsequent replay so there is a memory overhead in doing this which is linear with the number of matching documents. In most requests the volume of buckets generated is smaller than the number of documents that fall into them so the default depth_first collection mode is normally the best bet but occasionally the breadth_first strategy can be significantly more efficient. Currently elasticsearch will always use the depth_first collect_mode unless explicitly instructed to use breadth_first as in the above example. Note that the order parameter can still be used to refer to data from a child aggregation when using the breadth_first setting - the parent aggregation understands that this child aggregation will need to be called first before any of the other child aggregations.

There are different mechanisms by which terms aggregations can be executed:

Elasticsearch tries to have sensible defaults so this is something that generally doesn’t need to be configured.

map should only be considered when very few documents match a query. Otherwise the ordinals-based execution modes are significantly faster. By default, map is only used when running an aggregation on scripts, since they don’t have ordinals.

global_ordinals_low_cardinality only works for leaf terms aggregations but is usually the fastest execution mode. Memory usage is linear with the number of unique values in the field, so it is only enabled by default on low-cardinality fields.

global_ordinals is the second fastest option, but the fact that it preemptively allocates buckets can be memory-intensive, especially if you have one or more sub aggregations. It is used by default on top-level terms aggregations.

global_ordinals_hash on the contrary to global_ordinals and global_ordinals_low_cardinality allocates buckets dynamically so memory usage is linear to the number of values of the documents that are part of the aggregation scope. It is used by default in inner aggregations.

{
    "aggs" : {
        "tags" : {
             "terms" : {
                 "field" : "tags",
                 "execution_hint": "map" 
             }
         }
    }
}

Please note that Elasticsearch will ignore this execution hint if it is not applicable and that there is no backward compatibility guarantee on these hints.