COLLECT

The COLLECT keyword can be used to group an array by one or multiple group criteria.

The COLLECT statement will eliminate all local variables in the current scope. After COLLECT only the variables introduced by COLLECT itself are available.

The general syntaxes for COLLECT are:

COLLECT variableName = expression options
COLLECT variableName = expression INTO groupsVariable options
COLLECT variableName = expression INTO groupsVariable = projectionExpression options
COLLECT variableName = expression INTO groupsVariable KEEP keepVariable options
COLLECT variableName = expression WITH COUNT INTO countVariable options
COLLECT variableName = expression AGGREGATE variableName = aggregateExpression options
COLLECT AGGREGATE variableName = aggregateExpression options
COLLECT WITH COUNT INTO countVariable options

options is optional in all variants.

Grouping syntaxes

The first syntax form of COLLECT only groups the result by the defined group criteria specified in expression. In order to further process the results produced by COLLECT, a new variable (specified by variableName) is introduced. This variable contains the group value.

Here's an example query that find the distinct values in u.city and makes them available in variable city:

FOR u IN users
  COLLECT city = u.city
  RETURN { 
    "city" : city 
  }

The second form does the same as the first form, but additionally introduces a variable (specified by groupsVariable) that contains all elements that fell into the group. This works as follows: The groupsVariable variable is an array containing as many elements as there are in the group. Each member of that array is a JSON object in which the value of every variable that is defined in the AQL query is bound to the corresponding attribute. Note that this considers all variables that are defined before the COLLECT statement, but not those on the top level (outside of any FOR), unless the COLLECT statement is itself on the top level, in which case all variables are taken. Furthermore note that it is possible that the optimizer moves LET statements out of FOR statements to improve performance.

FOR u IN users
  COLLECT city = u.city INTO groups
  RETURN { 
    "city" : city, 
    "usersInCity" : groups 
  }

In the above example, the array users will be grouped by the attribute city. The result is a new array of documents, with one element per distinct u.city value. The elements from the original array (here: users) per city are made available in the variable groups. This is due to the INTO clause.

COLLECT also allows specifying multiple group criteria. Individual group criteria can be separated by commas:

FOR u IN users
  COLLECT country = u.country, city = u.city INTO groups
  RETURN { 
    "country" : country, 
    "city" : city, 
    "usersInCity" : groups 
  }

In the above example, the array users is grouped by country first and then by city, and for each distinct combination of country and city, the users will be returned.

Discarding obsolete variables

The third form of COLLECT allows rewriting the contents of the groupsVariable using an arbitrary projectionExpression:

FOR u IN users
  COLLECT country = u.country, city = u.city INTO groups = u.name
  RETURN { 
    "country" : country, 
    "city" : city, 
    "userNames" : groups 
  }

In the above example, only the projectionExpression is u.name. Therefore, only this attribute is copied into the groupsVariable for each document. This is probably much more efficient than copying all variables from the scope into the groupsVariable as it would happen without a projectionExpression.

The expression following INTO can also be used for arbitrary computations:

FOR u IN users
  COLLECT country = u.country, city = u.city INTO groups = { 
    "name" : u.name, 
    "isActive" : u.status == "active"
  }
  RETURN { 
    "country" : country, 
    "city" : city, 
    "usersInCity" : groups 
  }

COLLECT also provides an optional KEEP clause that can be used to control which variables will be copied into the variable created by INTO. If no KEEP clause is specified, all variables from the scope will be copied as sub-attributes into the groupsVariable. This is safe but can have a negative impact on performance if there are many variables in scope or the variables contain massive amounts of data.

The following example limits the variables that are copied into the groupsVariable to just name. The variables u and someCalculation also present in the scope will not be copied into groupsVariable because they are not listed in the KEEP clause:

FOR u IN users
  LET name = u.name
  LET someCalculation = u.value1 + u.value2
  COLLECT city = u.city INTO groups KEEP name 
  RETURN { 
    "city" : city, 
    "userNames" : groups[*].name 
  }

KEEP is only valid in combination with INTO. Only valid variable names can be used in the KEEP clause. KEEP supports the specification of multiple variable names.

Group length calculation

COLLECT also provides a special WITH COUNT clause that can be used to determine the number of group members efficiently.

The simplest form just returns the number of items that made it into the COLLECT:

FOR u IN users
  COLLECT WITH COUNT INTO length
  RETURN length

The above is equivalent to, but more efficient than:

RETURN LENGTH(
  FOR u IN users
    RETURN length
)

The WITH COUNT clause can also be used to efficiently count the number of items in each group:

FOR u IN users
  COLLECT age = u.age WITH COUNT INTO length
  RETURN { 
    "age" : age, 
    "count" : length 
  }

Note: the WITH COUNT clause can only be used together with an INTO clause.

Aggregation

A COLLECT statement can be used to perform aggregation of data per group. To only determine group lengths, the WITH COUNT INTO variant of COLLECT can be used as described before.

For other aggregations, it is possible to run aggregate functions on the COLLECT results:

FOR u IN users
  COLLECT ageGroup = FLOOR(u.age / 5) * 5 INTO g
  RETURN { 
    "ageGroup" : ageGroup,
    "minAge" : MIN(g[*].u.age),
    "maxAge" : MAX(g[*].u.age)
  }

The above however requires storing all group values during the collect operation for all groups, which can be inefficient.

The special AGGREGATE variant of COLLECT allows building the aggregate values incrementally during the collect operation, and is therefore often more efficient.

With the AGGREGATE variant the above query becomes:

FOR u IN users
  COLLECT ageGroup = FLOOR(u.age / 5) * 5 
  AGGREGATE minAge = MIN(u.age), maxAge = MAX(u.age)
  RETURN {
    ageGroup, 
    minAge, 
    maxAge 
  }

The AGGREGATE keyword can only be used after the COLLECT keyword. If used, it must directly follow the declaration of the grouping keys. If no grouping keys are used, it must follow the COLLECT keyword directly:

FOR u IN users
  COLLECT AGGREGATE minAge = MIN(u.age), maxAge = MAX(u.age)
  RETURN {
    minAge, 
    maxAge 
  }

Only specific expressions are allowed on the right-hand side of each AGGREGATE assignment:

  • on the top level, an aggregate expression must be a call to one of the supported aggregation functions LENGTH, MIN, MAX, SUM, AVERAGE, STDDEV_POPULATION, STDDEV_SAMPLE, VARIANCE_POPULATION, or VARIANCE_SAMPLE

  • an aggregate expression must not refer to variables introduced by the COLLECT itself

COLLECT variants

Since ArangoDB 2.6, there are two variants of COLLECT that the optimizer can choose from: the sorted variant and the hash variant. The hash variant only becomes a candidate for COLLECT statements that do not use an INTO clause.

The optimizer will always generate a plan that employs the sorted method. The sorted method requires its input to be sorted by the group criteria specified in the COLLECT clause. To ensure correctness of the result, the AQL optimizer will automatically insert a SORT statement into the query in front of the COLLECT statement. The optimizer may be able to optimize away that SORT statement later if a sorted index is present on the group criteria.

In case a COLLECT qualifies for using the hash variant, the optimizer will create an extra plan for it at the beginning of the planning phase. In this plan, no extra SORT statement will be added in front of the COLLECT. This is because the hash variant of COLLECT does not require sorted input. Instead, a SORT statement will be added after the COLLECT to sort its output. This SORT statement may be optimized away again in later stages. If the sort order of the COLLECT is irrelevant to the user, adding the extra instruction SORT null after the COLLECT will allow the optimizer to remove the sorts altogether:

FOR u IN users
  COLLECT age = u.age
  SORT null  /* note: will be optimized away */
  RETURN age

Which COLLECT variant is used by the optimizer depends on the optimizer's cost estimations. The created plans with the different COLLECT variants will be shipped through the regular optimization pipeline. In the end, the optimizer will pick the plan with the lowest estimated total cost as usual.

In general, the sorted variant of COLLECT should be preferred in cases when there is a sorted index present on the group criteria. In this case the optimizer can eliminate the SORT statement in front of the COLLECT, so that no SORT will be left.

If there is no sorted index available on the group criteria, the up-front sort required by the sorted variant can be expensive. In this case it is likely that the optimizer will prefer the hash variant of COLLECT, which does not require its input to be sorted.

Which variant of COLLECT was actually used can be figured out by looking into the execution plan of a query, specifically the AggregateNode and its aggregationOptions attribute.

Setting COLLECT options

options can be used in a COLLECT statement to inform the optimizer about the preferred COLLECT method. When specifying the following appendix to a COLLECT statement, the optimizer will always use the sorted variant of COLLECT and not even create a plan using the hash variant:

OPTIONS { method: "sorted" }

Note that specifying hash as method will not make the optimizer use the hash variant. This is because the hash variant is not eligible for all queries. Instead, if no options or any other method than sorted are specified in OPTIONS, the optimizer will use its regular cost estimations.

COLLECT vs. RETURN DISTINCT

In order to make a result set unique, one can either use COLLECT or RETURN DISTINCT. Behind the scenes, both variants will work by creating an AggregateNode. For both variants, the optimizer may try the sorted and the hashed variant of COLLECT. The difference is therefore mainly syntactical, with RETURN DISTINCT saving a bit of typing when compared to an equivalent COLLECT:

FOR u IN users
  RETURN DISTINCT u.age
FOR u IN users
  COLLECT age = u.age
  RETURN age

However, COLLECT is vastly more flexible than RETURN DISTINCT. Additionally, the order of results is undefined for a RETURN DISTINCT, whereas for a COLLECT the results will be sorted.