User-defined aggregate functions

For support during the preview, email bigquery-sql-preview-support@google.com.

This document describes how to create, call, and delete user-defined aggregate functions (UDAFs) in BigQuery.

A UDAF lets you create an aggregate function by using an expression that contains code. A UDAF accepts columns of input, performs a calculation on a group of rows at a time, and then returns the result of that calculation as a single value.

Create a SQL UDAF

This section describes the various ways that you can create a persistent or temporary SQL UDAF in BigQuery.

Create a persistent SQL UDAF

You can create a SQL UDAF that is persistent, meaning that you can reuse the UDAF across multiple queries. Persistent UDAFs are safe to call when they are shared between owners. UDAFs can't mutate data, talk to external systems, or send logs to Google Cloud Observability or similar applications.

To create a persistent UDAF, use the CREATE AGGREGATE FUNCTION statement without the TEMP or TEMPORARY keyword. You must include the dataset in the function path.

For example, the following query creates a persistent UDAF that's called ScaledAverage:

CREATE AGGREGATE FUNCTION myproject.mydataset.ScaledAverage(
  dividend FLOAT64,
  divisor FLOAT64)
RETURNS FLOAT64
AS (
  AVG(dividend / divisor)
);

Create a temporary SQL UDAF

You can create a SQL UDAF that is temporary, meaning that the UDAF only exists in the scope of a single query, script, session, or procedure.

To create a temporary UDAF, use the CREATE AGGREGATE FUNCTION statement with the TEMP or TEMPORARY keyword.

For example, the following query creates a temporary UDAF that's called ScaledAverage:

CREATE TEMP AGGREGATE FUNCTION ScaledAverage(
  dividend FLOAT64,
  divisor FLOAT64)
RETURNS FLOAT64
AS (
  AVG(dividend / divisor)
);

Use aggregate and non-aggregate parameters

You can create a SQL UDAF that has both aggregate and non-aggregate parameters.

UDAFs normally aggregate function parameters across all rows in a group. However, you can specify a function parameter as non-aggregate with the NOT AGGREGATE keyword.

A non-aggregate function parameter is a scalar function parameter with a constant value for all rows in a group. A valid non-aggregate function parameter must be a literal. Inside the UDAF definition, aggregate function parameters can only appear as function arguments to aggregate function calls. References to non-aggregate function parameters can appear anywhere in the UDAF definition.

For example, the following function contains an aggregate parameter that's called dividend, and a non-aggregate parameter called divisor:

-- Create the function.
CREATE TEMP AGGREGATE FUNCTION ScaledSum(
  dividend FLOAT64,
  divisor FLOAT64 NOT AGGREGATE)
RETURNS FLOAT64
AS (
  SUM(dividend) / divisor
);

Use the default project in the function body

In the body of a SQL UDAF, any references to BigQuery entities, such as tables or views, must include the project ID unless the entity resides in the same project that contains the UDAF.

For example, consider the following statement:

CREATE AGGREGATE FUNCTION project1.dataset_a.ScaledAverage(
  dividend FLOAT64,
  divisor FLOAT64)
RETURNS FLOAT64
AS (
  ( SELECT AVG(dividend / divisor) FROM dataset_a.my_table )
);

If you run the preceding statement in the project1 project, the statement succeeds because my_table exists in project1. However, if you run the preceding statement from a different project, the statement fails. To correct the error, include the project ID in the table reference:

CREATE AGGREGATE FUNCTION project1.dataset_a.ScaledAverage(
  dividend FLOAT64,
  divisor FLOAT64)
RETURNS FLOAT64
AS (
  ( SELECT AVG(dividend / divisor) FROM project1.dataset_a.my_table )
);

You can also reference an entity in a different project or dataset from the one where you create the function:

CREATE AGGREGATE FUNCTION project1.dataset_a.ScaledAverage(
  dividend FLOAT64,
  divisor FLOAT64)
RETURNS FLOAT64
AS (
  ( SELECT AVG(dividend / divisor) FROM project2.dataset_c.my_table )
);

Create a JavaScript UDAF

This section describes the various ways in which you can create a JavaScript UDAF in BigQuery. There are a few rules to observe when creating a JavaScript UDAF:

  • The body of the JavaScript UDAF must be a quoted string literal that represents the JavaScript code. To learn more about the different types of quoted string literals that you can use, see Formats for quoted literals.

  • Only certain type encodings are allowed. For more information, see Permitted SQL type encodings in a JavaScript UDAF.

  • The JavaScript function body must include four JavaScript functions that initialize, aggregate, merge, and finalize the results for the JavaScript UDAF (initialState, aggregate, merge, and finalize). For more information, see Permitted SQL type encodings in a JavaScript UDAF.

  • Any value returned by the initialState function or that is left in the state argument after the aggregate or merge function is called, must be serializable. If you want to work with non-serializable aggregation data, such as functions or symbol fields, you must use the included serialize and deserialize functions. To learn more, see Serialize and deserialize data in a JavaScript UDAF.

Create a persistent JavaScript UDAF

You can create a JavaScript UDAF that is persistent, meaning that you can reuse the UDAF across multiple queries. Persistent UDAFs are safe to call when they are shared between owners. UDAFs can't mutate data, talk to external systems, or send logs to Google Cloud Observability or similar applications.

To create a persistent UDAF, use the CREATE AGGREGATE FUNCTION statement without the TEMP or TEMPORARY keyword. You must include the dataset in the function path.

The following query creates a persistent JavaScript UDAF that's called SumPositive:

CREATE OR REPLACE AGGREGATE FUNCTION my_project.my_dataset.SumPositive(x FLOAT64)
RETURNS FLOAT64
LANGUAGE js
AS r'''

  export function initialState() {
    return {sum: 0}
  }
  export function aggregate(state, x) {
    if (x > 0) {
      state.sum += x;
    }
  }
  export function merge(state, partialState) {
    state.sum += partialState.sum;
  }
  export function finalize(state) {
    return state.sum;
  }

''';

-- Call the JavaScript UDAF.
WITH numbers AS (
  SELECT * FROM UNNEST([1.0, -1.0, 3.0, -3.0, 5.0, -5.0]) AS x)
SELECT my_project.my_dataset.SumPositive(x) AS sum FROM numbers;

/*-----*
 | sum |
 +-----+
 | 9.0 |
 *-----*/

Create a temporary JavaScript UDAF

You can create a JavaScript UDAF that is temporary, meaning that the UDAF only exists in the scope of a single query, script, session, or procedure.

To create a temporary UDAF, use the CREATE AGGREGATE FUNCTION statement with the TEMP or TEMPORARY keyword.

The following query creates a temporary JavaScript UDAF that's called SumPositive:

CREATE TEMP AGGREGATE FUNCTION SumPositive(x FLOAT64)
RETURNS FLOAT64
LANGUAGE js
AS r'''

  export function initialState() {
    return {sum: 0}
  }
  export function aggregate(state, x) {
    if (x > 0) {
      state.sum += x;
    }
  }
  export function merge(state, partialState) {
    state.sum += partialState.sum;
  }
  export function finalize(state) {
    return state.sum;
  }

''';

-- Call the JavaScript UDAF.
WITH numbers AS (
  SELECT * FROM UNNEST([1.0, -1.0, 3.0, -3.0, 5.0, -5.0]) AS x)
SELECT SumPositive(x) AS sum FROM numbers;

/*-----*
 | sum |
 +-----+
 | 9.0 |
 *-----*/

Include non-aggregate parameters in a JavaScript UDAF

You can create a JavaScript UDAF that has both aggregate and non-aggregate parameters.

UDAFs normally aggregate function parameters across all rows in a group. However, you can specify a function parameter as non-aggregate with the NOT AGGREGATE keyword.

A non-aggregate function parameter is a scalar function parameter with a constant value for all rows in a group. A valid non-aggregate function parameter must be a literal. Inside the UDAF definition, aggregate function parameters can only appear as function arguments to aggregate function calls. References to non-aggregate function parameters can appear anywhere in the UDAF definition.

In the following example, the JavaScript UDAF contains an aggregate parameter called s and a non-aggregate parameter called delimiter:

CREATE TEMP AGGREGATE FUNCTION JsStringAgg(
  s STRING,
  delimiter STRING NOT AGGREGATE)
RETURNS STRING
LANGUAGE js
AS r'''

  export function initialState() {
    return {strings: []}
  }
  export function aggregate(state, s) {
    state.strings.push(s);
  }
  export function merge(state, partialState) {
    state.strings = state.strings.concat(partialState.strings);
  }
  export function finalize(state, delimiter) {
    return state.strings.join(delimiter);
  }

''';

-- Call the JavaScript UDAF.
WITH strings AS (
  SELECT * FROM UNNEST(["aaa", "bbb", "ccc", "ddd"]) AS values)
SELECT JsStringAgg(values, '.') AS result FROM strings;

/*-----------------*
 | result          |
 +-----------------+
 | aaa.bbb.ccc.ddd |
 *-----------------*/

Serialize and deserialize data in a JavaScript UDAF

BigQuery must serialize any object returned by the initialState function or that is left in the state argument after the aggregate or merge function is called. BigQuery supports serializing an object if all fields are one of the following:

  • A JavaScript primitive value (for example: 2, "abc", null, undefined).
  • A JavaScript object for which BigQuery supports serializing all field values.
  • A JavaScript array for which BigQuery supports serializing all elements.

The following return values are serializable:

export function initialState() {
  return {a: "", b: 3, c: null, d: {x: 23} }
}
export function initialState() {
  return {value: 2.3};
}

The following return values are not serializable:

export function initialState() {
  return {
    value: function() {return 6;}
  }
}
export function initialState() {
  return 2.3;
}

If you want to work with non-serializable aggregation states, the JavaScript UDAF must include the serialize and deserialize functions. The serialize function converts the aggregation state into a serializable object; the deserialize function converts the serializable object back into an aggregation state.

In the following example, an external library calculates sums by using an interface:

export class SumAggregator {
 constructor() {
   this.sum = 0;
 }
 update(value) {
   this.sum += value;
 }
 getSum() {
   return this.sum;
 }
}

The following query doesn't execute because the SumAggregator class object is not BigQuery-serializable, due to the presence of functions inside of the class.

CREATE TEMP AGGREGATE FUNCTION F(x FLOAT64)
RETURNS FLOAT64
LANGUAGE js
AS r'''

  class SumAggregator {
   constructor() {
     this.sum = 0;
   }

   update(value) {
     this.sum += value;
   }

   getSum() {
     return this.sum;
   }
  }

  export function initialState() {
   return new SumAggregator();
  }

  export function aggregate(agg, value) {
   agg.update(value);
  }

  export function merge(agg1, agg2) {
   agg1.update(agg2.getSum());
  }

  export function finalize(agg) {
   return agg.getSum();
  }

''';

--Error: getSum is not a function
SELECT F(x) AS results FROM UNNEST([1,2,3,4]) AS x;

If you add the serialize and deserialize functions to the preceding query, the query runs because the SumAggregator class object is converted to an object that is BigQuery-serializable and then back to a SumAggregator class object again.

CREATE TEMP AGGREGATE FUNCTION F(x FLOAT64)
RETURNS FLOAT64
LANGUAGE js
AS r'''

  class SumAggregator {
   constructor() {
     this.sum = 0;
   }

   update(value) {
     this.sum += value;
   }

   getSum() {
     return this.sum;
   }
  }

  export function initialState() {
   return new SumAggregator();
  }

  export function aggregate(agg, value) {
   agg.update(value);
  }

  export function merge(agg1, agg2) {
   agg1.update(agg2.getSum());
  }

  export function finalize(agg) {
   return agg.getSum();
  }

  export function serialize(agg) {
   return {sum: agg.getSum()};
  }

  export function deserialize(serialized) {
   var agg = new SumAggregator();
   agg.update(serialized.sum);
   return agg;
  }

''';

SELECT F(x) AS results FROM UNNEST([1,2,3,4]) AS x;

/*-----------------*
 | results         |
 +-----------------+
 | 10.0            |
 *-----------------*/

To learn more about the serialization functions, see Optional JavaScript serialization functions.

Include global variables and custom functions in a JavaScript UDAF

The JavaScript function body can include custom JavaScript code such as JavaScript global variables and custom functions.

Global variables are executed when the JavaScript is loaded into BigQuery and before the initialState function is executed. Global variables might be useful if you need to perform one-time initialization work that shouldn't repeat for each aggregation group, as would be the case with the initialState, aggregate, merge, and finalize functions.

Don't use global variables to store aggregation state. Instead, limit aggregation state to objects passed to exported functions. Only use global variables to cache expensive operations that are not specific to any particular aggregation operation.

In the following query, the SumOfPrimes function calculates a sum, but only prime numbers are included in the calculation. In the JavaScript function body, there are two global variables, primes and maxTested, that are initialized first. In addition, there is a custom function called isPrime that checks if a number is prime.

CREATE TEMP AGGREGATE FUNCTION SumOfPrimes(x INT64)
RETURNS INT64
LANGUAGE js
AS r'''

  var primes = new Set([2]);
  var maxTested = 2;

  function isPrime(n) {
    if (primes.has(n)) {
      return true;
    }
    if (n <= maxTested) {
      return false;
    }
    for (var k = 2; k < n; ++k) {
      if (!isPrime(k)) {
        continue;
      }
      if ((n % k) == 0) {
        maxTested = n;
        return false;
      }
    }
    maxTested = n;
    primes.add(n);
    return true;
  }

  export function initialState() {
    return {sum: 0};
  }

  export function aggregate(state, x) {
    x = Number(x);
    if (isPrime(x)) {
      state.sum += x;
    }
  }

  export function merge(state, partialState) {
    state.sum += partialState.sum;
  }

  export function finalize(state) {
    return state.sum;
  }

''';

-- Call the JavaScript UDAF.
WITH numbers AS (
  SELECT * FROM UNNEST([10, 11, 13, 17, 19, 20]) AS x)
SELECT SumOfPrimes(x) AS sum FROM numbers;

/*-----*
 | sum |
 +-----+
 | 60  |
 *-----*/

Include JavaScript libraries

You can extend your JavaScript UDAFs with the library option in the OPTIONS clause. This option lets you specify external code libraries for the JavaScript UDAF, and then import those libraries with the import declaration.

In the following example, code in bar.js is available to any code in the function body of the JavaScript UDAF:

CREATE TEMP AGGREGATE FUNCTION JsAggFn(x FLOAT64)
RETURNS FLOAT64
LANGUAGE js
OPTIONS (library = ['gs://foo/bar.js'])
AS r'''

  import doInterestingStuff from 'bar.js';

  export function initialState() {
    return ...
  }
  export function aggregate(state, x) {
    var result = doInterestingStuff(x);
    ...
  }
  export function merge(state, partial_state) {
    ...
  }
  export function finalize(state) {
    return ...;
  }

''';

Required JavaScript structure

Unlike a JavaScript UDF, where the function body is free-form JavaScript that runs for every row, the function body for a JavaScript UDAF is a JavaScript module that contains some built-in exported functions, which are invoked at various stages in the aggregation process. Some of these built-in functions are required, while others are optional. You can also add your JavaScript functions.

Required JavaScript aggregation functions

You can include your JavaScript functions, but the JavaScript function body must include the following exportable JavaScript functions:

  • initialState([nonAggregateParam]): returns a JavaScript object that represents an aggregation state in which no rows have been aggregated yet.

  • aggregate(state, aggregateParam[, ...][, nonAggregateParam]): aggregates one row of data, updating state to store the result of the aggregation. Does not return a value.

  • merge(state, partialState, [nonAggregateParam]): merges aggregation state partialState into aggregation state state. This function is used when the engine aggregates different sections of data in parallel and needs to combine the results. Does not return a value.

  • finalize(finalState, [nonAggregateParam]): returns the final result of the aggregate function, given a final aggregation state finalState.

To learn more about the required functions, see Required functions in a JavaScript UDAF.

Optional JavaScript serialization functions

If you want to work with non-serializable aggregation states, the JavaScript UDAF must provide the serialize and deserialize functions. The serialize function converts the aggregation state to a BigQuery-serializable object; the deserialize function converts the BigQuery-serializable object back into an aggregation state.

  • serialize(state): returns a serializable object that contains the information in aggregation state, to be deserialized through the deserialize function.

  • deserialize(serializedState): deserializes serializedState (previously serialized by the serialize function) into an aggregation state that can be passed into the serialize, aggregate, merge, or finalize functions.

To learn more about the built-in JavaScript serialization functions, see Serialization functions for a JavaScript UDAF.

To learn how to serialize and deserialize data with a JavaScript UDAF, see Serialize and deserialize data in a JavaScript UDAF.

Permitted SQL type encodings in a JavaScript UDAF

In JavaScript UDAFs, the following supported GoogleSQL data types represent JavaScript data types as follows:

GoogleSQL
data type
JavaScript
data type
Notes
ARRAY Array An array of arrays is not supported. To get around this limitation, use the Array<Object<Array>> (JavaScript) and ARRAY<STRUCT<ARRAY>> (GoogleSQL) data types.
BIGNUMERIC Number or String Same as NUMERIC.
BOOL Boolean
BYTES Uint8Array
DATE Date
FLOAT64 Number
INT64 BigInt
JSON Various types The GoogleSQL JSON data type can be converted into a JavaScript Object, Array, or other GoogleSQL-supported JavaScript data type.
NUMERIC Number or String If a NUMERIC value can be represented exactly as an IEEE 754 floating-point value (range [-253, 253]), and has no fractional part, it is encoded as a Number data type, otherwise it is encoded as a String data type.
STRING String
STRUCT Object Each STRUCT field is a named property in the Object data type. An unnamed STRUCT field is not supported.
TIMESTAMP Date Date contains a microsecond field with the microsecond fraction of TIMESTAMP.

Call a UDAF

This section describes the various ways that you can call a persistent or temporary UDAF after you create it in BigQuery.

Call a persistent UDAF

You can call a persistent UDAF in the same way that you call a built-in aggregate function. For more information, see Aggregate function calls. You must include the dataset in the function path.

In the following example, the query calls a persistent UDAF that's called WeightedAverage:

SELECT my_project.my_dataset.WeightedAverage(item, weight, 2) AS weighted_average
FROM (
  SELECT 1 AS item, 2.45 AS weight UNION ALL
  SELECT 3 AS item, 0.11 AS weight UNION ALL
  SELECT 5 AS item, 7.02 AS weight
);

A table with the following results is produced:

/*------------------*
 | weighted_average |
 +------------------+
 | 4.5              |
 *------------------*/

Call a temporary UDAF

You can call a temporary UDAF in the same way that you call a built-in aggregate function. For more information, see Aggregate function calls.

The temporary function must be included in a multi-statement query or procedure that contains the UDAF function call.

In the following example, the query calls a temporary UDAF that's called WeightedAverage:

CREATE TEMP AGGREGATE FUNCTION WeightedAverage(...)

-- Temporary UDAF function call
SELECT WeightedAverage(item, weight, 2) AS weighted_average
FROM (
  SELECT 1 AS item, 2.45 AS weight UNION ALL
  SELECT 3 AS item, 0.11 AS weight UNION ALL
  SELECT 5 AS item, 7.02 AS weight
);

A table with the following results is produced:

/*------------------*
 | weighted_average |
 +------------------+
 | 4.5              |
 *------------------*/

Ignore or include rows with NULL values

When a JavaScript UDAF is called with the IGNORE NULLS argument, BigQuery automatically skips over rows for which any aggregate argument evaluates to NULL. Such rows are excluded from the aggregation completely and are not passed to the JavaScript aggregate function. When RESPECT NULLS argument is provided, NULL filtration is disabled, and every row is passed to the JavaScript UDAF, regardless of NULL values.

When neither the IGNORE NULLS nor RESPECT NULLS argument is provided, the default argument is IGNORE NULLS.

The following example illustrates the default NULL behavior, the IGNORE NULLS behavior, and the RESPECT NULLS behavior:

CREATE TEMP AGGREGATE FUNCTION SumPositive(x FLOAT64)
RETURNS FLOAT64
LANGUAGE js
AS r'''

  export function initialState() {
    return {sum: 0}
  }
  export function aggregate(state, x) {
    if (x == null) {
      // Use 1000 instead of 0 as placeholder for null so
      // that NULL values passed are visible in the result.
      state.sum += 1000;
      return;
    }
    if (x > 0) {
      state.sum += x;
    }
  }
  export function merge(state, partialState) {
    state.sum += partialState.sum;
  }
  export function finalize(state) {
    return state.sum;
  }

''';

-- Call the JavaScript UDAF.
WITH numbers AS (
  SELECT * FROM UNNEST([1.0, 2.0, NULL]) AS x)
SELECT
  SumPositive(x) AS sum,
  SumPositive(x IGNORE NULLS) AS sum_ignore_nulls,
  SumPositive(x RESPECT NULLS) AS sum_respect_nulls
FROM numbers;

/*-----+------------------+-------------------*
 | sum | sum_ignore_nulls | sum_respect_nulls |
 +-----+------------------+-------------------+
 | 3.0 | 3.0              | 1003.0            |
 *-----+------------------+-------------------*/

Delete a UDAF

This section describes the various ways that you can delete a persistent or temporary UDAF after you created it in BigQuery.

Delete a persistent UDAF

To delete a persistent UDAF, use the DROP FUNCTION statement. You must include the dataset in the function path.

In the following example, the query deletes a persistent UDAF that's called WeightedAverage:

DROP FUNCTION IF EXISTS my_project.my_dataset.WeightedAverage;

Delete a temporary UDAF

To delete a temporary UDAF, use the DROP FUNCTION statement.

In the following example, the query deletes a temporary UDAF that's called WeightedAverage:

DROP FUNCTION IF EXISTS WeightedAverage;

A temporary UDAF expires as soon as the query finishes. The UDAF doesn't need to be deleted unless you want to remove it early from a multi-statement query or procedure.

List UDAFs

UDAFs are a type of routine. To list all of the routines in a dataset, see List routines.

Performance tips

If you want to improve the performance of your queries, consider the following:

  • Prefilter your input. Processing data in JavaScript is more expensive than in SQL, so it is best to filter the input as much as possible in SQL first.

    The following query is less efficient because it filters the input by using x > 0 in the UDAF call:

    SELECT JsFunc(x) FROM t;
    

    The following query is more efficient because it prefilters the input by using WHERE x > 0 before the UDAF is called:

    SELECT JsFunc(x) FROM t WHERE x > 0;
    
  • Use built-in aggregate functions instead of JavaScript when possible. Re-implementing a built-in aggregate function in JavaScript is slower than calling a built-in aggregate function that does the same thing.

    The following query is less efficient because it implements a UDAF:

    SELECT SumSquare(x) FROM t;
    

    The following query is more efficient because it implements a built-in function that produces the same results as the previous query:

    SELECT SUM(x*x) FROM t;
    
  • JavaScript UDAFs are appropriate for more complex aggregation operations, which can't be expressed through built-in functions.

  • Use memory efficiently. The JavaScript processing environment has limited memory available for each query. JavaScript UDAF queries that accumulate too much local state might fail due to memory exhaustion. Be especially mindful about minimizing the size of aggregation state objects and avoid aggregation states which accumulate large numbers of rows.

    The following query is not efficient because the aggregate function uses an unbounded amount of memory when the number of rows processed gets large.

    export function initialState() {
      return {rows: []};
    }
    export function aggregate(state, x) {
      state.rows.push(x);
    }
    ...
    
  • Use partitioned tables when possible. JavaScript UDAFs typically run more efficiently when querying against a partitioned table compared to a non-partitioned table, because a partitioned table stores data in many smaller files compared to a non-partitioned table, thus enabling higher parallelism.

Limitations

  • UDAFs have the same limitations that apply to UDFs. For details, see UDF limitations.

  • Only literals, query parameters, and script variables can be passed in as non-aggregate arguments for a UDAF.

  • The use of the ORDER BY clause in a JavaScript UDAF function call is unsupported.

    SELECT MyUdaf(x ORDER BY y) FROM t; -- Error: ORDER BY is unsupported.
    

Pricing

UDAFs are billed using the standard BigQuery pricing model.

Quotas and limits

UDAFs have the same quotas and limits that apply to UDFs. For information about UDF quotas, see Quotas and limits.