*** DRAFT ***
Window Functions

1. Introduction to Window Functions

A window function is a special SQL function where the input values are taken from a "window" of one or more rows in the results set of a SELECT statement.

window-function-invocation:

syntax diagram window-function-invocation

expr:

filter:

window-defn:

Window functions are distinguished from ordinary SQL functions by the presence of an OVER clause. If a function invocation has an OVER clause then it is a window function, and if lacks a OVER clause it is an ordinary function. Window functions might also have a FILTER clause in between the function and the OVER clause.

Unlike ordinary functions, window functions cannot use the DISTINCT keyword. Also, Window functions may only appears in the result set and in the ORDER BY clause of a SELECT statement.

The following simple table is used to demonstrate how window functions work:

CREATE TABLE t0(x INTEGER PRIMARY KEY, y TEXT);
INSERT INTO t0 VALUES (1, 'aaa'), (2, 'ccc'), (3, 'bbb');

An example of using window functions:

-- The following SELECT statement returns:
-- 
--   x | y | row_number
-----------------------
--   1 | aaa | 1         
--   2 | ccc | 3         
--   3 | bbb | 2         
-- 
SELECT x, y, row_number() OVER (ORDER BY y) AS row_number FROM t0 ORDER BY x;

The example uses the built-in window function row_number(). The row_number() window function assigns a monotonically increasing integer to each row in order of the "ORDER BY" clause within the window-defn (in this case "ORDER BY y"). Note that this does not affect the order in which results are returned from the overall query. The order of the final output is still governed by the ORDER BY clause attached to the SELECT statement (in this case "ORDER BY x").

Named window-defn clauses may also be added to a SELECT statement using a WINDOW clause and then referred to by name within window function invocations. For example, the following SELECT statement contains two named window-defs clauses, "win1" and "win2":

SELECT x, y, row_number() OVER win1, rank() OVER win2 
FROM t0 
WINDOW win1 AS (ORDER BY y RANGE BETWEEN UNBOUNDED PRECEDING AND CURRENT ROW),
       win2 AS (PARTITION BY y ORDER BY x)
ORDER BY x;

The WINDOW clause, when one is present, comes after any HAVING clause and before any ORDER BY.

2. Aggregate Window Functions

The examples in this section all assume that the database is populated as follows:

CREATE TABLE t1(a INTEGER PRIMARY KEY, b, c);
INSERT INTO t1 VALUES   (1, 'A', 'one'  ),
                        (2, 'B', 'two'  ),
                        (3, 'C', 'three'),
                        (4, 'D', 'one'  ),
                        (5, 'E', 'two'  ),
                        (6, 'F', 'three'),
                        (7, 'G', 'one'  );

An aggregate window function is similar to an aggregate function, except adding it to a query does not change the number of rows returned. Instead, for each row the result of the aggregate window function is as if the corresponding aggregate were run over all rows in the "window frame".

-- The following SELECT statement returns:
-- 
--   a | b | group_concat
-------------------------
--   1 | A | A.B         
--   2 | B | A.B.C       
--   3 | C | B.C.D       
--   4 | D | C.D.E       
--   5 | E | D.E.F       
--   6 | F | E.F.G       
--   7 | G | F.G         
-- 
SELECT a, b, group_concat(b, '.') OVER (
  ORDER BY a ROWS BETWEEN 1 PRECEDING AND 1 FOLLOWING
) AS group_concat FROM t1;

In the example above, the window frame consists of all rows between the previous row ("1 PRECEDING") and the following row ("1 FOLLOWING"), inclusive, where rows are sorted according to the ORDER BY clause in the window-defn (in this case "ORDER BY a"). For example, the frame for the row with (a=3) consists of rows (2, 'B', 'two'), (3, 'C', 'three') and (4, 'D', 'one'). The result of group_concat(b, '.') for that row is therefore 'B.C.D'.

This means that, after sorting the rows returned by the SELECT according to the ORDER BY clause in the window-definition, the window frame consists of all rows between the first row and the last row with the same values as the current row for all ORDER BY expressions. This implies that rows that have the same values for all ORDER BY expressions will also have the same value for the result of the window function (as the window frame is the same). For example:

-- The following SELECT statement returns:
-- 
--   a | b | c | group_concat
-----------------------------
--   1 | A | one   | A.D.G       
--   2 | B | two   | A.D.G.C.F.B.E
--   3 | C | three | A.D.G.C.F   
--   4 | D | one   | A.D.G       
--   5 | E | two   | A.D.G.C.F.B.E
--   6 | F | three | A.D.G.C.F   
--   7 | G | one   | A.D.G       
-- 
SELECT a, b, c, 
       group_concat(b, '.') OVER (ORDER BY c) AS group_concat 
FROM t1 ORDER BY a;

All of SQLite's aggregate functions may be used as aggregate window functions. It is also possible to create user-defined aggregate window functions.

2.1. Frame Specifications

frame-spec:

syntax diagram frame-spec

expr:

The frame-spec determines which output rows are read by an aggregate window function. The frame-spec consists of three parts:

The ending frame boundary can be omitted, in which case it defaults to CURRENT ROW.

The default frame-spec is:

RANGE BETWEEN UNBOUNDED PRECEDING AND CURRENT ROW

The default means that aggregate window functions read all rows from the beginning of the partition up to and including the current row and its peers.

If the frame type is RANGE, then rows with the same values for all ORDER BY expressions are considered "peers". Or, if there are no ORDER BY terms, all rows are peers. Rows that are peers always have the same window frames.

There are five options for frame boundaries:

Frame Boundary Description
UNBOUNDED PRECEDING The start of the frame is the first row in the set.
<expr> PRECEDING <expr> is a constant expression that evaluates to a non-negative integer value. The start or end of the frame is <expr> rows before the current row. "0 PRECEDING" is the same as "CURRENT ROW". This frame boundary type may only be used with ROWS frames.
CURRENT ROW The current row. For RANGE frame types, all peers of the current row are also included in the window frame, regardless of whether CURRENT ROW is used as the starting or ending frame boundary.
<expr> FOLLOWING <expr> is a constant expression that must evaluate to a non-negative integer value. The start or end of the frame is <expr> rows after the current row. "0 FOLLOWING" is the same as "CURRENT ROW". This frame boundary type may only be used with ROWS frames.
UNBOUNDED FOLLOWING The end of the frame is the last row in the set.

The ending frame boundary must not take a form that appears higher in the above list than the starting frame boundary.

In the following example, the window frame for each row consists of all rows from the current row to the end of the set, where rows are sorted according to "ORDER BY a".

-- The following SELECT statement returns:
-- 
--   c     | a | b | group_concat
---------------------------------
--   one   | 1 | A | A.D.G.C.F.B.E
--   one   | 4 | D | D.G.C.F.B.E 
--   one   | 7 | G | G.C.F.B.E   
--   three | 3 | C | C.F.B.E     
--   three | 6 | F | F.B.E       
--   two   | 2 | B | B.E         
--   two   | 5 | E | E           
-- 
SELECT c, a, b, group_concat(b, '.') OVER (
  ORDER BY c, a ROWS BETWEEN CURRENT ROW AND UNBOUNDED FOLLOWING
) AS group_concat
FROM t1 ORDER BY c, a;

2.2. The PARTITION BY Clause

A window-defn may include a PARTITION BY clause. If so, the rows returned by the SELECT statement are divided into groups - partitions - with the same values for each PARTITION BY expression, and then window-function processing is performed separately for each partition. This is similar to the way the rows are grouped by the GROUP BY clause of an aggregate query.

For example:

-- The following SELECT statement returns:
-- 
--   c     | a | b | group_concat
---------------------------------
--   one   | 1 | A | A.D.G       
--   one   | 4 | D | D.G         
--   one   | 7 | G | G           
--   three | 3 | C | C.F         
--   three | 6 | F | F           
--   two   | 2 | B | B.E         
--   two   | 5 | E | E           
-- 
SELECT c, a, b, group_concat(b, '.') OVER (
  PARTITION BY c ORDER BY a RANGE BETWEEN CURRENT ROW AND UNBOUNDED FOLLOWING
) AS group_concat
FROM t1 ORDER BY c, a;

2.3. The FILTER Clause

filter:

syntax diagram filter

expr:

If a FILTER clause is provided, then only rows for which the expr is true are included in the window frame. The aggregate window still returns a value for every row, but those for which the FILTER expression evaluates to other than true are not included in the window frame for any row. For example:

-- The following SELECT statement returns:
-- 
--   c     | a | b | group_concat
---------------------------------
--   one   | 1 | A | A           
--   two   | 2 | B | A           
--   three | 3 | C | A.C         
--   one   | 4 | D | A.C.D       
--   two   | 5 | E | A.C.D       
--   three | 6 | F | A.C.D.F     
--   one   | 7 | G | A.C.D.F.G   
-- 
SELECT c, a, b, group_concat(b, '.') FILTER (WHERE c!='two') OVER (
  ORDER BY a
) AS group_concat
FROM t1 ORDER BY a;

3. Built-in Window Functions

As well as aggregate window functions, SQLite features a set of built-in window functions based on those supported by PostgreSQL.

Built-in window functions honor any PARTITION BY clause in the same way as aggregate window functions - each selected row is assigned to a partition and each partition is processed separately. The ways in which any ORDER BY clause affects each built-in window function is described below. Some of the window functions (rank(), dense_rank(), percent_rank() and ntile()) use the concept of "peer groups" (rows within the same partition that have the same values for all ORDER BY expressions). In these cases, it does not matter whether the frame-spec specifies ROWS or RANGE - for the purposes of built-in window function processing, rows with the same values for all ORDER BY expressions are considered peers regardless of the frame type.

Most built-in window functions ignore the frame-spec, the exceptions being first_value(), last_value() and nth_value(). It is a syntax error to specify a FILTER clause as part of a built-in window function invocation.

SQLite supports the following 11 built-in window functions:

row_number()

The number of the row within the current partition. Rows are numbered starting from 1 in the order defined by the ORDER BY clause in the window definition, or in arbitrary order otherwise.

rank()

The row_number() of the first peer in each group - the rank of the current row with gaps. If there is no ORDER BY clause, then all rows are considered peers and this function always returns 1.

dense_rank()

The number of the current row's peer group within its partition - the rank of the current row without gaps. Partitions are numbered starting from 1 in the order defined by the ORDER BY clause in the window definition. If there is no ORDER BY clause, then all rows are considered peers and this function always returns 1.

percent_rank()

Despite the name, this function always returns a value between 0.0 and 1.0 equal to (rank - 1)/(partition-rows - 1), where rank is the value returned by built-in window function rank() and partition-rows is the total number of rows in the partition. If the partition contains only one row, this function returns 0.0.

cume_dist()

The cumulative distribution. Calculated as row-number/partition-rows, where row-number is the value returned by row_number() for the last peer in the group and partition-rows the number of rows in the partition.

ntile(N)

Argument N is handled as an integer. This function divides the partition into N groups as evenly as possible and assigns an integer between 1 and N to each group, in the order defined by the ORDER BY clause, or in arbitrary order otherwise. If necessary, larger groups occur first. This function returns the integer value assigned to the group that the current row is a part of.

lag(expr)
lag(expr, offset)
lag(expr, offset, default)

The first form of the lag() function returns the result of evaluating expression expr against the previous row in the partition. Or, if there is no previous row (because the current row is the first), NULL.

If the offset argument is provided, then it must be a non-negative integer. In this case the value returned is the result of evaluating expr against the row offset rows before the current row within the partition. If offset is 0, then expr is evaluated against the current row. If there is no row offset rows before the current row, NULL is returned.

If default is also provided, then it is returned instead of NULL if row identified by offset does not exist.

lead(expr)
lead(expr, offset)
lead(expr, offset, default)

The first form of the lead() function returns the result of evaluating expression expr against the next row in the partition. Or, if there is no next row (because the current row is the last), NULL.

If the offset argument is provided, then it must be a non-negative integer. In this case the value returned is the result of evaluating expr against the row offset rows after the current row within the partition. If offset is 0, then expr is evaluated against the current row. If there is no row offset rows after the current row, NULL is returned.

If default is also provided, then it is returned instead of NULL if row identified by offset does not exist.

first_value(expr)

This built-in window function calculates the window frame for each row in the same way as an aggregate window function. It returns the value of expr evaluated against the first row in the window frame for each row.

last_value(expr)

This built-in window function calculates the window frame for each row in the same way as an aggregate window function. It returns the value of expr evaluated against the last row in the window frame for each row.

nth_value(expr, N)

This built-in window function calculates the window frame for each row in the same way as an aggregate window function. It returns the value of expr evaluated against the row N of the window frame. Rows are numbered within the window frame starting from 1 in the order defined by the ORDER BY clause if one is present, or in arbitrary order otherwise. If there is no Nth row in the partition, then NULL is returned.

The examples in this section all assume the following data:

CREATE TABLE t2(a, b);
INSERT INTO t2 VALUES('a', 'one'), 
                     ('a', 'two'), 
                     ('a', 'three'), 
                     ('b', 'four'), 
                     ('c', 'five'), 
                     ('c', 'six');

The following example illustrates the behaviour of the five ranking functions - row_number(), rank(), dense_rank(), percent_rank() and cume_dist().

-- The following SELECT statement returns:
-- 
--   a | row_number | rank | dense_rank | percent_rank | cume_dist
------------------------------------------------------------------
--   a |          1 |    1 |          1 |          0.0 |       0.5
--   a |          2 |    1 |          1 |          0.0 |       0.5
--   a |          3 |    1 |          1 |          0.0 |       0.5
--   b |          4 |    4 |          2 |          0.6 |       0.66
--   c |          5 |    5 |          3 |          0.8 |       1.0
--   c |          6 |    5 |          3 |          0.8 |       1.0
-- 
SELECT a                        AS a,
       row_number() OVER win    AS row_number,
       rank() OVER win          AS rank,
       dense_rank() OVER win    AS dense_rank,
       percent_rank() OVER win  AS percent_rank,
       cume_dist() OVER win     AS cume_dist
FROM t2
WINDOW win AS (ORDER BY a);

The example below uses ntile() to divide the six rows into two groups (the ntile(2) call) and into four groups (the ntile(4) call). For ntile(2), there are three rows assigned to each group. For ntile(4), there are two groups of two and two groups of one. The larger groups of two appear first.

-- The following SELECT statement returns:
-- 
--   a | b     | ntile_2 | ntile_4
----------------------------------
--   a | one   |       1 |       1
--   a | two   |       1 |       1
--   a | three |       1 |       2
--   b | four  |       2 |       2
--   c | five  |       2 |       3
--   c | six   |       2 |       4
-- 
SELECT a                        AS a,
       b                        AS b,
       ntile(2) OVER win        AS ntile_2,
       ntile(4) OVER win        AS ntile_4
FROM t2
WINDOW win AS (ORDER BY a);

The next example demonstrates lag(), lead(), first_value(), last_value() and nth_value(). The frame-spec is ignored by both lag() and lead(), but respected by first_value(), last_value() and nth_value().

-- The following SELECT statement returns:
-- 
--   b | lead | lag  | first_value | last_value | nth_value_3
-------------------------------------------------------------
--   A | C    | NULL | A           | A          | NULL       
--   B | D    | A    | A           | B          | NULL       
--   C | E    | B    | A           | C          | C          
--   D | F    | C    | A           | D          | C          
--   E | G    | D    | A           | E          | C          
--   F | n/a  | E    | A           | F          | C          
--   G | n/a  | F    | A           | G          | C          
-- 
SELECT b                          AS b,
       lead(b, 2, 'n/a') OVER win AS lead,
       lag(b) OVER win            AS lag,
       first_value(b) OVER win    AS first_value,
       last_value(b) OVER win     AS last_value,
       nth_value(b, 3) OVER win   AS nth_value_3
FROM t1
WINDOW win AS (ORDER BY b ROWS BETWEEN UNBOUNDED PRECEDING AND CURRENT ROW)

4. User-Defined Aggregate Window Functions

User-defined aggregate window functions may be created using the sqlite3_create_window_function() API. Implementing an aggregate window function is very similar to an ordinary aggregate function. Any user-defined aggregate window function may also be used as an ordinary aggregate. To implement a user-defined aggregate window function the application must supply four callback functions:

Callback Description
xStep This method is required by both window aggregate and legacy aggregate function implementations. It is invoked to add a row to the current window. The function arguments, if any, corresponding to the row being added are passed to the implementation of xStep.
xFinal This method is required by both window aggregate and legacy aggregate function implementations. It is invoked to return the current value of the aggregate (determined by the contents of the current window), and to free any resources allocated by earlier calls to xStep.
xValue This method is only required window aggregate functions, not legacy aggregate function implementations. It is invoked to return the current value of the aggregate. Unlike xFinal, the implementation should not delete any context.
xInverse This method is only required window aggregate functions, not legacy aggregate function implementations. It is invoked to remove a row from the current window. The function arguments, if any, correspond to the row being removed.

The C code below implements a simple window aggregate function named sumint(). This works in the same way as the built-in sum() function, except that it throws an exception if passed an argument that is not an integer value.

/*
** xStep for sumint().
**
** Add the value of the argument to the aggregate context (an integer).
*/
static void sumintStep(
  sqlite3_context *ctx, 
  int nArg, 
  sqlite3_value *apArg[]
){
  sqlite3_int64 *pInt;

  assert( nArg==1 );
  if( sqlite3_value_type(apArg[0])!=SQLITE_INTEGER ){
    sqlite3_result_error(ctx, "invalid argument", -1);
    return;
  }
  pInt = (sqlite3_int64*)sqlite3_aggregate_context(ctx, sizeof(sqlite3_int64));
  if( pInt ){
    *pInt += sqlite3_value_int64(apArg[0]);
  }
}

/*
** xInverse for sumint().
**
** This does the opposite of xStep() - subtracts the value of the argument
** from the current context value. The error checking can be omitted from
** this function, as it is only ever called after xStep() (so the aggregate
** context has already been allocated) and with a value that has already
** been passed to xStep() without error (so it must be an integer).
*/
static void sumintInverse(
  sqlite3_context *ctx, 
  int nArg, 
  sqlite3_value *apArg[]
){
  sqlite3_int64 *pInt;
  assert( sqlite3_value_type(apArg[0])==SQLITE_INTEGER );
  pInt = (sqlite3_int64*)sqlite3_aggregate_context(ctx, sizeof(sqlite3_int64));
  *pInt -= sqlite3_value_int64(apArg[0]);
}

/*
** xFinal for sumint().
**
** Return the current value of the aggregate window function. Because
** this implementation does not allocate any resources beyond the buffer
** returned by sqlite3_aggregate_context, which is automatically freed
** by the system, there are no resources to free. And so this method is
** identical to xValue().
*/
static void sumintFinal(sqlite3_context *ctx){
  sqlite3_int64 res = 0;
  sqlite3_int64 *pInt;
  pInt = (sqlite3_int64*)sqlite3_aggregate_context(ctx, 0);
  if( pInt ) res = *pInt;
  sqlite3_result_int64(ctx, res);
}

/*
** xValue for sumint().
**
** Return the current value of the aggregate window function. Because
*/
static void sumintValue(sqlite3_context *ctx){
  sqlite3_int64 res = 0;
  sqlite3_int64 *pInt;
  pInt = (sqlite3_int64*)sqlite3_aggregate_context(ctx, 0);
  if( pInt ) res = *pInt;
  sqlite3_result_int64(ctx, res);
}

/*
** Register sumint() window aggregate with database handle db. 
*/
int register_sumint(sqlite3 *db){
  return sqlite3_create_window_function(db, "sumint", 1, SQLITE_UTF8, 0,
      sumintStep, sumintFinal, sumintValue, sumintInverse, 0
  );
}

The following example uses the sumint() function implemented by the above C code. For each row, the window consists of the preceding row (if any), the current row and the following row (again, if any):

CREATE TABLE t3(x, y);
INSERT INTO t3 VALUES('a', 4),
                     ('b', 5),
                     ('c', 3),
                     ('d', 8),
                     ('e', 1);

-- Assuming the database is populated using the above script, the 
-- following SELECT statement returns:
-- 
--   x | sum_y
--------------
--   a | 9    
--   b | 12   
--   c | 16   
--   d | 12   
--   e | 9    
-- 
SELECT x, sumint(y) OVER (
  ORDER BY x ROWS BETWEEN 1 PRECEDING AND 1 FOLLOWING
) AS sum_y
FROM t3 ORDER BY x;

In processing the query above, SQLite invokes the sumint callbacks as follows:

  1. xStep(4) - add "4" to the current window.
  2. xStep(5) - add "5" to the current window.
  3. xValue() - invoke xValue() to obtain the value of sumint() for the row with (x='a'). The window currently consists of values 4 and 5, and so the result is 9.
  4. xStep(3) - add "3" to the current window.
  5. xValue() - invoke xValue() to obtain the value of sumint() for the row with (x='b'). The window currently consists of values 4, 5 and 3, and so the result is 12.
  6. xInverse(4) - remove "4" from the window.
  7. xStep(8) - add "8" to the current window. The window now consists of values 5, 3 and 8.
  8. xValue() - invoked to obtain the value for the row with (x='c'). In this case, 16.
  9. xInverse(5) - remove value "5" from the window.
  10. xStep(1) - add value "1" to the window.
  11. xValue() - invoked to obtain the value for row (x='d').
  12. xInverse(3) - remove value "3" from the window. The window now contains values 8 and 1 only.
  13. xValue() - invoked to obtain the value for row (x='d'). 9.

5. History

Window function support was added to SQLite with release version 3.25.0 (pending). The SQLite developers used the PostgreSQL window function documentation as their primary reference for how window functions ought to behave. Many test cases have been run against PostgreSQL to ensure that window functions operate the same way in both SQLite and PostgreSQL.

*** DRAFT ***