Window functions can only be used in the SELECT clause. To share OVER specifications between functions, use the statement’s WINDOW clause and use the OVER window-name syntax.

General-Purpose Window Functions

The table below shows the available general window functions.

Aggregate Window Functions

All aggregate functions can be used in a windowing context.

Evaluation

Windowing works by breaking a relation up into independent partitions, ordering those partitions, and then computing a new column for each row as a function the nearby values. Some window functions depend only on the partition boundary and the ordering, but a few (including all the aggregates) also use a frame. Frames are specified as a number of rows on either side (preceding or following) of the current row. The distance can either be specified as a number of rows or a range of values using the partition’s ordering value and a distance.

The full syntax is shown in the diagram at the top of the page, and this diagram visually illustrates computation environment:

Partition and Ordering

Partitioning breaks the relation up into independent, unrelated pieces. Partitioning is optional, and if none is specified then the entire relation is treated as a single partition. Window functions cannot access values outside of the partition containing the row they are being evaluated at.

Ordering is also optional, but without it the results are not well-defined. Each partition is ordered using the same ordering clause.

Here is a table of power generation data. After partitioning by plant and ordering by date, it will have this layout:

In what follows, we shall use this table (or small sections of it) to illustrate various pieces of window function evaluation.

The simplest window function is ROW_NUMBER(). This function just computes the 1-based row number within the partition using the query:

SELECT "Plant", "Date", ROW_NUMBER AS "Row"
FROM "History"
ORDER BY 1, 2

The result will be

Note that even though the function is computed with an ORDER BY clause, the result does not have to be sorted, so the SELECT also needs to be explicitly sorted if that is desired.

Framing

Framing specifies a set of rows relative to each row where the function is evaluated. The distance from the current row is given as an expression either PRECEDING or FOLLOWING the current row. This distance can either be specified as an integral number of ROWS or as a RANGE delta expression from the value of the ordering expression. For a RANGE specification, there must be only one ordering expression, and it has to support addition and subtraction (i.e., numbers or INTERVALs). The default values for frames are from UNBOUNDED PRECEDING to CURRENT ROW. It is invalid for a frame to start after it ends.

ROW Framing

Here is a simple ROW frame query, using an aggregate function:

SELECT points,
    SUM(points) OVER (
        ROWS BETWEEN 1 PRECEDING
                 AND 1 FOLLOWING) we
FROM results

This query computes the SUM of each point and the points on either side of it:

Notice that at the edge of the partition, there are only two values added together. This is because frames are cropped to the edge of the partition.

RANGE Framing

Returning to the power data, suppose the data is noisy. We might want to compute a 7 day moving average for each plant to smooth out the noise. To do this, we can use this window query:

SELECT "Plant", "Date",
    AVG("MWh") OVER (
        PARTITION BY "Plant"
        ORDER BY "Date" ASC
        RANGE BETWEEN INTERVAL 3 DAYS PRECEDING
                  AND INTERVAL 3 DAYS FOLLOWING)
        AS "MWh 7-day Moving Average"
FROM "Generation History"
ORDER BY 1, 2

This query partitions the data by Plant (to keep the different power plants’ data separate), orders each plant’s partition by Date (to put the energy measurements next to each other), and uses a RANGE frame of three days on either side of each day for the AVG (to handle any missing days). This is the result:

WINDOW Clauses

Multiple different OVER clauses can be specified in the same SELECT, and each will be computed separately. Often, however, we want to use the same layout for multiple window functions. The WINDOW clause can be used to define a named window that can be shared between multiple window functions:

SELECT "Plant", "Date",
    MIN("MWh") OVER seven AS "MWh 7-day Moving Minimum",
    AVG("MWh") OVER seven AS "MWh 7-day Moving Average",
    MAX("MWh") OVER seven AS "MWh 7-day Moving Maximum"
FROM "Generation History"
WINDOW seven AS (
    PARTITION BY "Plant"
    ORDER BY "Date" ASC
    RANGE BETWEEN INTERVAL 3 DAYS PRECEDING
              AND INTERVAL 3 DAYS FOLLOWING)
ORDER BY 1, 2

The three window functions will also share the data layout, which will improve performance.

Box and Whisker Queries

All aggregates can be used as windowing functions, including the complex statistical functions. These function implementations have been optimised for windowing, and we can use the window syntax to write queries that generate the data for moving box-and-whisker plots:

SELECT "Plant", "Date",
    MIN("MWh") OVER seven AS "MWh 7-day Moving Minimum",
    QUANTILE_CONT("MWh", [0.25, 0.5, 0.75]) OVER seven
        AS "MWh 7-day Moving IQR",
    MAX("MWh") OVER seven AS "MWh 7-day Moving Maximum",
FROM "Generation History"
WINDOW seven AS (
    PARTITION BY "Plant"
    ORDER BY "Date" ASC
    RANGE BETWEEN INTERVAL 3 DAYS PRECEDING
              AND INTERVAL 3 DAYS FOLLOWING)
ORDER BY 1, 2