Window functions

info

Apache Druid supports two query languages: Druid SQL and native queries. This document describes the SQL language.

Window functions are an experimental feature. Development and testing are still at early stage. Feel free to try window functions and provide your feedback. Windows functions are not currently supported by multi-stage-query engine so you cannot use them in SQL-based ingestion.

Set the context parameter enableWindowing: true to use window functions.

Window functions in Apache Druid produce values based upon the relationship of one row within a window of rows to the other rows within the same window. A window is a group of related rows within a result set. For example, rows with the same value for a specific dimension.

Window functions in Druid require a GROUP BY statement. Druid performs the row-level aggregations for the GROUP BY before performing the window function calculations.

The following example organizes results with the same channel value into windows. For each window, the query returns the rank of each row in ascending order based upon its changed value.

SELECT FLOOR(__time TO DAY) AS event_time,
    channel,
    ABS(delta) AS change,
    RANK() OVER w AS rank_value
FROM wikipedia
WHERE channel in ('#kk.wikipedia', '#lt.wikipedia')
AND '2016-06-28' > FLOOR(__time TO DAY) > '2016-06-26'
GROUP BY channel, ABS(delta), __time
WINDOW w AS (PARTITION BY channel ORDER BY ABS(delta) ASC)

View results

Window functions are similar to aggregation functions.

You can use the OVER clause to treat other Druid aggregation functions as window functions. For example, the sum of a value for rows within a window.

Window functions support aliasing.

Window function syntax

You can write a window function in Druid using either syntax below. The second syntax shows a window alias to reference a window that you can reuse.

window_function() OVER (
  [PARTITION BY partitioning expression]
  [ORDER BY order expression]
  [[ROWS, RANGE] BETWEEN range start AND range end])
FROM table
GROUP BY dimensions

window_function() OVER w
FROM table
WINDOW w AS ([PARTITION BY partitioning expression] [ORDER BY order expression]
  [[ROWS, RANGE] BETWEEN range start AND range end])
GROUP BY dimensions

The OVER clause defines the query windows for window functions as follows:

• PARTITION BY indicates the dimension that defines window boundaries
• ORDER BY specifies the order of the rows within the windows

An empty OVER clause or the absence of a PARTITION BY clause indicates that all data belongs to a single window.

In the following example, the following OVER clause example sets the window dimension to channel and orders the results by the absolute value of delta ascending:

...
RANK() OVER (PARTITION BY channel ORDER BY ABS(delta) ASC)
...

Window frames, set in ROWS and RANGE expressions, limit the set of rows used for the windowed aggregation.

ROWS and RANGE accept the following values for range start and range end:

• UNBOUNDED PRECEDING: from the beginning of the window as ordered by the order expression
• N ROWS PRECEDING: N rows before the current row as ordered by the order expression
• CURRENT ROW: the current row
• N ROWS FOLLOWING: N rows after the current row as ordered by the order expression
• UNBOUNDED FOLLOWING: to the end of the window as ordered by the order expression

See Example with window frames for more detail.

Druid applies the GROUP BY dimensions before calculating all non-window aggregation functions. Then it applies the window function over the aggregated results.

note

Sometimes windows are called partitions. However, the partitioning for window functions are a shuffle (partition) of the result set created at query time and is not to be confused with Druid’s segment partitioning feature which partitions data at ingest time.

ORDER BY windows

When the window definition only specifies ORDER BY and not PARTITION BY, it sorts the aggregate data set and applies the function in that order.

The following query uses ORDER BY SUM(delta) DESC to rank user hourly activity from the most changed the least changed within an hour:

SELECT
    TIME_FLOOR(__time, 'PT1H') as time_hour, 
    channel, 
    user,
    SUM(delta) net_user_changes,
    RANK() OVER (ORDER BY SUM(delta) DESC) AS editing_rank
FROM "wikipedia"
WHERE channel IN ('#kk.wikipedia', '#lt.wikipedia')
  AND __time BETWEEN '2016-06-27' AND '2016-06-28'
GROUP BY TIME_FLOOR(__time, 'PT1H'), channel, user
ORDER BY 5

View results

PARTITION BY windows

When a window only specifies PARTITION BY partition expression, Druid calculates the aggregate window function over all the rows that share a value within the selected dataset.

The following example demonstrates a query that uses two different windows—PARTITION BY channel and PARTITION BY user—to calculate the total activity in the channel and total activity by the user so that they can be compared to individual hourly activity:

SELECT
    TIME_FLOOR(__time, 'PT1H') as time_hour,
    channel,
    user,
    SUM(delta) AS hourly_user_changes,
    SUM(SUM(delta)) OVER (PARTITION BY user) AS total_user_changes,
    SUM(SUM(delta)) OVER (PARTITION BY channel) AS total_channel_changes
FROM "wikipedia"
WHERE channel IN ('#kk.wikipedia', '#lt.wikipedia')
  AND __time BETWEEN '2016-06-27' AND '2016-06-28'
GROUP BY TIME_FLOOR(__time, 'PT1H'), 2, 3
ORDER BY channel, TIME_FLOOR(__time, 'PT1H'), user

View results

In this example, the dataset is filtered for a single day. Therefore the window function results represent the total activity for the day, for the user and for the channel dimensions respectively.

This type of result helps you analyze the impact of an individual user’s hourly activity:

• the impact to the channel by comparing hourly_user_changes to total_channel_changes
• the impact of each user over the channel by total_user_changes to total_channel_changes
• the progress of each user’s individual activity by comparing hourly_user_changes to total_user_changes

Window frame guardrails

Druid has guardrail logic to prevent you from executing window function queries with window frame expressions that might return unexpected results.

For example:

• You cannot set expressions as bounds for window frames.
• You cannot use two FOLLOWING expressions in the window frame. For example: ROWS BETWEEN 2 ROWS FOLLOWING and 3 ROWS FOLLOWING.
• You can only use a RANGE frames when both endpoints are unbounded or current row.

If you write a query that violates one of these conditions, Druid throws an error: “The query contains a window frame which may return incorrect results. To disregard this warning, set windowingStrictValidation to false in the query context.”

Window function reference

Examples

The following example illustrates all of the built-in window functions to compare the number of characters changed per event for a channel in the Wikipedia data set.

SELECT FLOOR(__time TO DAY) AS event_time,
    channel,
    ABS(delta) AS change,
    ROW_NUMBER() OVER w AS row_no,
    RANK() OVER w AS rank_no,
    DENSE_RANK() OVER w AS dense_rank_no,
    PERCENT_RANK() OVER w AS pct_rank,
    CUME_DIST() OVER w AS cumulative_dist,
    NTILE(4) OVER w AS ntile_val,
    LAG(ABS(delta), 1, 0) OVER w AS lag_val,
    LEAD(ABS(delta), 1, 0) OVER w AS lead_val,
    FIRST_VALUE(ABS(delta)) OVER w AS first_val,
    LAST_VALUE(ABS(delta)) OVER w AS last_val
FROM wikipedia
WHERE channel IN ('#kk.wikipedia', '#lt.wikipedia')
GROUP BY channel, ABS(delta), FLOOR(__time TO DAY) 
WINDOW w AS (PARTITION BY channel ORDER BY ABS(delta) ASC)

View results

The following example demonstrates applying the SUM() function over the values in a window to calculate the cumulative changes to a channel over time:

SELECT
    FLOOR(__time TO MINUTE) as "time",
    channel,
    ABS(delta) AS changes,
    sum(ABS(delta)) OVER (PARTITION BY channel ORDER BY FLOOR(__time TO MINUTE) ASC) AS cum_changes
FROM wikipedia
WHERE channel IN ('#kk.wikipedia', '#lt.wikipedia')
GROUP BY channel, __time, delta

View results

Example with window frames

The following query uses a few different window frames to calculate overall activity by channel:

SELECT
    channel, 
    TIME_FLOOR(__time, 'PT1H')      AS time_hour, 
    SUM(delta)                      AS hourly_channel_changes,
    SUM(SUM(delta)) OVER cumulative AS cumulative_activity_in_channel,
    SUM(SUM(delta)) OVER moving5    AS csum5,
    COUNT(1) OVER moving5           AS count5
FROM "wikipedia"
WHERE channel = '#en.wikipedia'
  AND __time BETWEEN '2016-06-27' AND '2016-06-28'
GROUP BY 1, TIME_FLOOR(__time, 'PT1H')
WINDOW cumulative AS (   
                         PARTITION BY channel 
                         ORDER BY TIME_FLOOR(__time, 'PT1H') 
                         ROWS BETWEEN UNBOUNDED PRECEDING AND CURRENT ROW
                     )
                     ,
        moving5 AS ( 
                    PARTITION BY channel 
                    ORDER BY TIME_FLOOR(__time, 'PT1H') 
                    ROWS BETWEEN 4 PRECEDING AND CURRENT ROW
                  )

View results

The example defines multiple window specifications in the WINDOW clause that you can use for various window function calculations.

The query uses two windows:

• cumulative is partitioned by channel and includes all rows from the beginning of partition up to the current row as ordered by __time to enable cumulative aggregation
• moving5 is also partitioned by channel but only includes up to the last four rows and the current row as ordered by time

The number of rows considered for the moving5 window for the count5 column:

• starts at a single row because there are no rows before the current one
• grows up to five rows as defined by ROWS BETWEEN 4 ROWS PRECEDING AND CURRENT ROW

Known issues

The following are known issues with window functions:

• Aggregates with ORDER BY specified are processed in the window: ROWS BETWEEN UNBOUNDED PRECEDING AND CURRENT ROW
  This behavior differs from other databases that use the default of RANGE BETWEEN UNBOUNDED PRECEDING AND CURRENT ROW.
  In cases where the order column is unique there is no difference between RANGE / ROWS; windows with RANGE specifications are handled as ROWS.
• LEAD/LAG ignores the default value
• LAST_VALUE returns the last value of the window even when you include an ORDER BY clause