Grouping Functions

Functions for creating special groupings (also known as bucketing); as such these need to be used as part of the grouping.

HISTOGRAM

Synopsis:

  1. HISTOGRAM(
  2. numeric_exp,
  3. numeric_interval)
  4. HISTOGRAM(
  5. date_exp,
  6. date_time_interval)

Input:

numeric expression (typically a field)

numeric interval

date/time expression (typically a field)

date/time interval

Output: non-empty buckets or groups of the given expression divided according to the given interval

Description: The histogram function takes all matching values and divides them into buckets with fixed size matching the given interval, using (roughly) the following formula:

  1. bucket_key = Math.floor(value / interval) * interval

The histogram in SQL does NOT return empty buckets for missing intervals as the traditional histogram and date histogram. Such behavior does not fit conceptually in SQL which treats all missing values as NULL; as such the histogram places all missing values in the NULL group.

Histogram can be applied on either numeric fields:

  1. SELECT HISTOGRAM(salary, 5000) AS h FROM emp GROUP BY h;
  2. h
  3. ---------------
  4. 25000
  5. 30000
  6. 35000
  7. 40000
  8. 45000
  9. 50000
  10. 55000
  11. 60000
  12. 65000
  13. 70000

or date/time fields:

  1. SELECT HISTOGRAM(birth_date, INTERVAL 1 YEAR) AS h, COUNT(*) AS c FROM emp GROUP BY h;
  2. h | c
  3. ------------------------+---------------
  4. null |10
  5. 1952-01-01T00:00:00.000Z|8
  6. 1953-01-01T00:00:00.000Z|11
  7. 1954-01-01T00:00:00.000Z|8
  8. 1955-01-01T00:00:00.000Z|4
  9. 1956-01-01T00:00:00.000Z|5
  10. 1957-01-01T00:00:00.000Z|4
  11. 1958-01-01T00:00:00.000Z|7
  12. 1959-01-01T00:00:00.000Z|9
  13. 1960-01-01T00:00:00.000Z|8
  14. 1961-01-01T00:00:00.000Z|8
  15. 1962-01-01T00:00:00.000Z|6
  16. 1963-01-01T00:00:00.000Z|7
  17. 1964-01-01T00:00:00.000Z|4
  18. 1965-01-01T00:00:00.000Z|1

Expressions inside the histogram are also supported as long as the return type is numeric:

  1. SELECT HISTOGRAM(salary % 100, 10) AS h, COUNT(*) AS c FROM emp GROUP BY h;
  2. h | c
  3. ---------------+---------------
  4. 0 |10
  5. 10 |15
  6. 20 |10
  7. 30 |14
  8. 40 |9
  9. 50 |9
  10. 60 |8
  11. 70 |13
  12. 80 |3
  13. 90 |9

Do note that histograms (and grouping functions in general) allow custom expressions but cannot have any functions applied to them in the GROUP BY. In other words, the following statement is NOT allowed:

  1. SELECT MONTH(HISTOGRAM(birth_date), 2)) AS h, COUNT(*) as c FROM emp GROUP BY h ORDER BY h DESC;

as it requires two groupings (one for histogram followed by a second for applying the function on top of the histogram groups).

Instead one can rewrite the query to move the expression on the histogram inside of it:

  1. SELECT HISTOGRAM(MONTH(birth_date), 2) AS h, COUNT(*) as c FROM emp GROUP BY h ORDER BY h DESC;
  2. h | c
  3. ---------------+---------------
  4. 12 |7
  5. 10 |17
  6. 8 |16
  7. 6 |16
  8. 4 |18
  9. 2 |10
  10. 0 |6
  11. null |10

When the histogram in SQL is applied on DATE type instead of DATETIME, the interval specified is truncated to the multiple of a day. E.g.: for HISTOGRAM(CAST(birth_date AS DATE), INTERVAL '2 3:04' DAY TO MINUTE) the interval actually used will be INTERVAL '2' DAY. If the interval specified is less than 1 day, e.g.: HISTOGRAM(CAST(birth_date AS DATE), INTERVAL '20' HOUR) then the interval used will be INTERVAL '1' DAY.

All intervals specified for a date/time HISTOGRAM will use a fixed interval in their date_histogram aggregation definition, with the notable exceptions of INTERVAL '1' YEAR, INTERVAL '1' MONTH and INTERVAL '1' DAY where a calendar interval is used. The choice for a calendar interval was made for having a more intuitive result for YEAR, MONTH and DAY groupings. In the case of YEAR, for example, the calendar intervals consider a one year bucket as the one starting on January 1st that specific year, whereas a fixed interval one-year-bucket considers one year as a number of milliseconds (for example, 31536000000ms corresponding to 365 days, 24 hours per day, 60 minutes per hour etc.). With fixed intervals, the day of February 5th, 2019 for example, belongs to a bucket that starts on December 20th, 2018 and Elasticsearch (and implicitly Elasticsearch SQL) would have returned the year 2018 for a date that’s actually in 2019. With calendar interval this behavior is more intuitive, having the day of February 5th, 2019 actually belonging to the 2019 year bucket.

Histogram in SQL cannot be applied applied on TIME type. E.g.: HISTOGRAM(CAST(birth_date AS TIME), INTERVAL '10' MINUTES) is currently not supported.