Dynamic row to column pivotation/transpose in Postgres made simple.
Author: Hannes Landeholm [email protected]
colpivot.sql
defines a single Postgres function:
create or replace function colpivot(
out_table varchar, in_query varchar,
key_cols varchar[], class_cols varchar[],
value_e varchar, col_order varchar
) returns void
The colpivot()
function groups data by a specific key and returns a column
for every unique class with a specified value expression in a new destination
temporary output table. The result is returned in a new temporary table.
This is the similar to the problem solved by
crosstab.
However, colpivot()
can be used with completely dynamic data. With crosstab
you MUST know what categories/classes you have before hand. The crosstab
function is also incompatible with multiple key or category/class columns.
Overall, the benefits of colpivot()
benefits are:
- Completely dynamic, no row specification required.
- Supports multiple rows and classes/attributes columns.
- Gives complete control over output columns order and limit.
- Easier to understand and use.
- Does not require loading extra modules.
- Does not require writing extra functions.
See also this thread on Stack Overflow.
Essentially the function transposes query results with columns like:
key1, key2, ..., keyN, class1, class2, ... classN, *
to:
key1, key2, ..., keyN, classC1, classC2, ..., classCN
The output is undefined if the input query has more than one row with the
same (key1, key2, ..., keyN, class1, class2, ... classN)
value. In most
real world use cases of colpivot()
it is sensible to have a primary key
or unique index over these columns when the input query selects a table
directly or using an input query with a distinct/group by over these columns
when uniqueness is not guaranteed.
- The output will have one row per unique key combination.
- The output will have one class column per unique class combination.
- The value of the corresponding class column is the result of the specified
expression or
null
if the corresponding class combination has no associated key.
To avoid having to specify an output column definition (since there is many real world cases where this is not possible to say beforehand) the result is stored in a temporary table with the specified name. It is impossible to return a result set with dynamic columns in Postgres without using temporary tables.
out_table
- Unquoted new temporary table to create. The table is deleted when the transaction ends (on commit drop
).in_query
[1] - Query to run that generates source data tocolpivot()
.key_cols
- Array of unquoted key columns.class_cols
- Array of unquoted class columns.value_e
[1] - Value expression. You must use the#
token as an alias if you are referencing a column in the input result. For example, specify#.salary'
instead of'salary'
.col_order
[1] - Column order. Specify as null to simply use the sorted classes. This is useful if you want another specific column order. For example, you may want to have the class with the highest salary as the first column. In that case you can specifysum(salary) desc
. You can also (ab)use this parameter to limit the number of columns returned like:sum(salary) desc limit 10
to only get the 10 highest salaries.
[1] These parameters are concatenated directly in evaluated queries to allow maximum flexibility for the caller and therefore unsafe. Ensure that you are not feeding dirty/non-validated/unquoted data into these parameters as that will allow a SQL injection exploit.
begin;
create temp table _test (
year int,
month int,
country varchar,
state varchar,
income int
) on commit drop;
insert into _test values
(1985, 01, 'sweden', '', 10),
(1985, 01, 'denmark', '', 11),
(1985, 01, 'usa', 'washington', 13),
(1985, 02, 'sweden', '', 20),
(1985, 02, 'usa', 'washington', 21),
(1985, 03, 'sweden', '', 34),
(1985, 03, 'denmark', '', 31),
(1985, 03, 'usa', 'washington', 39),
(1990, 12, 'sweden', '', 42),
(1990, 12, 'denmark', '', 43),
(1990, 12, 'usa', 'washington', 49),
(1990, 12, 'germany', '', 45);
select colpivot('_test_pivoted', 'select * from _test',
array['year', 'month'], array['country', 'state'], '#.income', null);
select * from _test_pivoted order by year, month;
-- returns:
-- year | month | 'denmark', '' | 'germany', '' | 'sweden', '' | 'usa', 'washington'
-- ------+-------+---------------+---------------+--------------+---------------------
-- 1985 | 1 | 11 | | 10 | 13
-- 1985 | 2 | | | 20 | 21
-- 1985 | 3 | 31 | | 34 | 39
-- 1990 | 12 | 43 | 45 | 42 | 49
-- (4 rows)
rollback;
MPLv2 (https://www.mozilla.org/MPL/2.0/)