PostgreSQL - best way to return an array of key-value pairs

ARRAYs can only hold elements of the same type

Your example displays a text and an integer value (no single quotes around 1). It is generally impossible to mix types in an array. To get those values into an array you have to create a composite type and then form an ARRAY of that composite type like you already mentioned yourself.

Alternatively you can use the data types json in Postgres 9.2+, jsonb in Postgres 9.4+ or hstore for key-value pairs.

Of course, you can cast the integer to text, and work with a two-dimensional text array. Consider the two syntax variants for a array input in the demo below and consult the manual on array input.

There is a limitation to overcome. If you try to aggregate an ARRAY (build from key and value) into a two-dimensional array, the aggregate function array_agg() or the ARRAY constructor error out:

ERROR:  could not find array type for data type text[]

There are ways around it, though.

Aggregate key-value pairs into a 2-dimensional array

PostgreSQL 9.1 with standard_conforming_strings= on:

CREATE TEMP TABLE tbl( id     int,txt    text,txtarr text[]);

The column txtarr is just there to demonstrate syntax variants in the INSERT command. The third row is spiked with meta-characters:

INSERT INTO tbl VALUES (1, 'foo', '{{1,foo1},{2,bar1},{3,baz1}}'),(2, 'bar', ARRAY[['1','foo2'],['2','bar2'],['3','baz2']]),(3, '}b",a{r''', '{{1,foo3},{2,bar3},{3,baz3}}'); -- txt has meta-charactersSELECT * FROM tbl;

Simple case: aggregate two integer (I use the same twice) into a two-dimensional int array:

Update: Better with custom aggregate function

With the polymorphic type anyarray it works for all base types:

CREATE AGGREGATE array_agg_mult (anyarray)  (    SFUNC     = array_cat   ,STYPE     = anyarray   ,INITCOND  = '{}');

Call:

SELECT array_agg_mult(ARRAY[ARRAY[id,id]]) AS x        -- for int      ,array_agg_mult(ARRAY[ARRAY[id::text,txt]]) AS y -- or textFROM   tbl;

Note the additional ARRAY[] layer to make it a multidimensional array.

Update for Postgres 9.5+

Postgres now ships a variant of array_agg() accepting array input and you can replace my custom function from above with this:

The manual:

array_agg(expression)
...
input arrays concatenated into array of one higher dimension (inputs must all have same dimensionality, and cannot be empty or NULL)

I suspect that without having more knowledge of your application I'm not going to be able to get you all the way to the result you need. But we can get pretty far. For starters, there is the ROW function:

# SELECT 'foo', ROW(3, 'Bob'); ?column? |   row   ----------+--------- foo      | (3,Bob)(1 row)

So that right there lets you bundle a whole row into a cell. You could also make things more explicit by making a type for it:

# CREATE TYPE person(id INTEGER, name VARCHAR);CREATE TYPE# SELECT now(), row(3, 'Bob')::person;              now              |   row   -------------------------------+--------- 2012-02-03 10:46:13.279512-07 | (3,Bob)(1 row)

Incidentally, whenever you make a table, PostgreSQL makes a type of the same name, so if you already have a table like this you also have a type. For example:

# DROP TYPE person;DROP TYPE# CREATE TABLE people (id SERIAL, name VARCHAR);NOTICE:  CREATE TABLE will create implicit sequence "people_id_seq" for serial column "people.id"CREATE TABLE# SELECT 'foo', row(3, 'Bob')::people; ?column? |   row   ----------+--------- foo      | (3,Bob)(1 row)

See in the third query there I used people just like a type.

Now this is not likely to be as much help as you'd think for two reasons:

1. I can't find any convenient syntax for pulling data out of the nested row.

   I may be missing something, but I just don't see many people using this syntax. The only example I see in the documentation is a function taking a row value as an argument and doing something with it. I don't see an example of pulling the row out of the cell and querying against parts of it. It seems like you can package the data up this way, but it's hard to deconstruct after that. You'll wind up having to make a lot of stored procedures.

2. Your language's PostgreSQL driver may not be able to handle row-valued data nested in a row.

   I can't speak for NPGSQL, but since this is a very PostgreSQL-specific feature you're not going to find support for it in libraries that support other databases. For example, Hibernate isn't going to be able to handle fetching an object stored as a cell value in a row. I'm not even sure the JDBC would be able to give Hibernate the information usefully, so the problem could go quite deep.

So, what you're doing here is feasible provided you can live without a lot of the niceties. I would recommend against pursuing it though, because it's going to be an uphill battle the whole way, unless I'm really misinformed.