I am looking into a way of computing argmin or argmax aggregation of multiple rows in Snowflake similar to Hive or Presto.
In Hive, one can use a workaround with (named) structs because the aggregation function gets applied to the first element of the struct. Here is an example:
SELECT max(named_struct('y', y, 'x', x)).x FROM t
Now I am asking myself if there is a similar way to do this in Snowflake.
In Snowflake we have an OBJECT datatype with similar properties. Can I use the following code to compute argmin or argmax like in the Hive example? Are min/max aggregations for objects also performed on the first element of the object?
SELECT max(object_construct('y', y, 'x', x)).x FROM t
Running the above code returns an error: SQL compilation error: Function MAX does not support OBJECT argument type.. It actually doesn't support any complex type.
If I have correctly understood the logic of argmin(), then you could implement it as a javascript UDF like so:
create or replace function argmin("a" object, "b" object)
returns object
language javascript
as
$$
for (let [k,v] of Object.entries(a))
if (v==b[k])
continue
else
return v < b[k] ? a : b
return b
$$;
And applied like this:
with t as (
select
object_construct('x',1, 'y',2) a,
object_construct('x',2, 'y',1) b
)
select argmin(t.a,t.b):y from t;
Actually this functionality is built in:
select
object_construct('x',1, 'y',2) a,
object_construct('x',2, 'y',1) b,
iff(a<b, a, b) : y
;
Even more concisely:
select
object_construct('x',1, 'y',2) a,
object_construct('x',2, 'y',1) b,
least(a,b):y,
greatest(a,b):y;
Related
I need to search with LIKE or ILIKE - applied to elements of a JSONB[] column.
I can use unnest() to display nested JSONB elements separately, but I am having trouble combining that with my LIKE expression. My query:
SELECT
UNNEST( column_jsonb[] )
FROM
table1
WHERE
UNNEST( column_jsonb[] ) ->> 'member_type' LIKE '%member%'
Results in an error:
argument of WHERE must not return a set
How to get this right?
If it's indeed a jsonb[] column:
SELECT * -- or whatever
FROM t, unnest(jsonb_col) j
WHERE j->>'member_type' LIKE '%member%';
That's an implicit CROSS JOIN LATERAL.
If it's a JSON array in a plain jsonb column (more likely):
SELECT *
FROM t1, jsonb_array_elements(jsonb_col) j
WHERE j->>'member_type' LIKE '%member%';
Related:
Query for array elements inside JSON type
Is the only way to pass an extra parameter to the final function of a PostgreSQL aggregate to create a special TYPE for the state value?
e.g.:
CREATE TYPE geomvaltext AS (
geom public.geometry,
val double precision,
txt text
);
And then to use this type as the state variable so that the third parameter (text) finally reaches the final function?
Why aggregates can't pass extra parameters to the final function themselves? Any implementation reason?
So we could easily construct, for example, aggregates taking a method:
SELECT ST_MyAgg(accum_number, 'COMPUTE_METHOD') FROM blablabla
Thanks
You can define an aggregate with more than one parameter.
I don't know if that solves your problem, but you could use it like this:
CREATE OR REPLACE FUNCTION myaggsfunc(integer, integer, text) RETURNS integer
IMMUTABLE STRICT LANGUAGE sql AS
$f$
SELECT CASE $3
WHEN '+' THEN $1 + $2
WHEN '*' THEN $1 * $2
ELSE NULL
END
$f$;
CREATE AGGREGATE myagg(integer, text) (
SFUNC = myaggsfunc(integer, integer, text),
STYPE = integer
);
It could be used like this:
CREATE TABLE mytab
AS SELECT * FROM generate_series(1, 10) i;
SELECT myagg(i, '+') FROM mytab;
myagg
-------
55
(1 row)
SELECT myagg(i, '*') FROM mytab;
myagg
---------
3628800
(1 row)
I solved a similar issue by making a custom aggregate function that did all the operations at once and stored their states in an array.
CREATE AGGREGATE myagg(integer)
(
INITCOND = '{ 0, 1 }',
STYPE = integer[],
SFUNC = myaggsfunc
);
and:
CREATE OR REPLACE FUNCTION myaggsfunc(agg_state integer[], agg_next integer)
RETURNS integer[] IMMUTABLE STRICT LANGUAGE 'plpgsql' AS $$
BEGIN
agg_state[1] := agg_state[1] + agg_next;
agg_state[2] := agg_state[2] * agg_next;
RETURN agg_state;
END;
$$;
Then made another function that selected one of the results based on the second argument:
CREATE OR REPLACE FUNCTION myagg_pick(agg_state integer[], agg_fn character varying)
RETURNS integer IMMUTABLE STRICT LANGUAGE 'plpgsql' AS $$
BEGIN
CASE agg_fn
WHEN '+' THEN RETURN agg_state[1];
WHEN '*' THEN RETURN agg_state[2];
ELSE RETURN 0;
END CASE;
END;
$$;
Usage:
SELECT myagg_pick(myagg("accum_number"), 'COMPUTE_METHOD') FROM "mytable" GROUP BY ...
Obvious downside of this is the overhead of performing all the functions instead of just one. However when dealing with simple operations such as adding, multiplying etc. it should be acceptable in most cases.
You would have to rewrite the final function itself, and in that case you might as well write a set of new aggregate functions, one for each possible COMPUTE_METHOD. If the COMPUTE_METHOD is a data value or implied by a data value, then a CASE statement can be used to select the appropriate aggregate method. Alternatively, you may want to create a custom composite type with fields for accum_number and COMPUTE_METHOD, and write a single new aggregate function that uses this new data type.
What I need is a string/text json-compliant array of arrays like the following:
[
[1421420100, 14214201003487],
[1421420101, 14214201003487],
[1421420109, 14214201003487]
...
]
in order to return it from the server to the client without any further elaboration.
I have a function that queries on a table that returns the two fields, but I would like to return an array of arrays in order to call the function and then call array_to_json. Then, I would like to convert it into text/string.
Is that possible? Can an array of arrays have two different types?
Postgres arrays cannot hold elements of different type.
But you can return an array of composite type (bigint, double precision) instead. The type needs to be registered in the database instance. Either with an explicit CREATE TYPE statement, or implicitly as row type of a table or view.
Example:
CREATE TYPE int8_float AS (b bigint, f float);
CREATE OR REPLACE FUNCTION f_combo_arr()
RETURNS int8_float[]
LANGUAGE sql AS
$func$
SELECT ARRAY(SELECT (int8_col, float_col)::int8_float FROM tbl);
$func$;
Call:
SELECT f_combo_arr();
float is an alias for double precision in Postgres.
There may be a smarter solution, like returning data type json or a set instead of the array.
Related:
Return type for function with array_agg()
PostgreSQL - best way to return an array of key-value pairs
I have two arrays in PostgreSQL that I need to union. For example:
{1,2,3} union {1,4,5} would return {1,2,3,4,5}
Using the concatenate (||) operator would not remove duplicate entries, i.e. it returns {1,2,3,1,4,5}
I found one solution on the web, however I do not like how it needs to unnest both arrays:
select ARRAY(select unnest(ARRAY[1,2,3]) as a UNION select unnest(ARRAY[2,3,4,5]) as a)
Is there an operator or built-in function that will cleanly union two arrays?
If your problem is to unnest twice this will unnest only once
select array_agg(a order by a)
from (
select distinct unnest(array[1,2,3] || array[2,3,4,5]) as a
) s;
There is a extension intarray (in contrib package) that contains some useful functions and operators:
postgres=# create extension intarray ;
CREATE EXTENSION
with single pipe operator:
postgres=# select array[1,2,3] | array[3,4,5];
?column?
─────────────
{1,2,3,4,5}
(1 row)
or with uniq function:
postgres=# select uniq(ARRAY[1,2,3] || ARRAY[3,4,5]);
uniq
─────────────
{1,2,3,4,5}
(1 row)
ANSI/SQL knows a multiset, but it is not supported by PostgreSQL yet.
Can be done like so...
select uniq(sort(array_remove(array_cat(ARRAY[1,2,3], ARRAY[1,4,5]), NULL)))
gives:
{1,2,3,4,5}
array_remove is needed because your can't sort arrays with NULLS.
Sort is needed because uniq de-duplicates only if adjacent elements are found.
A benefit of this approach over #Clodoaldo Neto's is that works entire within the select, and doesn't the unnest in the FROM clause. This makes it straightforward to operate on multiple arrays columns at the same time, and in a single table-scan. (Although see Ryan Guill version as a function in the comment).
Also, this pattern works for all array types (who's elements are sortable).
A downside is that, feasibly, its a little slower for longer arrays (due to the sort and the 3 intermediate array allocations).
I think both this and the accept answer fail if you want to keep NULL in the result.
The intarray-based answers don't work when you're trying to take the set union of an array-valued column from a group of rows. The accepted array_agg-based answer can be modified to work, e.g.
SELECT selector_column, array_agg(a ORDER BY a) AS array_valued_column
FROM (
SELECT DISTINCT selector_column, UNNEST(array_valued_column) AS a FROM table
) _ GROUP BY selector_column;
but, if this is buried deep in a complex query, the planner won't be able to push outer WHERE expressions past it, even when they would substantially reduce the number of rows that have to be processed. The right solution in that case is to define a custom aggregate:
CREATE FUNCTION array_union_step (s ANYARRAY, n ANYARRAY) RETURNS ANYARRAY
AS $$ SELECT s || n; $$
LANGUAGE SQL IMMUTABLE LEAKPROOF PARALLEL SAFE;
CREATE FUNCTION array_union_final (s ANYARRAY) RETURNS ANYARRAY
AS $$
SELECT array_agg(i ORDER BY i) FROM (
SELECT DISTINCT UNNEST(x) AS i FROM (VALUES(s)) AS v(x)
) AS w WHERE i IS NOT NULL;
$$
LANGUAGE SQL IMMUTABLE LEAKPROOF PARALLEL SAFE;
CREATE AGGREGATE array_union (ANYARRAY) (
SFUNC = array_union_step,
STYPE = ANYARRAY,
FINALFUNC = array_union_final,
INITCOND = '{}',
PARALLEL = SAFE
);
Usage is
SELECT selector_column, array_union(array_valued_column) AS array_valued_column
FROM table
GROUP BY selector_column;
It's doing the same thing "under the hood", but because it's packaged into an aggregate function, the planner can see through it.
It's possible that this could be made more efficient by having the step function do the UNNEST and append the rows to a temporary table, rather than a scratch array, but I don't know how to do that and this is good enough for my use case.
I have a table with 4 array columns.. the results are like:
ids signed_ids new_ids new_ids_signed
{1,2,3} | {2,1,3} | {4,5,6} | {6,5,4}
Anyway to compare ids and signed_ids so that they come out equal, by ignoring the order of the elements?
You can use contained by operator:
(array1 <# array2 and array1 #> array2)
The additional module intarray provides operators for arrays of integer, which are typically (much) faster. Install once per database with (in Postgres 9.1 or later):
CREATE EXTENSION intarray;
Then you can:
SELECT uniq(sort(ids)) = uniq(sort(signed_ids));
Or:
SELECT ids #> signed_ids AND ids <# signed_ids;
Bold emphasis on functions and operators from intarray.
In the second example, operator resolution arrives at the specialized intarray operators if left and right argument are type integer[].
Both expressions will ignore order and duplicity of elements. Further reading in the helpful manual here.
intarray operators only work for arrays of integer (int4), not bigint (int8) or smallint (int2) or any other data type.
Unlike the default generic operators, intarray operators do not accept NULL values in arrays. NULL in any involved array raises an exception. If you need to work with NULL values, you can default to the standard, generic operators by schema-qualifying the operator with the OPERATOR construct:
SELECT ARRAY[1,4,null,3]::int[] OPERATOR(pg_catalog.#>) ARRAY[3,1]::int[]
The generic operators can't use indexes with an intarray operator class and vice versa.
Related:
GIN index on smallint[] column not used or error "operator is not unique"
The simplest thing to do is sort them and compare them sorted. See sorting arrays in PostgreSQL.
Given sample data:
CREATE TABLE aa(ids integer[], signed_ids integer[]);
INSERT INTO aa(ids, signed_ids) VALUES (ARRAY[1,2,3], ARRAY[2,1,3]);
the best thing to do is to if the array entries are always integers is to use the intarray extension, as Erwin explains in his answer. It's a lot faster than any pure-SQL formulation.
Otherwise, for a general version that works for any data type, define an array_sort(anyarray):
CREATE OR REPLACE FUNCTION array_sort(anyarray) RETURNS anyarray AS $$
SELECT array_agg(x order by x) FROM unnest($1) x;
$$ LANGUAGE 'SQL';
and use it sort and compare the sorted arrays:
SELECT array_sort(ids) = array_sort(signed_ids) FROM aa;
There's an important caveat:
SELECT array_sort( ARRAY[1,2,2,4,4] ) = array_sort( ARRAY[1,2,4] );
will be false. This may or may not be what you want, depending on your intentions.
Alternately, define a function array_compare_as_set:
CREATE OR REPLACE FUNCTION array_compare_as_set(anyarray,anyarray) RETURNS boolean AS $$
SELECT CASE
WHEN array_dims($1) <> array_dims($2) THEN
'f'
WHEN array_length($1,1) <> array_length($2,1) THEN
'f'
ELSE
NOT EXISTS (
SELECT 1
FROM unnest($1) a
FULL JOIN unnest($2) b ON (a=b)
WHERE a IS NULL or b IS NULL
)
END
$$ LANGUAGE 'SQL' IMMUTABLE;
and then:
SELECT array_compare_as_set(ids, signed_ids) FROM aa;
This is subtly different from comparing two array_sorted values. array_compare_as_set will eliminate duplicates, making array_compare_as_set(ARRAY[1,2,3,3],ARRAY[1,2,3]) true, whereas array_sort(ARRAY[1,2,3,3]) = array_sort(ARRAY[1,2,3]) will be false.
Both of these approaches will have pretty bad performance. Consider ensuring that you always store your arrays sorted in the first place.
If your arrays have no duplicates and are of the same dimension:
use array contains #>
AND array_length where the length must match the size you want on both sides
select (string_agg(a,',' order by a) = string_agg(b,',' order by b)) from
(select unnest(array[1,2,3,2])::text as a,unnest(array[2,2,3,1])::text as b) A