Convert flattened key/value table into hierarchical JSON in PostgreSQL - json

Question

I have a PostgreSQL table with unique key/value pairs, which were originally in a JSON format, but have been normalized and melted:
key | value
-----------------------------
name | Bob
address.city | Vancouver
address.country | Canada
I need to turn this into a hierarchical JSON:
{
"name": "Bob",
"address": {
"city": "Vancouver",
"country": "Canada"
}
}
Is there a way to do this easily within SQL?

Answer 1

jsonb_set() almost does everything for you, but unfortunately it can only create missing leafs (i.e. missing last keys on a path), but not whole missing branches. To overcome this, here is a modified version of it, which can set values on any missing levels:
create function jsonb_set_rec(jsonb, jsonb, text[])
returns jsonb
language sql
as $$
select case
when array_length($3, 1) > 1 and ($1 #> $3[:array_upper($3, 1) - 1]) is null
then jsonb_set_rec($1, jsonb_build_object($3[array_upper($3, 1)], $2), $3[:array_upper($3, 1) - 1])
else jsonb_set($1, $3, $2, true)
end
$$;
Now you only need to apply this function one-by-one to your rows, starting with an empty json object: {}. You can do this with either recursive CTEs:
with recursive props as (
(select distinct on (grp)
pk, grp, jsonb_set_rec('{}', to_jsonb(value), string_to_array(key, '.')) json_object
from eav_tbl
order by grp, pk)
union all
(select distinct on (grp)
eav_tbl.pk, grp, jsonb_set_rec(json_object, to_jsonb(value), string_to_array(key, '.'))
from props
join eav_tbl using (grp)
where eav_tbl.pk > props.pk
order by grp, eav_tbl.pk)
)
select distinct on (grp)
grp, json_object
from props
order by grp, pk desc;
Or, with a custom aggregate defined as:
create aggregate jsonb_set_agg(jsonb, text[]) (
sfunc = jsonb_set_rec,
stype = jsonb,
initcond = '{}'
);
your query could became as simple as:
select grp, jsonb_set_agg(to_jsonb(value), string_to_array(key, '.'))
from eav_tbl
group by grp;
https://rextester.com/TULNU73750

Answer 2

There are no ready to use tools for this. The function generates a hierarchical json object based on a path:
create or replace function jsonb_build_object_from_path(path text, value text)
returns jsonb language plpgsql as $$
declare
obj jsonb;
keys text[] := string_to_array(path, '.');
level int := cardinality(keys);
begin
obj := jsonb_build_object(keys[level], value);
while level > 1 loop
level := level- 1;
obj := jsonb_build_object(keys[level], obj);
end loop;
return obj;
end $$;
You also need the aggregate function jsonb_merge_agg(jsonb) described in this answer. The query:
with my_table (path, value) as (
values
('name', 'Bob'),
('address.city', 'Vancouver'),
('address.country', 'Canada'),
('first.second.third', 'value')
)
select jsonb_merge_agg(jsonb_build_object_from_path(path, value))
from my_table;
gives this object:
{
"name": "Bob",
"first":
{
"second":
{
"third": "value"
}
},
"address":
{
"city": "Vancouver",
"country": "Canada"
}
}
The function do not recognize json arrays.

Answer 3

I can't really think of something simpler, although I think there should be an easier way.
I assume there is some additional column that can be used to bring the keys that belong to one "person" together, I used p_id for that in my example.
select p_id,
jsonb_object_agg(k, case level when 1 then v -> k else v end)
from (
select p_id,
elements[1] k,
jsonb_object_agg(case cardinality(elements) when 1 then ky else elements[2] end, value) v,
max(cardinality(elements)) as level
from (
select p_id,
"key" as ky,
string_to_array("key", '.') as elements, value
from kv
) t1
group by p_id, k
) t2
group by p_id;
The innermost query just converts the dot notation to an array for easier access later.
The next level then builds JSON objects depending on the "key". For the "single level" keys, it just uses key/value, for the others it uses the second element + the value and then aggregates those that belong together.
The second query level returns the following:
p_id | k | v | level
-----+---------+--------------------------------------------+------
1 | address | {"city": "Vancouver", "country": "Canada"} | 2
1 | name | {"name": "Bob"} | 1
2 | address | {"city": "Munich", "country": "Germany"} | 2
2 | name | {"name": "John"} | 1
The aggregation done in the second step, leaves one level too much for the "single element" keys, and that's what we need level for.
If that distinction wasn't made, the final aggregation would return {"name": {"name": "Bob"}, "address": {"city": "Vancouver", "country": "Canada"}} instead of the wanted: {"name": "Bob", "address": {"city": "Vancouver", "country": "Canada"}}.
The expression case level when 1 then v -> k else v end essentially turns {"name": "Bob"} back to "Bob".
So, with the following sample data:
create table kv (p_id integer, "key" text, value text);
insert into kv
values
(1, 'name','Bob'),
(1, 'address.city','Vancouver'),
(1, 'address.country','Canada'),
(2, 'name','John'),
(2, 'address.city','Munich'),
(2, 'address.country','Germany');
then query returns:
p_id | jsonb_object_agg
-----+-----------------------------------------------------------------------
1 | {"name": "Bob", "address": {"city": "Vancouver", "country": "Canada"}}
2 | {"name": "John", "address": {"city": "Munich", "country": "Germany"}}
Online example: https://rextester.com/SJOTCD7977

Answer 4

create table kv (key text, value text);
insert into kv
values
('name','Bob'),
('address.city','Vancouver'),
('address.country','Canada'),
('name','John'),
('address.city','Munich'),
('address.country','Germany');
create view v_kv as select row_number() over() as nRec, key, value from kv;
create view v_datos as
select k1.nrec, k1.value as name, k2.value as address_city, k3.value as address_country
from v_kv k1 inner join v_kv k2 on (k1.nrec + 1 = k2.nrec)
inner join v_kv k3 on ((k1.nrec + 2= k3.nrec) and (k2.nrec + 1 = k3.nrec))
where mod(k1.nrec, 3) = 1;
select json_agg(json_build_object('name',name, 'address', json_build_object('city',address_city, 'country', address_country)))
from v_datos;