Add support for hstore

[bullno1]

Currently typespec is:

hstore() :: [{key(), value()}] | {[key(), value()}.
key() :: atom() | binary() | string() | number().
value() :: null | atom() | binary() | string() | number().

"jiffy-like" format is required for array support since there's no trivial way to distinguish between an array of hstore and a hstore.
Return type for hstore is always: [{binary(), binary() | null}] for convenience. Maps in 17.0 will solve this problem though.

Opened a pull request because I need some discussion about this.

[seriyps]

Maybe, as you say, define hstore strictly on jiffy-like format {[ {key(), value()} ]} and drop [ {key(), value()} ]? Including return value. This should fix any ambiguity.

[davidw]

I would vote for avoiding ambiguity.

[bullno1]

Would 17.0's map be better?
I don't have any urgent need for it right now anyway.

[davidw]

Given that R17 isn't even released yet, and the conservative nature of many in the Erlang community, I think a solution that only works for that is not optimal. I think I would try and do something like:

• A reasonable solution for older Erlangs. Essential.
• A map-based solution for new Erlangs. Nice to have.

[bullno1]

jiffy-like format is enforced now

[davidw]

This has been merged into the devel branch! Thank you!

[davidw]

I'm actually getting some errors with the tests. What version are you running of Postgres?

[davidw]

   pgsql_tests: all_test_ (hstore_type_test(pgsql))...*failed*
in function pgsql_tests:'-check_type/6-fun-2-'/7 (test/pgsql_tests.erl, line 829)
in call from pgsql_tests:with_connection/4 (test/pgsql_tests.erl, line 788)
in call from pgsql_tests:hstore_type_test/1 (test/pgsql_tests.erl, line 557)
**error:{badmatch,{ok,[{column,<<"hstore">>,{unknown_oid,255845},-1,-1,0}],[{<<>>}]}}

[bullno1]

Did you update the test schema?

It needs "CREATE EXTENSION hstore" and an extra column to do the serialization test.

bullno1@7e37aff#diff-b89470b98ce374506061ea6587f445edR38

[davidw]

Yes, I did. I'm on Postgres 9.1.13

[bullno1]

My make test result:

Failed: 6.  Skipped: 0.  Passed: 51.

As usual, those are ssl tests.
I'm running postgres 9.3.4-1 on ArchLinux. Let me get a VM to test 9.1.13.
Still, why would they change oid between versions?

[davidw]

SELECT n.nspname as "Schema",
  t.oid,
  pg_catalog.format_type(t.oid, NULL) AS "Name",
  pg_catalog.obj_description(t.oid, 'pg_type') as "Description"
FROM pg_catalog.pg_type t
     LEFT JOIN pg_catalog.pg_namespace n ON n.oid = t.typnamespace
WHERE (t.typrelid = 0 OR (SELECT c.relkind = 'c' FROM pg_catalog.pg_class c WHERE c.oid = t.typrelid))
  AND NOT EXISTS(SELECT 1 FROM pg_catalog.pg_type el WHERE el.oid = t.typelem AND el.typarray = t.oid)
  AND pg_catalog.pg_type_is_visible(t.oid)
ORDER BY 1, 2;

Shows this:

public     | 256714 | hstore                      | 
public     | 256793 | ghstore                     |

[bullno1]

Yes, it's totally random:

SELECT typname, typelem, typarray FROM pg_type WHERE typname LIKE '%hstore%';

shows

 typname  | typelem | typarray 
----------+---------+----------
 hstore   |       0 |    16672
 _hstore  |   16667 |        0
 ghstore  |       0 |    16749
 _ghstore |   16746 |        0

[davidw]

Idea: API to fill in type cache (ets or the process dictionary?) by querying the database, that is called at connect time, and whenever the user wants. So, if you know you are going to fool around with extensions during your session, you can regenerate the type cache.

[davidw]

I have pushed a version of this to the dynamic_types branch - all the tests pass, so have a look at it and see what you think.

[davidw]

We have hstore support now. Thanks!