Re: Pushing down sorted joins

Поиск
Список
Период
Сортировка
От Ashutosh Bapat
Тема Re: Pushing down sorted joins
Дата
Msg-id CAFjFpReY+wtjcO78k3NHRLtoG3F4JLY35sVsP_Q=N=_S8LX=Ow@mail.gmail.com
обсуждение исходный текст
Ответ на Pushing down sorted joins  (Ashutosh Bapat <ashutosh.bapat@enterprisedb.com>)
Ответы Re: Pushing down sorted joins  (Robert Haas <robertmhaas@gmail.com>)
Список pgsql-hackers
Rushabh pointed out that declarations of helper functions get_useful_ecs_for_relation and get_useful_pathkeys_for_relation() are part of FDW routines declarations rather than helper function declaration. Since those functions are related to this patch, the attached patch moves those declaration in their right place.

On Wed, Feb 17, 2016 at 5:37 PM, Ashutosh Bapat <ashutosh.bapat@enterprisedb.com> wrote:
Hi All,
Now that we have join pushdown support in postgres_fdw, we can leverage the sort pushdown mechanism for base relations to work for pushed down joins as well. PFA patch for the same.

The code to find useful set of pathkeys and then generate paths for each list of pathkeys is moved into a function which is called for base relations and join relations, while creating respective paths. The useful pathkeys are same as the base relation i.e. root->query_pathkeys and pathkeys useful for merge join as discussed in [1].

I measured performance of pushing down sort for merge joins for query
SELECT lt1.val, ft1.val, ft2.val FROM lt1 join (ft1 join ft2 on (ft1.val = ft2.val)) on (lt1.val = ft1.val) where ft1, ft2 are foreign tables, join between which gets pushed down to the foreign server and lt is the local table.

Without the patch servers prefers local merge join between foreign tables followed by merge join with local table by getting the data sorted from the foreign server. But with the patch, it pushes down the foreign join and also gets the data sorted for local merge join. The times measured over 10 runs of query with and without patch are

With patch
 avg_exe_time | std_dev_exe_time | min_exe_time | max_exe_time
--------------+------------------+--------------+--------------
   60310.0369 | 251.075471210925 |    59895.064 |    60746.496

Without patch
 avg_exe_time | std_dev_exe_time | min_exe_time | max_exe_time
--------------+------------------+--------------+--------------
   86396.6001 |  254.30988131848 |    85906.606 |    86742.311

With the patch the execution time of the query reduces by 30%.

The scripts to setup and run query and outputs of running query with and without patch are attached.


[1] http://www.postgresql.org/message-id/CAFjFpRfeKHiCmwJ72p4=ZvuZRQsaU9tbfyW7vwr-5PPvrCbcQQ@mail.gmail.com
--
Best Wishes,
Ashutosh Bapat
EnterpriseDB Corporation
The Postgres Database Company



--
Best Wishes,
Ashutosh Bapat
EnterpriseDB Corporation
The Postgres Database Company
Вложения

В списке pgsql-hackers по дате отправления:

Предыдущее
От: Robert Haas
Дата:
Сообщение: Re: postgres_fdw vs. force_parallel_mode on ppc
Следующее
От: Fujii Masao
Дата:
Сообщение: tab completion for CREATE USER MAPPING