Re: [Proposal] Global temporary tables

> 2019年10月11日 下午9:50，Konstantin Knizhnik <k(dot)knizhnik(at)postgrespro(dot)ru> 写道：
>
>
>
> On 11.10.2019 15:15, 曾文旌(义从) wrote:
>> Dear Hackers,
>>
>> This propose a way to develop global temporary tables in PostgreSQL.
>>
>> I noticed that there is an "Allow temporary tables to exist as empty by default in all sessions" in the postgresql todolist.
>> https://wiki.postgresql.org/wiki/Todo <https://wiki.postgresql.org/wiki/Todo>
>>
>> In recent years, PG community had many discussions about global temp table (GTT) support. Previous discussion covered the following topics:
>> (1) The main benefit or function: GTT offers features like “persistent schema, ephemeral data”, which avoids catalog bloat and reduces catalog vacuum.
>> (2) Whether follows ANSI concept of temporary tables
>> (3) How to deal with statistics, single copy of schema definition, relcache
>> (4) More can be seen in https://www.postgresql.org/message-id/73954ab7-44d3-b37b-81a3-69bdcbb446f7%40postgrespro.ru <https://www.postgresql.org/message-id/73954ab7-44d3-b37b-81a3-69bdcbb446f7%40postgrespro.ru>
>> (5) A recent implementation and design from Konstantin Knizhnik covered many functions of GTT: https://www.postgresql.org/message-id/attachment/103265/global_private_temp-1.patch <https://www.postgresql.org/message-id/attachment/103265/global_private_temp-1.patch>
>>
>> However, as pointed by Konstantin himself, the implementation still needs functions related to CLOG, vacuum, and MVCC visibility.
>>
>
> Just to clarify.
> I have now proposed several different solutions for GTT:
>
> Shared vs. private buffers for GTT:
> 1. Private buffers. This is least invasive patch, requiring no changes in relfilenodes.
> 2. Shared buffers. Requires changing relfilenode but supports parallel query execution for GTT.
>
> Access to GTT at replica:
> 1. Access is prohibited (as for original temp tables). No changes at all.
> 2. Tuples of temp tables are marked with forzen XID. Minimal changes, rollbacks are not possible.
> 3. Providing special XIDs for GTT at replica. No changes in CLOG are required, but special MVCC visibility rules are used for GTT. Current limitation: number of transactions accessing GTT at replica is limited by 2^32
> and bitmap of correspondent size has to be maintained (tuples of GTT are not proceeded by vacuum and not frozen, so XID horizon never moved).
>
> So except the limitation mentioned above (which I do not consider as critical) there is only one problem which was not addressed: maintaining statistics for GTT.
> If all of the following conditions are true:
>
> 1) GTT are used in joins
> 2) There are indexes defined for GTT
> 3) Size and histogram of GTT in different backends can significantly vary.
> 4) ANALYZE was explicitly called for GTT
>
> then query execution plan built in one backend will be also used for other backends where it can be inefficient.
> I also do not consider this problem as "show stopper" for adding GTT to Postgres.
When session A writes 10000000 rows of data to gtt X, session B also uses X at the same time and it has 100 rows of different data. If B uses analyze to count the statistics of 100000 rows of data and updates it to catalog.
Obviously, session A will get inaccurate query plan based on misaligned statistics when calculating the query plan for X related queries. Session A may think that table X is too small to be worth using index scan, but it is not. Each session needs to get the statistics of the self data to make the query plan.

> I still do not understand the opinion of community which functionality of GTT is considered to be most important.
> But the patch with local buffers and no replica support is small enough to become good starting point.
Yes ,the first step, we focus on complete basic functions of gtt (dml ddl index on gtt (MVCC visibility rules) storage).
Abnormal statistics can cause problems with index selection on gtt, so index on gtt and accurate statistical information is necessary.

>
> --
> Konstantin Knizhnik
> Postgres Professional: http://www.postgrespro.com <http://www.postgrespro.com/>
> The Russian Postgres Company