Re: Proposal: Local indexes for partitioned table

On Tue, Apr 18, 2017 at 4:43 PM, Maksim Milyutin
<m(dot)milyutin(at)postgrespro(dot)ru> wrote:
> On 18.04.2017 13:08, Amit Langote wrote:
>>
>> Hi,
>>
>
> Hi, Amit!
>
>> On 2017/04/17 23:00, Maksim Milyutin wrote:
>>>
>>>
>>> Ok, thanks for the note.
>>>
>>> But I want to discuss the relevancy of introduction of a new relkind for
>>> partitioned index. I could to change the control flow in partitioned
>>> index
>>> creation (specify conditional statement in the 'index_create' routine in
>>> attached patch) and not enter to the 'heap_create' routine. This case
>>> releases us from integrating new relkind into different places of
>>> Postgres
>>> code. But we have to copy-paste some specific code from 'heap_create'
>>> function, e.g., definition of relfilenode and tablespaceid for the new
>>> index and perhaps something more when 'heap_create' routine will be
>>> extended.
>>
>>
>> I may be missing something, but isn't it that a new relkind will be needed
>> anyway? How does the rest of the code distinguish such index objects once
>> they are created?
>
>
> Local partitioned indexes can be recognized through the check on the relkind
> of table to which the index refers. Something like this:
>
> heap = relation_open(IndexGetRelation(indexid, false), heapLockmode);
> if (heap->rd_rel->relkind == RELKIND_PARTITIONED_TABLE)
> /* indexid is local index on partitioned table */

An index on partitioned table can be global index (yet to be
implemented) or a local index. We can not differentiate between those
just by looking at the relation on which they are built.

>
>> Is it possible that some other code may try to access
>> the storage for an index whose indrelid is a partitioned table?
>>
>
> Thеsе cases must be caught. But as much as partitioned tables doesn't
> participate in query plans their indexes are unaccessible by executor.
> Reindex operation is overloaded with my patch.
>

A global index would have storage for a partitioned table whereas a
local index wouldn't have any storage for a partitioned table.

I agree with Amit that we need new relkinds for local as well as global indexes.
--
Best Wishes,
Ashutosh Bapat
EnterpriseDB Corporation
The Postgres Database Company