Обсуждение: [GENERAL] Indexes being ignored after upgrade to 9.5

On Tue, Jul 25, 2017 at 10:34 PM, Nick Brennan <nbrennan02@gmail.com> wrote:
> We've added duplicate indexes and analyzing, however the new indexes are
> still ignored unless we force using enable_seqscan=no or reduce
> random_page_cost to 2. The query response times using the new indexes are
> still as slow when we do this. Checking pg_stat_user_indexes the number of
> tuples returned per idx_scan is far greater after the upgrade than before.
> All indexes show valid in pg_indexes.
>
>
> We have tried increasing effective_cache_size but no effect (the queries
> appear to go slower). The DB is 24x7 so we cannot reindex the tables/
> partitions.
>
>
> Can anyone suggest why this would be happening?

Are the indexes bloated? Are they larger than before, as indicated by
psql's \di+ or similar? Did you notice that this happened immediately,
or did it take a while? Are these unique indexes or not? Do you have a
workload with many UPDATEs?

I ask all these questions because I think it's possible that this is
explained by a regression in 9.5's handling of index bloat, described
here:

http://postgr.es/m/CAH2-Wz=SfAKVMv1x9Jh19EJ8am8TZn9f-yECipS9HrrRqSswnA@mail.gmail.com

--
Peter Geoghegan

On Wed, Jul 26, 2017 at 2:05 PM, Peter Geoghegan <pg@bowt.ie> wrote:
> On Tue, Jul 25, 2017 at 10:34 PM, Nick Brennan <nbrennan02@gmail.com> wrote:
>> We've added duplicate indexes and analyzing, however the new indexes are
>> still ignored unless we force using enable_seqscan=no or reduce
>> random_page_cost to 2. The query response times using the new indexes are
>> still as slow when we do this. Checking pg_stat_user_indexes the number of
>> tuples returned per idx_scan is far greater after the upgrade than before.
>> All indexes show valid in pg_indexes.

I assume that you mean that pg_stat_user_indexes.idx_tup_read is a lot
higher than before, in proportion to pg_stat_user_indexes.idx_scan.
What about the ratio between pg_stat_user_indexes.idx_tup_read and
pg_stat_user_indexes.idx_tup_fetch? How much has that changed by?

--
Peter Geoghegan

Hi Peter,

Many thanks for your response. I tried to cancel the thread, it was unfortunately stupidity that was the issue. We'd
beenforced to manually analyze our tables due to time constraints, and one of the table partitions read in the query
wasmissed. It was reporting a bitmap index scan on the parent so we thought all was ok, and was then causing other
tablesto sequential scan. 

A further misunderstanding was that an explain analyze would initiate stats gathering on all queried tables, however
thisis not the case. 

Thanks again for your response, we'll check the behaviour you report.

Best regards
Nick


> On 27 Jul 2017, at 00:40, Peter Geoghegan <pg@bowt.ie> wrote:
>
>> On Wed, Jul 26, 2017 at 2:05 PM, Peter Geoghegan <pg@bowt.ie> wrote:
>>> On Tue, Jul 25, 2017 at 10:34 PM, Nick Brennan <nbrennan02@gmail.com> wrote:
>>> We've added duplicate indexes and analyzing, however the new indexes are
>>> still ignored unless we force using enable_seqscan=no or reduce
>>> random_page_cost to 2. The query response times using the new indexes are
>>> still as slow when we do this. Checking pg_stat_user_indexes the number of
>>> tuples returned per idx_scan is far greater after the upgrade than before.
>>> All indexes show valid in pg_indexes.
>
> I assume that you mean that pg_stat_user_indexes.idx_tup_read is a lot
> higher than before, in proportion to pg_stat_user_indexes.idx_scan.
> What about the ratio between pg_stat_user_indexes.idx_tup_read and
> pg_stat_user_indexes.idx_tup_fetch? How much has that changed by?
>
> --
> Peter Geoghegan