Re: Index no longer being used, destroying and recreating it restores use.

I've collected all relevant info(I think so at least) and put it here:

The table in question is used to keep filepath data, of files on a
harddrive.
The query in question is used to retrieve items which should be backed up,
but have not yet been.

The relevant columns of the table:

Table "public.item"
Column | Type |
Collation | Nullable | Default
------------------------------------+-----------------------------+-----------+----------+----------------------------------------------
id | bigint |
| not null | nextval('item_id_seq'::regclass)
shouldbebackedup | boolean |
| not null | true
backupperformed | boolean |
| not null | false
itemCreated | timestamp without time zone |
| | now()
filepath | text |
| |

The existing index, which no longer gets used:
"index_in_question" btree (shouldbebackedup, backupperformed, itemCreated,
filepath) WHERE shouldbebackedup = true AND backupperformed = false

The new index, made out of the exact same columns and conditions, get used
immediately after creation:
CREATE INDEX CONCURRENTLY index_test ON item USING btree (shouldbebackedup,
backupperformed, itemCreated, filepath) WHERE shouldbebackedup = true AND
backupperformed = false;

The query in question will look something like this:
select * from item where shouldbebackedup=true and itemCreated<='2020-06-05
00:00:00.000' and backupperformed=false order by filepath asc, id asc limit
100 offset 10400;

Having done a count, there are around 13000 items here, without the offset
and limit.
That being said, the amount is entirely dependant on what was added on a
previous day.

I tried creating an extended statistic, like this, but it had no effect:
CREATE STATISTICS s1 (dependencies) ON shouldbebackedup, backupperformed
FROM item;

Settings from the conf file I think are related:

shared_buffers = 1024MB
effective_cache_size = 2048MB
random_page_cost = 1.1
effective_io_concurrency = 200
work_mem = 32MB

Finally, I state again that this database gets a nightly "vacuum analyze".

My thanks for looking at this and any suggestions one might have.

Regards,
Koen

On Thu, Jun 4, 2020 at 7:08 PM Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> wrote:

> Adrian Klaver <adrian(dot)klaver(at)aklaver(dot)com> writes:
> > On 6/4/20 9:43 AM, Tom Lane wrote:
> >> It's possible that the index had bloated to the point where the planner
> >> thought it was cheaper to use a seqscan. Did you make a note of the
> >> cost estimates for the different plans?
>
> > I missed the part where the OP pointed to a SO question. In that
> > question where links to explain.depesz.com output.
>
> Ah, I didn't bother to chase that link either.
>
> So the cost estimates are only a fraction of a percent apart, making
> it unsurprising for not-so-large changes in the index size to cause
> a flip in the apparently-cheapest plan. The real question then is
> why the cost estimates aren't actually modeling the real execution
> times very well; and I'd venture that that question boils down to
> why is this rowcount estimate so far off:
>
> > -> Parallel Seq Scan on oscar mike_three
> > (cost=0.000..1934568.500 rows=2385585 width=3141) (actual
> > time=159.800..158018.961 rows=23586 loops=3)
> > Filter: (four AND (NOT bravo) AND (zulu <=
> > 'echo'::timestamp without time zone))
> > Rows Removed by Filter: 8610174
>
> We're not going to be able to answer that if the OP doesn't wish
> to decloak his data a bit more ... but a reasonable guess is that
> those filter conditions are correlated. With late-model Postgres
> you might be able to improve matters by creating extended statistics
> for this table.
>
> regards, tom lane
>