Adrian Klaver <adrian.klaver@aklaver.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