Ayo <ayo(at)blicky(dot)net> writes:
> On 2021-02-23, Tom Lane wrote:
>> The slow query isn't using the chars_pkey1 index, which makes one
>> wonder if you have a corresponding index in the custom-type case,
> The index exists and is usable in the custom-type case:
Hmm. The next most likely theory seems to be something wrong with
cost estimation, causing the planner to avoid the nestloop-with-
inner-indexscan plan even though it would work. Have you tried
"enable_seqscan = off" (and maybe also disable merge and hash joins)
to see if you can force choice of that plan?
regards, tom lane