Re: BUG #15115: temporary AND unlogged tables for parallel db import

That sounds right.
Every session (72 for now) creates and drops several (temp-) tables and
indexes. Every session runs for about 30min, then does a final insert
and exits afterwards.
Depending on parallelism chosen, we run 8, 16 or 24 sessions in parallel.
The main, (global-) table which persists has been declared as either
logged or unlogged, there is much WAL traffic in any case.

On 3/16/18 08:25, Tom Lane wrote:
> Andrew Gierth <andrew(at)tao11(dot)riddles(dot)org(dot)uk> writes:
>> "Enno" == Enno Wein <ennowein(at)yahoo(dot)com> writes:
>> Enno> However, it doesn't seem to work like that. Even with ONLY either
>> Enno> unlogged (the global ones) and temprorary tables (the local ones), the
>> Enno> system still produces a huge amount of WAL/checkpoint writes.
>
>> So what you just said above is that you get a lot of WAL/checkpoint
>> writes when using unlogged non-temporary tables. What this implies is
>> that you have something _else_ going on, involving persistent tables
>> (i.e. non-temporary non-unlogged tables - these are the only kind that
>> get their data changes WAL-logged).
>
> One theory is that the OP's workload involves a huge number of short-lived
> temp tables, and then the WAL traffic is coming from the system catalog
> changes involved in creating/dropping those tables (since even a temp
> table must have entries in pg_class, pg_attribute, etc). That's only
> a theory though, given the lack of any hard evidence in this thread.
>
> regards, tom lane
>