Re: Parallel append plan instability/randomness

On 8 January 2018 at 13:35, Amit Kapila <amit(dot)kapila16(at)gmail(dot)com> wrote:
> On Mon, Jan 8, 2018 at 11:26 AM, Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> wrote:
>> Amit Khandekar <amitdkhan(dot)pg(at)gmail(dot)com> writes:
>>> The fact that b_star gets moved from 5th position to the first
>>> position in the scans, indicates that it's cost shoots up from 1.04 to
>>> a value greater than 1.16. It does not look like a case where two
>>> costs are almost same due to which their positions swap sometimes. I
>>> am trying to figure out what else can it be ...
>>
>
> That occurred to me as well, but still, the change in plan can happen
> due to the similar costs.

Agreed. But I think we should first fix the issue due to which the
test might be failing in this case. BTW, for your patch, I am thinking
we can have a separate factor other than STD_FUZZ_FACTOR ? This way,
we can make it much smaller than 1.01 also. And anyways,
STD_FUZZ_FACTOR is used only for comparing paths on the same relation,
whereas in our case, our comparision goal is different.

> Another possibility as indicated in the
> previous email is that if somehow the stats of table (reltuples,
> relpages) is not appropriate, say due to some reason analyze doesn't
> happen on the table.

Yes, I am also thinking on the same lines. E.g., if the relpages is 0
(due to no analyze run yet), tuple density calculation follows a
different logic, due to which reltuples can be quite bigger. I suspect
this also might be the reason. So yes, I think it's worth having
ANALYZE on *_star.

> For example, if you just manually update the
> value of reltuples for b_star in pg_class to 20 or so, you will see
> the plan as indicated in the failure. If that is true, then probably
> doing Analyze before Parallel Append should do the trick.

Or better still, we can have Analyze in create_misc.sql and
create_table.sql where the table is populated.

>
>> The gut feeling I had upon seeing the failure was that the plan shape
>> depends on the order in which rows happen to be read from the system
>> catalogs by a heapscan. I've not tried to run that idea to ground yet.
>>
>
> I don't see how something like that can happen because we internally
> sort the subpaths for parallel append.

True. Or may I didn't understand. Tom, are you referring to reading
pg_class.reltuples or similar, when say "read from the system
catalogs" ?

>
>
> --
> With Regards,
> Amit Kapila.
> EnterpriseDB: http://www.enterprisedb.com

--
Thanks,
-Amit Khandekar
EnterpriseDB Corporation
The Postgres Database Company