On 01/25/2018 11:31 PM, Todd A. Cook wrote:
> On 11/27/17 14:17, Tomas Vondra wrote:
>> Hi,
>>
>> On 11/27/2017 07:57 PM, tcook@blackducksoftware.com wrote:
>>> The following bug has been logged on the website:
>>>
>>> Bug reference: 14932
>>> Logged by: Todd Cook
>>> Email address: tcook@blackducksoftware.com
>>> PostgreSQL version: 10.1
>>> Operating system: CentOS Linux release 7.4.1708 (Core)
>>> Description:
>>>
>>> It hangs on a table with 167834 rows, though it works fine with only
>>> 167833
>>> rows. When it hangs, CTRL-C does not interrupt it, and the backend
>>> has to
>>> be killed to stop it.
>>>
>>
>> Can you share the query and data, so that we can reproduce the issue?
>>
>> Based on the stack traces this smells like a bug in the simplehash,
>> introduced in PostgreSQL 10. Perhaps somewhere in tuplehash_grow(),
>> which gets triggered for 167834 rows (but not for 167833).
>
> FWIW, changing the guts of hashint8() to
>
> + if (val >= INT32_MIN && val <= INT32_MAX)
> + return hash_uint32((uint32) val);
> + else
> + return hash_any((unsigned char *) &val, sizeof(val));
>
> allows us to process a full-sized data set of around 900 million rows.
> However,
> memory usage seemed to be rather excessive (we can only run 7 of these
> jobs in parallel
> on a 128GB system before the OOM killer kicked in, rather than the usual
> 24); if there's
> any interest, I can try to measure exactly how excessive.
>
I suspect you're right the hash is biased to lohalf bits, as you wrote
in the 19/12 message. In fact, I think it's a direct consequence of the
requirement that hashint8() needs to produce the same hash for logically
equivalent int2 and int4 values.
Out of curiosity, could you try replacing the hash_any call in hashint8
with a hash function like murmur3, and see if it improves the behavior?
That obviously breaks the hashint8 for cross-type hash joins, but it
would be interesting bit of information I think.
regards
--
Tomas Vondra http://www.2ndQuadrant.com
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services