Re: [rfc] overhauling pgstat.stat

(2013/09/09 8:19), Tomas Vondra wrote:
> On 8.9.2013 23:04, Jeff Janes wrote:
>> On Tue, Sep 3, 2013 at 10:09 PM, Satoshi Nagayasu <snaga(at)uptime(dot)jp>
>> wrote:
>>> Hi,
>>>
>>>
>>> (2013/09/04 13:07), Alvaro Herrera wrote:
>>>>
>>>> Satoshi Nagayasu wrote:
>>>>
>>>>> As you may know, this file could be handreds of MB in size,
>>>>> because pgstat.stat holds all access statistics in each
>>>>> database, and it needs to read/write an entire pgstat.stat
>>>>> frequently.
>>>>>
>>>>> As a result, pgstat.stat often generates massive I/O operation,
>>>>> particularly when having a large number of tables in the
>>>>> database.
>>>>
>>>>
>>>> We already changed it:
>>>
>>>>
>>>> commit 187492b6c2e8cafc5b39063ca3b67846e8155d24 Author: Alvaro
>>>> Herrera <alvherre(at)alvh(dot)no-ip(dot)org> Date: Mon Feb 18 17:56:08
>>>> 2013 -0300
>>>>
>>>> Split pgstat file in smaller pieces
>>>
>>> Thanks for the comments. I forgot to mention that.
>>>
>>> Yes, we have already split single pgstat.stat file into several
>>> pieces.
>>>
>>> However, we still need to read/write large amount of statistics
>>> data when we have a large number of tables in single database or
>>> multiple databases being accessed. Right?
>>
>> Do you have a test case for measuring this? I vaguely remember from
>> when I was testing the split patch, that I thought that after that
>> improvement the load that was left was so low that there was little
>> point in optimizing it further.
>
> This is actually a pretty good point. Creating a synthetic test case is
> quite simple - just create 1.000.000 tables in a single database, but
> I'm wondering if it's actually realistic. Do we have a real-world
> example where the current "one stat file per db" is not enough?

I have several assumptions for that.

- Single shared database contains thousands of customers.
- Each customer has hundreds of tables and indexes.
- Customers are separated by schemas (namespaces) in single database.
- Application server uses connection pooling for performance reason.
- Workload (locality in the table access) can not be predicted.

Looks reasonable?

> The reason why I worked on the split patch is that our application is
> slightly crazy and creates a lot of tables (+ indexes) on the fly, and
> as we have up to a thousand databases on each host, we often ended up
> with a huge stat file.
>
> Splitting the stat file improved that considerably, although that's
> partially because we have the stats on a tmpfs, so I/O is not a problem,
> and the CPU overhead is negligible thanks to splitting the stats per
> database.

I agree that splitting a single large database into several pieces,
like thousands of tiny databases, could be an option in some cases.

However, what I intend here is eliminating those limitations on
database design.

In fact, when considering connection pooling, splitting a database
is not a good idea, because AFAIK, many connection poolers manage
connections per database.

So, I'd like to support 100k tables in single database.

Any comments?

Regards,

> But AFAIK there are operating systems where creating a filesystem in RAM
> is not that simple - e.g. Windows. In such cases even a moderate number
> of objects may be a significant issue I/O-wise. But then again, I can't
> really think of reasonable a system creating that many objects in a
> single database (except for e.g. a shared database using schemas instead
> of databases).
>
> Tomas
>
>

--
Satoshi Nagayasu <snaga(at)uptime(dot)jp>
Uptime Technologies, LLC. http://www.uptime.jp