On 04/27/2017 07:35 PM, Andres Freund wrote:
> On 2017-04-27 10:29:48 -0700, Joshua D. Drake wrote:
>> On 04/27/2017 09:34 AM, Andres Freund wrote:
>>> On 2017-04-27 09:31:34 -0700, Joshua D. Drake wrote:
>>>> On 04/27/2017 08:59 AM, Andres Freund wrote:
>>>>
>>
>>>> I would agree it isn't yet a widespread issue.
>>>
>>> I'm not yet sure about that actually. I suspect a large
>>> percentage of people with such workloads aren't lingering lots on
>>> the lists.
>>
>> That would probably be true. I was thinking of it more as the
>> "most new users are in the cloud" and the "cloud" is going to be
>> rare that a cloud user is going to be able to hit that level of
>> writes. (at least not without spending LOTS of money)
>
> You can get pretty decent NVMe SSD drives on serveral cloud
> providers these days, without immediately bancrupting you. Sure, it's
> instance storage, but with a decent replication and archival setup,
> that's not necessarily an issue.
>
> It's not that hard to get to the point where postgres can't keep up
> with storage, at least for some workloads.
>
I can confirm this observation. I bought the Intel 750 NVMe SSD last
year, the device has 1GB DDR3 cache on it (power-loss protected), can do
~1GB/s of sustained O_DIRECT sequential writes. But when running
pgbench, I can't push more than ~300MB/s of WAL to it, no matter what I
do because of WALWriteLock.
regards
--
Tomas Vondra http://www.2ndQuadrant.com
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services