On Sat, Jun 5, 2010 at 6:07 PM, Jon Schewe <jpschewe@mtu.net> wrote:
> On 06/05/2010 07:02 PM, Scott Marlowe wrote:
>> On Sat, Jun 5, 2010 at 5:58 PM, Jon Schewe <jpschewe@mtu.net> wrote:
>>
>>> On 06/05/2010 06:54 PM, Scott Marlowe wrote:
>>>
>>>> On Sat, Jun 5, 2010 at 5:03 PM, Jon Schewe <jpschewe@mtu.net> wrote:
>>>>
>>>>
>>>>> On 06/05/2010 05:52 PM, Greg Smith wrote:
>>>>>
>>>>>
>>>>>> Jon Schewe wrote:
>>>>>>
>>>>>>
>>>>>>>> If that's the case, what you've measured is which filesystems are
>>>>>>>> safe because they default to flushing drive cache (the ones that take
>>>>>>>> around 15 minutes) and which do not (the ones that take >=around 2
>>>>>>>> hours). You can't make ext3 flush the cache correctly no matter what
>>>>>>>> you do with barriers, they just don't work on ext3 the way PostgreSQL
>>>>>>>> needs them to.
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>> So the 15 minute runs are doing it correctly and safely, but the slow
>>>>>>> ones are doing the wrong thing? That would imply that ext3 is the safe
>>>>>>> one. But your last statement suggests that ext3 is doing the wrong
>>>>>>> thing.
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>> I goofed and reversed the two times when writing that. As is always
>>>>>> the case with this sort of thing, the unsafe runs are the fast ones.
>>>>>> ext3 does not ever do the right thing no matter how you configure it,
>>>>>> you have to compensate for its limitations with correct hardware setup
>>>>>> to make database writes reliable.
>>>>>>
>>>>>>
>>>>>>
>>>>> OK, so if I want the 15 minute speed, I need to give up safety (OK in
>>>>> this case as this is just research testing), or see if I can tune
>>>>> postgres better.
>>>>>
>>>>>
>>>> Or use a trustworthy hardware caching battery backed RAID controller,
>>>> either in RAID mode or JBOD mode.
>>>>
>>>>
>>> Right, because the real danger here is if the power goes out you can end
>>> up with a scrambled database, correct?
>>>
>> Correct. Assuming you can get power applied again before the battery
>> in the RAID controller dies, it will then flush out its cache and your
>> data will still be coherent.
>>
> Or if you really don't care if your database is scrambled after a power
> outage you can go without the battery backed RAID controller.
I do that all the time. On slony replication slaves. You can use a
considerably less powerful machine, IO wise, with fsync disabled and a
handful of cheap SATA drives.