On 12/21/2014 02:12 PM, Tom Lane wrote:
> Andrew Dunstan <andrew@dunslane.net> writes:
>> On 12/21/2014 01:23 PM, Alvaro Herrera wrote:
>>> The point, I think, is that without atomic instructions you have to hold
>>> a lock while incrementing the counters.
>> Hmm, do we do that now?
> We already have a spinlock mutex around the counter adjustment code, so
> I'm not sure why this discussion is being held.
Because Peter suggested we might be able to use atomics. I'm a bit
dubious that we can for min and max anyway.
>
>> I would like someone more versed in numerical analysis than me to
>> tell me how safe using sum of squares actually is in our case.
> That, on the other hand, might be a real issue. I'm afraid that
> accumulating across a very long series of statements could lead
> to severe roundoff error in the reported values, unless we use
> a method chosen for numerical stability.
>
>
Right.
The next question along those lines is whether we need to keep a running
mean or whether that can safely be calculated on the fly. The code at
<http://www.johndcook.com/blog/standard_deviation/> does keep a running
mean, and maybe that's required to prevent ill conditioned results,
although I'm quite sure I see how it would. But this isn't my area of
expertise.
cheers
andrew