Statistics dumps.

[Erik Rozendaal]

Måns Nilsson wrote:
> 
> --On torsdag 28 juli 2005 16.56 +0200 Mans Nilsson <mansaxel at sunet.se>
> wrote:
> 
> 
>>Subject: Re: Statistics dumps. Date: Thu, Jul 28, 2005 at 04:38:07PM
>>+0200 Quoting Erik Rozendaal (erik at nlnetlabs.nl):
>>
>>>NSD also dumps statistics when the process exits. So the shorter dump 
>>>interval can be explained by database reloads (which causes NSD to 
>>>restart itself).
>>
>>Ok, this explains things. 
> 
> 
> An additional question: As I understand the log format, the second time_t
> value on a ?STATS line is the start of the global period, while the first
> is dumptime, and all querycounts are since the second time_t. This means
> that one can derive the average query count since initiation per second by
> doing something like 
> 
>     qcount_since_start
>    --------------------
> (time_field_1 - time_field_2)

True, except for the case a reload has happened.  In this case all the 
counters are reset to 0, but the second time_t is not reset.  This may 
be a bug...

> ...but if I want to get the qps in 5 minute intervals (and take a stats
> dump every 300 secs to achieve that resolution.) I can do post-processing
> by saving the last dump time and doing some awk math on that (which would
> catch the not-at-300-sec-intervals dumps) but I still get strange value
> dips (again refering to the graph URL in my last mail) as soon as I've
> gotten an extra reload. It might be my math, of course, but your opinions
> on the best sampling method to get qps in 5 minute or better resolution
> from NSD stats dumps would be most appreciated. 

I'm not sure if you can get the math right without also knowing a 
reboot/reload happened.  And detecting a reboot/reload is not easy 
without the pid information.

Erik