On 08.01.2013 17:09, Jim Jagielski wrote: > Rainer, does the below address your concern and make > it OK that the current behavior acts the way it does > and we're OK to keep it that way.
Yes sort of. We now know/remember why there are subtleties and since some statistical counters are dual purpose - statistics/monitoring and LB decision data it seems we need to give up a 100% consistency goal. I'm OK with that as it is now. Did you have a look at the other topic about the "Used" shm slot counter? Regards, Rainer > On Jan 8, 2013, at 10:25 AM, Jim Jagielski <[email protected]> wrote: > >> Yeppers... >> >> actually, transferred and read *are* persisted, it's just >> that the bytraffic lbmethod is smart enough to know that >> when the config is changed, those counters need to >> be reset and, by default, a restart is a "change". >> Since elected is never used for any sort of lbmethod, >> it never is reset and so the persist "holds". >> >> So the idea is that lbmethods should reset any potential >> persisted values iff it makes sense for that lbmethod. >> bytraffic resets transferred/read but byrequests doesn't, >> for example. >> >> On Jan 8, 2013, at 9:49 AM, Jim Jagielski <[email protected]> wrote: >> >>> Hold on a tic... let me go thru my notes. I seem to recall >>> an issue I hit. >>> >>> On Jan 8, 2013, at 9:28 AM, Jim Jagielski <[email protected]> wrote: >>> >>>> +1 for both being reset... >>>> >>>> On Jan 7, 2013, at 4:40 AM, Rainer Jung <[email protected]> wrote: >>>> >>>>> On 06.01.2013 17:48, Jim Jagielski wrote: >>>>>> I had thought that I has responded to that orig >>>>>> email that both below issues where "by design" >>>>>> but could be adjusted if need be or desired. >>>>>> >>>>>> I have no opinions either way, but to be Used seems >>>>>> more session based and Elected is more long-term >>>>>> statistical. >>>>> >>>>> Hmmm, but the "Used" field should give information about shm slots >>>>> occupied by configured balancers. Isn't the info (=0) wrong after a >>>>> restart? >>>>> >>>>> Concerning the long term statistical values I'm undecided like you, but >>>>> I think we should handle them in a consistent way, so "traffic" and >>>>> "elected" should either be both persisted or both reset. >>>>> >>>>> Regards, >>>>> >>>>> Rainer >>>>> >>>>>> On Jan 5, 2013, at 1:41 PM, Rainer Jung <[email protected]> wrote: >>>>>> >>>>>>> On 04.01.2013 19:48, Jim Jagielski wrote: >>>>>>>> Have people had a chance to test, review and try the balancer >>>>>>>> persist and inheritance stuff in trunk? I want to make >>>>>>>> sure that we have some level of verification and agreement >>>>>>>> there before I work on the backports for 2.4 ;) >>>>>>> >>>>>>> I didn't see any changes with respect to two of my original comments: >>>>>>> >>>>>>> The "Used" counters (number of shared mem slots) in balancer manager >>>>>>> drops to "0" after restart/reboot´with persist on. This seems strange. >>>>>>> Not that "Used" does *not* have anything to do with request counting. >>>>>>> >>>>>>> Second the "Elected" counter is persisted but e.g. the traffic counter's >>>>>>> not. This seems somewhat inconsistent. I have no strong opinion whether >>>>>>> to persist statistics or not, but we might want to behave consistently. >>>>>>> >>>>>>> I did not observe any functional changes after my Mail from Dec. 14, >>>>>>> right? >>>>>>> >>>>>>> Regards, >>>>>>> >>>>>>> Rainer
