Commits on Oct 11, 2015

1. promql: Remove extrapolation from rate/increase/delta. …

   This is an improvement from both a theoretical and practical
   standpoint.
   
   Theory:
   For simplicty I'll take the example of increase().
   
   Counters are fundamentally lossy, if there's a counter reset
   or instance failure between one scrape and the next we'll
   lose information about the period up to the reset/failure.
   Thus the samples we have allow us to calculate a lower-bound
   on the increase() in a time period.
   
   Extrapolation multiples this by an amount depending on timing which is
   an upper bound based on what might have happened if everything continued
   as it was.
   This mix of upper and lower bounds means that we can't make any
   meaningful statements about the output of increase() in relation to what
   actually happened.
   
   By removing the extrapolation, we can once again reason that the result
   of increase() is a lower bound on the true increase of the counter.
   
   Practical:
   Fixes #581.
   The extrapolation presumes that it's looking at a range within a
   continuous stream of samples. If in fact the time series starts or
   end within this range, this leads to an over-correction.
   
   For discrete counters and gauges, extrapolating to invalid values in
   their domain can be confusing and prevent rules being written that
   depend on exact values.
   
   For those looking to graph things more accurately irate() is a better
   choice than extrapolation on rate().
   For those looking to calculate how a gauge is trending deriv() is a
   better choice than delta().

   

   brian-brazil committed Oct 11, 2015
   Configuration menu

   • View commit details
   • Copy full SHA for 06c55d7
   • Browse repository at this point

   Copy the full SHA

   06c55d7 View commit details

   Browse the repository at this point in the history