Open ashrayjain opened 5 years ago
@rcrowley @mihasya
Friendly ping @rcrowley @mihasya
This looks like a very good change, but it makes me wonder if folks have come to rely on the old behavior, and if changing this is a "backwards incompatible" change 🤔 Also trying to think of a case where this might cause a performance issue.. In a set up where you have lots of rarely updated counters that are read regularly, you could suddenly find yourself recomputing a bunch of EWMAs for things that were previously "free." Hmmm.
I do see a bunch of votes for this, and have definitely been annoyed by it in the past.
I'm worried the fix is not complete - Snapshot()
is a read, so it needs rescaling, too. Otherwise, the reads performed on a Snapshoted metric will not trigger the rescaling.
If it seems out of scope of this MR, I could create another one. @ashrayjain, @mihasya what do you folks think about it?
any news on this ?
@p-kozlowski you are correct. I have updated this to also rescale on Snapshot()
Previously, we were only rescaling on updates. This meant that histograms would latch on to last updated values and stay at those values until the next update for a reader who reads the histogram periodically instead of decaying with time.
Fixes #215