On Thu, Jul 11, 2013 at 10:48 AM, Branko Čibej <br...@wandisco.com> wrote: > I suspect that's completely irrelevant, since the probe issues OPTIONS > requests which are very small and should never spill to disk. The big > requests would be GET/PUT of large files and/or deltas, IIRC -- but > those shouldn't have to be retried if the probe does the right thing.
I think you misunderstood the purpose of the test. The tests I did won't show any significant change from the extra OPTIONS request (I could have just did no and yes but I did auto for completeness) because the latency is so low (<1ms) and as you point out the spillbuff doesn't matter for OPTIONS. Greg is concerned that people may set the tristate to no and then eat performance loses from not using chunking when they could be. So the point here is to show what that impact would be, which based on the tests I did is minimal.
