In addition to choosing the right way to look at the distribution of
total action times, I strongly recommend breaking down the transactions
into component parts and looking at the details. The exchange I had
with Sam Hartman was a good example. On 5/23. Sam wrote:
> I'd think two months would be doing good for IESG processing. That
> includes AD review, IETf last call, telechat delays and a bit of slop
> for interaction with the authors. However that's time to travel
> through the queue, not actually time spent on that document.
>
> Because of the delays involved with telechats and last call, getting a
> WG document done in less than a month of wall time or an individual
> document done in less than 1.5 months is very unlikely.
What catches my attention -- and brings shudders as I remember my time
as an AD -- is that he paints a picture that looks like
T[IESG] = T[AD] + T[last call] + T[telechat delay] + T[authors]
I see two problems with this formulation:
(1) T[AD] is in fact composed of several separate things, and the
differences between them are almost certainly critical in understanding
process delays. In particular, AD time is spent in several different ways:
(a) Initial document review (done by all ADs)
(b) AD followups (done by the sheparding AD)
(c) AD rereview and discussion to clear discussses (done by all ADs)
There are timeouts associated with (a) - once a document is on the
IESG's agenda, ADs are required to have their review done by the next
telechat. They can get an extension, but only once. This sohuld therefore
end up being similar to telechat timinig.
There are, however, no timeouts associated with (b) and (c). And more
to the point, the pressures on sheparding AD are quite different than
those on the other ADs to act in a timely fashion.
My experience has been that the distribution for (b) is bimodal (or
trimodal if you count the case where the document just passes without
any discusses). In some cases the sheparding AD only has to
pass on the discuss comments to the authors, done. In others, however,
the AD has to spend a bunch of time in effect figuring out what
amounts to a compromise. This can take a lot of time.
In my experience (c) is the truly problematic case. In my experience
it is both highly variable and the average time taken seems to be much
longer than it should be. Additionally, as sheparding AD I often found
it was only possible to clear discusses by explicitly bringing the
document up on a subsequent telechat, so there's a periodicity here as
well.
I have in the past advocated putting a timeout on (c), and I continue to
think this is a really good idea.
(2) When I was an AD I went to a good deal of trouble to overlap the
various phases of the process whenever possible in order to save time.
Nothing prevents ADs from entering comments early. Nothing prevents
documents from being revised during last call. I don't think I was
terribly successful in this, but much of my time as AD was spent without
the tools we have now. I therefore think the additive nature of your
formula has to be characterized as a worst case, and it needs to be
noted that substantial gains can be had by overlapping terms.
Now each of these terms has quite different characteristics. The last
call and telechat delays add up to a few weeks, to the overall time is
at least a couple of months. But those terms probably have moderately
low variance. On the other hand, the T[AD] and T[authors] terms have
very low minimums but extremely high variance, and it's within those
terms that the real issues emerge.
Agreed.
I suspect if you break those terms down even further, interesting and
useful dynamics will emerge.
Exactly so. Breaking them down further is IMO essential to getting
anything useful out of the analysis.
Ned
_______________________________________________
Ietf mailing list
Ietf(_at_)ietf(_dot_)org
https://www1.ietf.org/mailman/listinfo/ietf