Why do ppl assume that estimates are necessary? I don't see the need for them myself. They're solid waterfall thinking—big upfront plans, milestones, months-long delivery cycles. Not much agility. It's deployable right now. It will be deployable w/ this new capability tomorrow.
@allenholub We always do at least subconscious estimates because there are always more likely improvements than we have the time to make.
Consider [this now-obsolete extreme example from xkcd](https://xkcd.com/1425/). Wouldn't have been reasonable to equally prioritize working on both of these possible features; an afternoon or a week spent on the bird-recognition feature would have been only wasted.
In less extreme cases, an hour or two of estimation work can tell us whether a feature is more likely to take two days or a month. Many features that are worth trading off two days of other features for are not worth trading off a month of other features for.
Unfortunately, almost no companies have incentives set up to reward doing that estimation work, so formal estimates add little more value than subconscious ones.
@allenholub I agree that is more extreme than most cases in practice, but is an extreme version of a universal problem in programming: we never *really* know what's hard and what isn't until we're done. Our Bayesian belief about completion time has a lognormal distribution with usually enormous variance, sometimes until the very moment of completion.
Often, though, a spike provides much information to reduce that variance, and is often valuable.
@radehi
Agreed, but often we _do_ know what we don't know. Unknown unknowns are rare. Many have observed that the distribution goes down as we approach zero. The shorter the estimate, the more accurate it seems to be. That's a motivation to narrow the scope of all stories as much as possible.
@allenholub In a Bayesian sense there's no such thing as a completely unknown unknown because there's no uniform probability distribution over the whole real line or the whole positive real line. You probably don't know the volume of Lake Superior (to take an example from earlier today) but probably know is more than a liter and less than a googol liters. An unknown is always just a question of the probability distribution.
I agree that narrowing the scope of stories to the minimal possible scope is very valuable.
@radehi That comic is really about how customers don't know what's hard and what isn't, though. If you look at the vast majority of stories, narrowing as much as possible can be done without estimation. Then sort the stories into two buckets: we can do this, and we don't know enough to know if we can do this. The latter requires a research project, which is pbly not estimable.