It's not uncommon IME that doing the work to come up with a reasonable estimate is 80% of the final work.
E.g., addressing a scalability issue might require a bunch of profiling work, then a bunch of experimentation / thinking about developing a lock free algorithm. Or it might turn out that reducing false sharing a bit will give large enough benefit on its own. Without having done most of that I won't know whether it's a few days or many weeks.
>E.g., addressing a scalability issue might require a bunch of profiling work, then a bunch of experimentation / thinking about developing a lock free algorithm. Or it might turn out that reducing false sharing a bit will give large enough benefit on its own.
so this sounds like my example of my friend that I gave above who would say I will have to do a bunch of profiling work first and then a few days of investigation before I tell you how much it is actually going to take.
I don't say you can't run into uncertainties and be unable to estimate until you have clarified those uncertainties, I also don't say that you can't run into uncertainties in the actual implementation and have your estimate be woefully off, not every estimate needs to be correct.
> It's not uncommon IME that doing the work to come up with a reasonable estimate is 80% of the final work.
That estimate you do after finishing 80% of the work is probably orders of magnitude smaller than the real thing.
More realistically, if you don't have the experience for an estimate beforehand, you also don't have the experience to know when you covered enough of the high-risk subproblems, so you have no basis at all to decide how complete your research is. You will only have a reasonably safe estimate by the time you finish the task.
There's tasks where experience doesn't tell you ahead of time how long something takes. If you need to figure out a [novel take on an] algorithm, it'll be hard to predict the time even if you did things like that plenty times before. And you'll be able to judge the remaining work after pretty decently.
E.g., addressing a scalability issue might require a bunch of profiling work, then a bunch of experimentation / thinking about developing a lock free algorithm. Or it might turn out that reducing false sharing a bit will give large enough benefit on its own. Without having done most of that I won't know whether it's a few days or many weeks.