Counters don't work because of the asynchronous nature of Storm. For example, consider a topology that looks like this:
A -> B -> C
\-> D
Let's say A emits 2 tuples (+2 differential), B processed those and emits 2 to C and 3 to D (+3 differential), C processes 2 tuples (-2 differential), and D processes 3 tuples (-3 differential).Everything's asynchronous, so the acker could receive the acks in this order: A, C, B, D
The counter would then look like this: 2, 0, 3, 0
So it would think that the tuple was complete before it actually was, which means the counter algorithm doesn't work.
Which means, due to the birthday paradox, that you should expect your first collision around 2^32 tuples. Process a few million tuples a day (that's only about 25 per second) and you should expect your first error within your first year.
EDIT: I'm wrong, see the explanations below.
By the time you reach your 2^32 ack, your chance of having all successes is ((2^64-1)/(2^64))^2^32, which is < 0.5.
EDIT: My test program seems to indicate that I'm wrong, but I can't see the flaw either in it or in my reasoning. Can you explain why the birthday paradox doesn't apply here?
EDIT': In the birthday paradox, it would be ((2^64-1)/2^64)) * ((2^64-2)/2^64) * ((2^64-3)/2^64), etc.
Hilariously, you got your equation right (which isn't the birthday paradox equation btw) but got your result wrong. http://www.wolframalpha.com/input/?i=%28%282%5E64-1%29%2F%28...
So the chance of success after 2^32 acks is >0.999999999, not <0.5.
It takes about 2^60 acks before there's a significant chance of a mistake. That's a lot of acks, so it will take an insanely long time for a mistake to be made.
Suppose that the sequence of values you have after each ack is A, B, C, D, E (five acks total). So long as A, B, C, and D are all nonzero, we're OK. With the birthday paradox, we'd be looking at A=B, A=C, A=D, A=E, B=C, B=D, etc. -- many more combinations.