I ignored GB advice and am suffering the consequen...
# ask-questions
I ignored GB advice and am suffering the consequences. Can y'all just confirm one thing for me? I changed traffic allocation mid-test to shut down one experiment variant--without sticky bucketing, without starting new phase / re-seeding. It appears as though the hash range freed by the stopped variant has NOT been evenly allocated across remaining variants. Is this the case?
Yes, that's correct. As an example, imagine you went from a 33/33/34 test to a 50/50/0 test (essentially turning off the last variation). The hash ranges would change like this:
Copy code
Original hash ranges:
[0, 0.33], [0.33, 0.66], [0.66, 1.0]

New hash ranges:
[0, 0.5], [0.5, 1], [1, 1]
So in this case, anyone with the range
[0.33, 0.5]
would switch from variation 1 to the control. Everyone from
[0.66, 1]
would switch from the stopped variation to variation 1.
Yep, that checks out. Oopsy!
The reason we still allow this in the UI despite the warnings is because of 2 scenarios where this isn't that big of a deal: 1. Really minor adjustment of traffic weights. Like going from 50/50 to 49/51. Barely anyone is going to switch variations in this case, so it's not a big deal. 2. If you catch a bug super early into a test, like in the first hour. That small amount of traffic will be a drop in the bucket compared to what you'll collect over the next week+, so again not a big deal if a few of those users switch variations.
Actually, one more scenario. If you had turned off the middle variation above instead, it would have worked out by coincidence that the stopped variation's traffic would evenly distribute between the remaining variations, just due to how the ranges happen to be calculated
Yep. Now I understand the reason for the strong warnings. Without thinking too hard about it, I assumed the allocation would be partitioned out evenly, but I guess that doesn't make sense considering the many different ways people could try to change allocations.
I need to get sticky bucketing deployed stat!
Thanks for the explanations @future-teacher-7046