Hello Ionut,
I actually wrote a long explanation here: Leaderboard comparability after the June 8 real-time data-access fix — were pre-fix scores rescored? - #2 by enzo
In short, trying to persist state across time series will likely result in your code not being deterministic based on when it starts.
- Depending on the number of
INFER_PARALLELISM, if you run the full dataset, the runner will start at regular intervals. - However, during the determinism dataset, those offsets will be unique.
Here is a visual representation:
(Feedbacks are welcome as I am not the best designer.)
There seems to be a general misunderstanding of this mechanism. I should add it to the documentation.
P.S.: It’s only 10%, not 30%. I recently found out that the documentation was outdated, but the change happened a month after the competition started.