Rather than “buckets”, I bet they just have a flag on each tweet in the database for whether it’s deleted. So if you delete the most recent 3600, next time they do a query to see recent tweets, it retrieves 3600, but they’re all deleted, so it filters them all out. Buckets are a little less likely implementation wise because you’d either sometimes see that the most recent bucket had very few tweets in it, or they’d have to be constantly shuffling tweets between buckets.