Comment by sh1mmer

I think this was a nice article and helped walk through some of the maybe unintuitive maths but I feel like they missed the meat a bit.

With random placement, as was said in the article, each flip is independent. With PoTRC it’s using knowledge of current queue depth.

The author didn’t explain the challenges around tracking queue depth, or why PoTRC is better than say keeping an ordered list of queues depth and picking the smallest one.

I’m assuming the trade off is about cost of managing the sorted list of queues depths vs just O(1) look ups of the random nodes’ depths.

It just would’ve been nice to cover that in the article.