Comment by sestep

The shuffle was only unacceptably slow for data too big to fit in memory. For data that fits in memory, Fisher-Yates is totally fine; this is why it's fine for the two-pass shuffle to use buckets that fit in RAM but not in cache.