How to shuffle a big dataset (2018)
blog.janestreet.com
How to shuffle a big dataset (2018)
1–10 of 15 posts
Re: How to shuffle a big dataset (2018)
#2If you have a system with a fast sort/order by operation but a slow (or missing) shuffle, one trick that I sometimes use is to generate a new random number for each record, then sort by that random number.
Uniform sampling is also quite easy (and principled) - you can just take the first n records.
Re: How to shuffle a big dataset (2018)
#3I did not see it mentioned in the article, but I wonder why or if the team rejected a permutation function on the index. The shuffled dataset would then be Xs[i] = X[f(i)]
Re: How to shuffle a big dataset (2018)
#4I did not see it mentioned in the article, but I wonder why or if the team rejected a permutation function on the index. The shuffled dataset would then be Xs[i] = X[f(i)]
Do you have a cheap permutation function for large n? It seems like you still have to do it in two passes if you do this.
One reason cited in TFA for the half-shuffled files approach is that it's easy to rotate old data out of and new data into the half-shuffled files.
Re: How to shuffle a big dataset (2018)
#5My immediate thought before reading the article is that shuffling is the inverse of sorting. To put another way, you can take a sorted list and apply a merge sort on it, but the comparison operation is random. Merge sort can work on data sets that are too large to fit in memory since you sort small sized blocks that do fit in memory, then merge using streaming
Re: How to shuffle a big dataset (2018)
#6I did not see it mentioned in the article, but I wonder why or if the team rejected a permutation function on the index. The shuffled dataset would then be Xs[i] = X[f(i)]
Do you have a cheap permutation function for large n? It seems like you still have to do it in two passes if you do this. One reason cited in TFA for the half-shuffled files approach is that it's easy to rotate old data out of and new data into the half-shuffled files.
The permutation function is essentially a block cipher on the index. It's stateless and there is no need for intermediate files or buffers. The cost of running the function should be comparable to RNG.
Re: How to shuffle a big dataset (2018)
#7Can you randomize memory pointers or something?
Re: How to shuffle a big dataset (2018)
#8My immediate thought before reading the article is that shuffling is the inverse of sorting. To put another way, you can take a sorted list and apply a merge sort on it, but the comparison operation is random. Merge sort can work on data sets that are too large to fit in memory since you sort small sized blocks that do fit in memory, then merge using streaming
You can't just randomize the comparison operation, since that'd violate the requirements of a comparison function. MS made that mistake in their browser download dialog, which led to significant biases. You'd need to assign a random number to each item. But then you'd need to figure out a way to store this random number.
Also, for in-memory data comparison-based sorting is much slower than shuffling. Both asymptotically (O(n log n) vs O(n)) and in practice. I did not investigate if sorting algorithms not based on comparisons (e.g. radix sort) have competitive performance.
https://www.computerworld.com/article/2762287/microsoft-s-eu...
Re: How to shuffle a big dataset (2018)
#9I did not see it mentioned in the article, but I wonder why or if the team rejected a permutation function on the index. The shuffled dataset would then be Xs[i] = X[f(i)]
Do you have a cheap permutation function for large n? It seems like you still have to do it in two passes if you do this. One reason cited in TFA for the half-shuffled files approach is that it's easy to rotate old data out of and new data into the half-shuffled files.
There are format preserving encryption algorithms which act as a pseudo random permutation of integers with a variable upper bound. These slower than Fisher-Yates for in memory data, but for on disk data the random access overhead should exceed their cost. The advantage of this approach is minimal memory use and no need to modify the data.
The downside is that it still needs one random read access per element, so cache friendly hierarchical algorithms, like the one described in the post, are probably still faster for on disk data.
Re: How to shuffle a big dataset (2018)
#10My immediate thought before reading the article is that shuffling is the inverse of sorting. To put another way, you can take a sorted list and apply a merge sort on it, but the comparison operation is random. Merge sort can work on data sets that are too large to fit in memory since you sort small sized blocks that do fit in memory, then merge using streaming
Shuffling and sorting don’t seem related at all. Naively, I would expect shuffle to be linear in time complexity.