reshape task (pivot, unpivot)

h2oai / db-benchmark

reproducible benchmark of database-like ops

https://h2oai.github.io/db-benchmark

Mozilla Public License 2.0

321 stars 85 forks source link

reshape task (pivot, unpivot) #175

Open jangorecki opened 3 years ago

jangorecki commented 3 years ago

nice examples by @grantmcdermott can be found https://grantmcdermott.com/reshape-benchmarks/ https://grantmcdermott.com/even-more-reshape/

grantmcdermott commented 3 years ago

Agree: I think that a reshaping benchmark is an important addition to the list. Happy to add a PR with my examples as-is if that helps? Some quick thoughts/issues:

I'd drop the Stata runs — I'm guessing you don't have the license reqs — leaving DT, dplyr, pandas, and DataFrames.jl implementations.
My examples are only wide-to-long, but easy enough to add a complement going long-to-wide (i.e. back to the original dataset).
My dataset is deliberately sparse (lots of missing obs). Would you want the same thing for this benchmark?

jangorecki commented 3 years ago

@grantmcdermott Thank you for your comment. Yes, Stata needs to be dropped, we stick to open source software. No need PR, but eventually some assistance in reviewing design may be useful.

Ideally reshape task should test:

melt
dcast
95%, 5%, 0% missing
different functions applied during dcast
multiple columns on id side
multiple columns on measure side
probably quite few other features (need to look at common usage patterns on SO)

All that needs to be categorized into: 5 "basic" and 5 "advanced" queries. So the scope will be way bigger than your posts, yet your posts are very useful working example to start on them.

grantmcdermott commented 3 years ago

Sounds good. Lmk if and when you'd like someone to cast an extra eye over the tests.

skanskan commented 3 years ago

I came here to suggest we also need to benchmark reshaping times. Glad to see other thought the same.