note: the data.table resutls shown above are not being reproduced by others even when they run 1 billion rows -- this is under investigation to understand why

Hi!

• This is the repo inspired by Gunnar Morlng's 1 billion row challenge to see which functions / libraries are quickest in summarizing the mean, min and max of a 1 billion rows of record

• Actually, Gunnar's efforts are more robust as its a challenge to the Java community to optimize their code to reduce the processing speed whereas I'm just taking available packages and comparing them

  • I tried to download the orignal Java repo but since I know absolutely nothing about Java this predictiblity failed
  • So this isn't the same dataset as the original challenge but it is billion rows of record which I think is the important part of the challenge

• I used the micro benchmark library as the benchmarking library

• Please feel free to fork or copy the repo and reproduce the findings

• A few notes:

  • The underlying file is large (~20GB) so that isn't in the repo, however the generate_data.R script will produce this for you from your command line (see instructions)
  • I ran the analysis in a Saturn Cloud cluster, my computer struggled with benchmarking billion rows of data multiple times (you can specify how many rows of data you want generated if you want to run it locally)
  • Cluster details are: A10G-G54XLarge - 16 cores - 64 GB RAM - 1 GPU - 40Gi Disk
  • I'm fairly new to tidypolars and data.table so if you see any issues with the script that I wrote please tell me!
  • The general format is that each function test is aggregate by their data workflow eg. since dplyr and collapse use regular data.frames as their input they are kept together and since duckdb and arrow can leverage the same datasource, they are also kept together
  • I added collect() or whatever the equivalent function is to most of the functions so that they are all returned a tibble format however this may add some variation in the results

Instructions

• Using any flavor of a linux command line run ./generate_data.R in your command line with the input of rows you want created eg:

$ ./generate_data.R 1e9

note: you may need to change the file permission $ chmod +x generate_data.R

• This will create a new folder called "data" which will contain the measurement.txt file

• From there, you can either run each test script individually (eg. dt.R) or you can run them all together via the 1br.R which will source all the scripts

• The format of each test script is largely the same

  • load libraries
  • read data (using data.table::fread)
  • convert data appropriate format (if necessary)
  • create functions that summarize mean, min, max by the respective libraries
  • benchmark results 10 times
  • save results in csv file

• Additionally there is a function to covert the ggplot theme to the bbc plot theme

• There were alot of questions about data.table's benchmark so I'm trying to build an alternative benchmark function using the tictoc package but its wip (benchmark_fn.R)

• If you see any issues or have suggestions of improvements, please let me know!