This project provides tests and benchmarks of the Go library "stats" using the data set Anscombe's quartet. It includes benchmarks of the same operations in Python and R for comparison.
This project can be cloned through git with the following command:
git clone https://github.com/Meowcenary/stats_benchmark.git
In order to run the Python and R benchmarks you will need to install Python
3.7.11 and R 4.3.1. Additional dependencies for Python can be installed from
within the anscombe_quartet_benchmark_py/ with:
pip install -r requirements.txt.
The tests for the stats package can be viewed in in stats_test.go. The tests
can be run From the root directory of the project with go test
Go benchmarks, which are included alongside the tests in stats_test.go, can be
run from the root directory of the project with go test -bench=.
The Python benchmarks can be run by switching to the directory
anscombe_quartet_benchmark_py/, installing dependencies from requirements.txt
(if you have not already done so), and finally running the script with
python3 anscombe_quartet_benchmark.py
The R benchmarks can be run by switching to the directory
anscombe_quartet_benchmark_r/ and running the script with
Rscript anscombe_quartet_benchmark.R
The Go library "stats" calculates linear regression coeffcients, intercepts, and
gradients for the data set "Anscombe's quartet" that align with the output from
the provided Python and R scripts. The tests and data can be viewed within
stats_test.go.
Each script was benchmarked ten times with the fastest, slowest, and average time recorded below.
The Go "stats" library had an average run time of 0.0000025 ns with the slowest time at 0.0000029 ns and the fastest time at 0.0000020 ns.
The Python script had an average run time of 5.838 ns with the slowest time at 6.01 ns and the fastest time at 5.69 ns.
Finally, the R script had an average run time of 19190070 ns with the slowest time at 20943880 ns and the fastest time at 18280030 ns.
These comparisons are from relatively small data sets and at scale the gap between Go's run time versus both Python and R will only widen.
One of the most common concerns with migrating from languages like Python or R to Go is the learning curve of Go's syntax, best practices, and tools. While Go can be intimidating to look at compared to the syntax of Python or R, the relatively small language specification and philosophy of having a single solution to a problem leads to simple solutions that can be easily understood even with little experience in the language. There are also many books written on learning Go and thorough documentation that reduce the learning curve through examples. Regarding tooling, Go's installation includes standard tools for packaging, testing, and linting. By providing a standard out of the box there is no need to look for external tools to handle these common tasks which is convenient, removes the need to argue for one tool versus another, and avoids adding external dependencies.
Another common concern with migrating programming languages is finding replacements for libraries. The library "stats" tested in this repository with Go's standard packages provides everything needed to produce the output from the Python and R scripts. Though R does not require additional libraries, Go only require one while Python requires three. While R has the functionality built in, the run time is significantly slower. With the expectation that large data sets will need to be analyzed, efficiency is the main priority and the cost of a single dependency is outweighed by Go's superior efficiency.
Based on the efficiency, small language specification, readily available learning materials, and testing of the stats library it is my recommendation to use the Go programming language for future development.