leftPadDeez

nuts

Introduction

leftPad, an 11 line piece of code that broke the internet turns out had a second incident... and it involves me... ... getting rolled

https://qz.com/646467/how-one-programmer-broke-the-internet-by-deleting-a-tiny-piece-of-code

So... maybe i made a mistake where i accidentally code a "faster" of leftpad live where the results... were not faster...

https://www.youtube.com/watch?v=NmHUjxKpD90

ooffff you hate to see it

this inspired memes...

https://twitter.com/ThePrimeagen/status/1653154841668071424

then it started to happen first ember - https://twitter.com/EmberHext/status/1653096327675346945

it hurt, but it was true

then trav - https://twitter.com/techsavvytravvy/status/1653672117081108483

then others https://twitter.com/ESArnau/status/1653151823274942472

then the competition

Even stack overflow had their own bespoke solution... and it didn't even take advantage of tail recursion... rook

https://stackoverflow.com/questions/13859538/simplest-inline-method-to-left-pad-a-string/13861999

The Problem

So this all started due to something called microbenchmarks.

show what a microbenchmark is

Now the primary feature of a micro benchmark is that they lie to you

Run the benchmarks...

So is travvy's solution over 100x faster than the original leftpad? Was my live solution solution 50% slower than the original leftpad? Is my specialCase solution the SAME speed as the native implementation provided by padLeft?

But why do they lie to you?

they are often ran on "someone's computer." Its not a great testing environment

pgrep node pgrep rust-analyzer pgrep zls
- ladies and arch users, calm down *
not only that, but i am on a laptop... powermanagement?
the duration of the tests often hide implementation details, ie GC
usage makes a difference...

what a string usually is
rope https://en.wikipedia.org/wiki/Rope_(data_structure)
consequence of rope

The Experiment

So how did i determine what leftpad was ackshually the fastest?

create a server

go a head, complain about it in the comments

setup apache bench script and ensure a proper sleep between each and manual gc execution on the server if no requests (previous tests wont have as much initial effect)
send 50 million requests that took 36 hours to complete.

i would personally like to thank linode for not rate limiting me and giving me an instance to test on :) Thanks Andrew

create a parser to parse apache bench results and calculate the average of the medians
import results into google spreadsheets
charts

The Conclusion

What is each data point? Each data point represents 75k requests. The value is the average of the median value of 25k requests.

First, what do these charts mean? (go over axes) Second, which was the fastest? Third, What is our conclusion?