Scaling APIs with Performance Testing

Thank you

Austin HomeGrown API's

and Capital Factory

We skipped the beta,

deployed once,

everybody signed up,

and never had a problem.'

- No one ever

Promise: Scaling API's With Performance Testing

Scaling API's is about:

Optimizing Code
Architecture
Operations
Sensible Features

Performance is about:

Measuring performance (testing)
Identifying the slow bits (profiling)
Optimizing what matters (optimization)

Promise: Scaling API's With Performance Testing

Scaling API's is about:

Optimizing Code
Architecture
Operations
Sensible Features

Performance is about:

Measuring performance (testing) ← This talk!
Identifying the slow bits (profiling)
Optimizing what matters (optimization)

Who is Michael Cole?

Fullstack JS, DevOps, Freelancer, Digital Nomad.

bigSmall.io Micro-Consulting Platform

Micro-consulting platform
Video chat + invoices + teams
No job's too small if it's 100% billable
Sustainable Open Source
Helping people should be awesome!

Why Performance Testing?

Unit testing = feature confidence for changes

Performance testing = feature confidence for # of users

Performance testing came from Enterprise software (don't google "example test plan")

Agile Performance Testing means leaner, smaller, and iterative

What's your purpose?

Can I make this use case faster? In 2 weeks.
Can each server to serve 1000 requests/second, without crashing or errors? In 2 weeks.
We have a marketing event. App must serve 70,000 requests/second. In 2 weeks.

Measure

Change

Predict

In 2 weeks.

Perf Testing Menagerie

Load testing
How long does it take to process 10,000 requests?
Endurance/Soak testing
Measure resource leaks over sustained use.
How often will the server "randomly" crash?
Stress testing
Upper performance limit before failure.
Can it support 10,000 users simultaneously?
Spike testing
Sudden increased load and measures results.
Slashdot effect testing

Questions?

Today!

Build a small web app
Deploy it
Performance test it

Today!

Build a small web app
https://github.com/MichaelJCole/evil_kittens
Deploy it
Locally, then Heroku
Performance test it

Evil Kittens must REST!

Text

GET /tiny-kittens-in-your-memory/:id   -> text/plain: filename 

GET /big-kittens-in-your-memory/:id    -> image/jpeg: [kitten jpg data]

GET /evil-kittens-in-your-memory/:id   -> lock 1 mb of memory for 1 sec

GET /evil-kittens-in-your-cpu/:id      -> Calculate fib(30) badly

GET /evil-kittens-in-your-disk/:id     -> writes 1 mb to disk

GET /evil-kittens-in-your-network/:id  -> sends 512k of random data

Text

https://github.com/MichaelJCole/evil_kittens

A server to test

# If you'd like to play along on your laptop: 
# google `github evil_kittens`

# Get the code
git clone git@github.com:MichaelJCole/evil_kittens.git
cd evil_kittens

# Install dependencies
npm install

# Start server
node single.js

# Make requests
curl http://localhost:8000/evil-kittens-in-your-cpu/hello 
wget -qO- http://localhost:8000/evil-kittens-in-your-cpu/hello

Text

https://github.com/MichaelJCole/evil_kittens

A simple Load Test

# Install ApacheBench
sudo apt-get install apache2-utils

# MEASURE:  -n 1000 requests, using -c 10 threads, -l to relax
node single.js
ab -l -n 1000 -c 10 http://localhost:8000/evil-kittens-in-your-cpu/single


# Hmmmm... How could it be faster?  Lets turn off logging
node singleNoLog.js
ab -l -n 1000 -c 10 http://localhost:8000/evil-kittens-in-your-cpu/single


# Hmmmn... What if we used the cluster api?
node cluster.js
ab -l -n 1000 -c 10 http://localhost:8000/evil-kittens-in-your-cpu/single

# Yes!

A simple Endurance Test

# MEASURE 10000 requests with 1mb memory leak.
node single.js
ab -l -n 10000 -c 10 \
   http://localhost:8000/evil-kittens-in-your-memory/single

# MEASURE 10000 requests with 1mb memory leak.
node single.js
ab -l -n 10000 -c 10 \
   http://localhost:8000/evil-kittens-in-your-memory/single

My laptop has 16 gig of memory.

The "cloud" usually has about 0.5 -> 2 gig.

It's important to test on localhost and in a production-like environment - aka "stage".

If we use `ab` to test staging, all the requests come from localhost. We need some new tools.

A simple Endurance Test

Questions?

READER BEWARE:

"evil kittens" could be considered a denial-of-service attack, ESPECIALLY a shared hosting provider like Heroku.

Inappropriate use could cause others, and you, great suffering in downtime, account lock-outs, and/or legal aggression.

Production Like Performance Testing

aka. Don't come cryin to me!

READER BEWARE:

"evil kittens" could be considered a denial-of-service attack, ESPECIALLY a shared hosting provider like Heroku.

Inappropriate use could cause others, and you, great suffering in downtime, account lock-outs, and/or legal aggression.

A simple Load Test

(and for my final trick of the evening)

# Deploy your app.
git push heroku

# Install beeswithmachineguns
pip install git+git://github.com/newsapps/beeswithmachineguns.git

# make a ~/.boto file with your AWS keys.
# Details: https://github.com/newsapps/beeswithmachineguns

# I did this already because it is really slow!  
bees up -s 8 -g public -k evilkittens -z us-east-1e 

#  Wait till bees are no longer "Initailizing" in AWS console.
bees report

# Do the test
bees attack -n 10000 -c 40 \
  -u https://evilkittens.herokuapp.com/tiny-kittens-in-your-memory/bees

# forget this and you will get a big bill.
bees down

Questions?

Conclusion

Performance Testing isn't magic beans.
It's continually finding and fixing the slowest thing.
Test locally, then globally.
Remember: Cloud Computing = Cheap Hardware

Measure

Change

Predict

In 2 weeks.

Put a button on your site
Get paid to help people
New paying clients
Premium Support
Sustainable Open Source
Helping people should be awesome!

bigSmall.io Micro-Consulting Platform

Thanks for coming!

michael@bigSmall.io