Heroku Under the Hood workshop

Agenda 1. Performance tuning and dyno sizing 2. Buildpacks 3. app.json and the Heroku button 4. Platform API

Performance tuning

Performance basics Which WSGI server should I use? Recommendations: gunicorn, uWSGI Worth trying: waitress, twisted Not even once: runserver How many workers should I run? Start with WEB_CONCURRENCY Tune appropriately (more on this later) How should I serve static assets? Small scale: use Whitenoise Large scale, option A: S3 + CDN Large scale, option B: Whitenoise + CDN

Dyno types Type RAM Compute Cost Free / Hobby 512 MB 1-4x $0 / $7 Standard-1X 512 MB 1-4x $25 Standard-2X 1 GB 2x - 4x $50 Performance-M 2.5 GB 12x $250 Performance-L 14 GB 50x $500

Optimizing dyno types Assumption: your application is RAM-bound. Recommendation: optimize for cost per process. You’ll need to measure: RAM per process number of required processes Math: cost_per_process = (dyno_ram / process_ram) * dyno_cost H/T: https://medium.com/swlh/running-a-high-traffic-rails-app-on-heroku-s-performance-dynos-d9e6833d34c

Example 1 Dyno Processes / dyno Required # of dynos Total cost Cost / process 1X 4 5 $125 $6 2X 8 3 $150 $8 P-M 21 1 $250 $13 P-L 119 1 $500 $25 RAM per process: 120 Total processes: 20

Example 2 Dyno Processes / dyno Required # of dynos Total cost Cost / process 1X 1 60 $1,500 $25 2X 2 30 $1,500 $25 P-M 5 12 $3,000 $50 P-L 31 2 $1,000 $17 RAM per process: 450 Total processes: 60

Making measurements 1. Observe memory performance: log-runtime-metrics Heroku Dashboard Librato 2. Load test Siege 
 https://www.joedog.org/siege-home/ Bees with Machine Guns 
 https://github.com/newsapps/beeswithmachineguns Load testing addons 
 https://elements.heroku.com/addons#testing

Exercise 1: performance tuning A. What’s your RAM per process? B. What’s your number of required processes? C. What’s your optimal dyno formation?

Buildpacks

$ git push heroku master ... -----> Python app detected -----> Installing runtime (python-2.7.10) -----> Installing dependencies using pip ... What’s a buildpack?

How buildpacks work bin/detect — will this buildpack build this app? bin/compile — compile source into app bin/release — emit default config/addons

Python buildpack https://github.com/heroku/heroku-buildpack-python

Other interesting buildpacks Conda https://github.com/kennethreitz/conda-buildpack Null https://github.com/ryandotsmith/null-buildpack pgbouncer https://elements.heroku.com/buildpacks/heroku/heroku-buildpack-pgbouncer nginx https://elements.heroku.com/buildpacks/ryandotsmith/nginx-buildpack

Multiple buildpacks Use cases: nginx / pgbouncer asset collection via Gulp/Grunt/etc. …? $ heroku buildpacks === aqueous-shore-5792 Buildpack URL heroku/python $ heroku buildpacks:add --index 2 heroku/nodejs Buildpack added. Next release on aqueous-shore-5792 will use: 1. heroku/python 2. heroku/nodejs Run `git push heroku master` to create a new release using these buildpacks.

Exercise 2: try out multi-buildpack A. Try out nginx or pgbouncer. 
 Do they improve performance? B. Try a Node asset manager:
 https://github.com/beaugunderson/django-gulp

app.json, 
 Heroku buttons, 
 and Pipelines

$ cat app.json { "name": "Appointment Reminders (Django)", "description": "Appointment Reminders in Django with Twilio", "repository": "https://github.com/atbaker/appointment-reminders-django", "addons": [ "heroku-postgresql:hobby-dev", "redistogo:nano" ], "env": { "TWILIO_ACCOUNT_SID": { "description": "Your Twilio account secret ID", "value": "enter_your_account_sid_here", "required": true }, ... } } app.json

[](https://heroku.com/deploy? template=https://github.com/atbaker/appointment- reminders-django) Heroku button

Pipelines / PR apps Live demo (yipes)

Exercise 3: app.json / PR apps A. Create an app.json (and Heroku button?) for your app. B. Try out pull request apps.

Platform API https://devcenter.heroku.com/articles/platform-api-quickstart https://devcenter.heroku.com/articles/platform-api-reference

>>> import netrc, requests, json >>> token = netrc.netrc().hosts['api.heroku.com'][2] >>> h = requests.session() >>> h.headers['Authorization'] = 'Bearer ' + token >>> h.headers['Accept'] = 'application/vnd.heroku+json; version=3' >>> h.headers['Content-Type'] = 'application/json' >>> formation = h.get('https://api.heroku.com/apps/NAME/formation').json() >>> {f['type']: f['quantity'] for f in formation} {'web': 1, 'worker': 0} >>> payload = {'quantity': 4, 'size': 'standard-1X'} >>> h.patch('https://api.heroku.com/apps/NAME/formation/web', data=json.dumps(payload)) >>> formation = h.get('https://api.heroku.com/apps/NAME/formation').json() >>> {f['type']: f['quantity'] for f in formation} {'web': 4, 'worker': 0} Intro

Authentication Authentication uses bearer tokens: For local dev and/or fooling around, reading a token from .netrc is fine. Better: direct authentication with token exchange: Best: proper use of OAuth + scopes: 
 https://devcenter.heroku.com/articles/oauth >>> h.post('https://api.heroku.com/oauth/authorizations', ... data=json.dumps({'description': 'demo auth'})).json() {'access_token': {'expires_in': None, 'id': '4d6e04f0-693b-4118-bb9b-646e723ff7fa', 'token': ‘4a1637b4-9bc8-4ad9-8747-43ec7c744621'}, ...} >>> h.headers['Authorization'] = 'Bearer ' + token

Build & Slug APIs Deploy to Heroku without git push heroku master! To build without releasing, use a build app and copy slugs:
 https://devcenter.heroku.com/articles/platform-api-copying-slugs You can also create slugs from scratch: 
 https://devcenter.heroku.com/articles/platform-api-deploying-slugs >>> source = h.post('https://api.heroku.com/apps/aqueous-shore-5792/sources').json() >>> requests.put(source['source_blob']['put_url'], data=open('foo.tgz').read()) >>> payload = {'source_blob': {'url': source['source_blob']['get_url'], ... 'version': 'abcd1234'}} >>> build = h.post('https://api.heroku.com/apps/aqueous-shore-5792//builds', 
 data=json.dumps(payload)).json()

Exercise 4: platform API A. Build an auto-scaler. B. Build an alternate deploy flow. C. Re-create pull-request apps (!)