Principles and Practices in Continuous Deployment at Etsy

Post on 23-Aug-2014

10190 Views

Category:

Engineering

1 Downloads

Preview:

Click to see full reader

DESCRIPTION

Presented at ALM Forum 2014. Like what you've read? We're frequently hiring for a variety of engineering roles at Etsy. If you're interested, drop me a line or send me your resume: mike@etsy.com. http://www.etsy.com/careers

Transcript

Principles and Practices in Continuous Deployment

Mike Brittain

Engineering Director, Etsy

@mikebrittain mikebrittain.com/talks

“Continuous Deployment”Process by which our team deploys software changes to production services over 30 times per day.

Where we started

Principles for our Engineering team

Continuous Deployment

Business case

Five years ago…2-3 weeks of code changes

Release and rollback plans

Traffic and infrastructure management (Ops)

6-14 hours

Five years ago…“Deployment Army”

Stressful, especially when things go wrong

Long days and late nights

Scheduled downtime

pro·duc·tion [pruh-duhk-shuhn] (n)

1. This complex system of application code, distributed services, servers, networking gear, etc., upon which we’re going to try to carefully apply a complicated set of changes and hope that nothing goes wrong. Cross your fingers… here goes.

Software for large-scale web sites has been traditionally written by one group of people, then released and operated by a different group.

These two groups have very different levels of visibility into how the software works.

Stagnation

“…frequent and prolonged outages.”2010 CAPACITY PLAN

First, Principles.

Innovate or die

Innovate or die

Resolve scaling hurdles

Innovate or die

Resolve scaling hurdles

Mean-time-to-recovery

“Quality is not just testing pre-release. It also includes our adaptability and response time.”

- Jeff Sussna at ALM Forum, 2014

Innovate or die

Resolve scaling hurdles

Mean-time-to-recovery

Innovate or die

Resolve scaling hurdles

Mean-time-to-recovery

Healthy and talented engineering team

Autonomy, Mastery, Purpose.

“Drive: The surprising truth about what motivates us.” ~Dan Pink, at RSA http://youtu.be/u6XAPnuFjJc

Innovate or die

Resolve scaling hurdles

Mean-time-to-recovery

Healthy and talented engineering team

Stop stressing about releases

First, Principles.

http://timothyfitz.com/2009/02/10/continuous-deployment-at-imvu-doing-the-impossible-fifty-times-a-day/

In a software release process Fail Fast means releasing undeployed code as fast as possible, instead of waiting for a weekly release to break.

http://youtu.be/LdOe18KhtT4

SECRET WEAPON: Hired as VP, Tech-Ops at Etsy

Continuous Deployment

Continuous Delivery~ vs ~

Frequent check-ins directly to mainline.

Continuous Deployment Continuous Delivery✓ ✓

Continuous Integration and Automated tests.

Continuous Deployment Continuous Delivery✓ ✓

Keep the build green. We’re always ready to release.

Continuous Deployment Continuous Delivery✓ ✓

“One button” deploys.

Continuous Deployment Continuous Delivery✓ ✓

Business dictates when a build is deployed.

Continuous Deployment Continuous Delivery✓

Every passing build is deployed to production.

Continuous Deployment Continuous Delivery✓

All enhancements are gated by Config Flags. (“Branch in code”)

Continuous Deployment Continuous Delivery✓ ?

Most of the builds we deploy are “dark” changes.

CSS rules and properties

Copy in templates (e.g. typos)

New, un-referenced code (e.g. classes, funcs, templates)

Code paths behind disabled config flags

etc…

Dev Team Version Control Build & Unit Tests

Automated Acceptance Tests

User Acceptance Tests Release

Check inTrigger

Feedback

Source: http://en.wikipedia.org/wiki/Continuous_delivery

Continuous Delivery release pipeline

Dev Team Version Control Build & Unit Tests

Automated Acceptance Tests

User Acceptance Tests Release

Check inTrigger

Feedback

Check inTrigger

Feedback Trigger

Feedback

Source: http://en.wikipedia.org/wiki/Continuous_delivery

Continuous Delivery release pipeline

Dev Team Version Control Build & Unit Tests

Automated Acceptance Tests

User Acceptance Tests Release

Check inTrigger

Feedback

Check inTrigger

Feedback Trigger

Feedback

Check inTrigger

Feedback Trigger

Feedback Approval

ApprovalFeedback

Source: http://en.wikipedia.org/wiki/Continuous_delivery

Continuous Delivery release pipeline

Dev Team Version Control Build & Unit Tests

Automated Acceptance Tests

User Acceptance Tests Release

Check inTrigger

Feedback

Check inTrigger

Feedback Trigger

Feedback

Check inTrigger

Feedback Trigger

Feedback Approval

ApprovalFeedback

Continuous Delivery release pipeline

Dev / Integration Staging Production

Dev Team Version Control Build & Unit Tests

Automated Acceptance Tests

User Acceptance Tests Release

Check inTrigger

Feedback

Check inTrigger

Feedback Trigger

Feedback

Check inTrigger

Feedback Trigger

Feedback Approval

ApprovalFeedback

Continuous Delivery release pipeline

Dev / Integration Staging Production

Assumptions:

Staging is a perfect reflection of Production, with respect to hardware, configurations, data, overall load, capacity, etc.

Deploy process is infallible.

“What do you mean, ‘it’s not working in production?’ I TESTED IT BEFORE WE RELEASED!”

Dev Team Version Control Build & Unit Tests

Automated Acceptance Tests

User Acceptance Tests Release

Check inTrigger

TriggerApproval

Continuous Delivery release pipeline

Dev / Integration Staging Production

Check inTrigger

TriggerApproval

Check inTrigger

TriggerApproval

Check inTrigger

TriggerApproval

Check inTrigger

TriggerApproval

Check inTrigger

TriggerApproval

Dev Team Version Control Build & Unit Tests

Automated Acceptance Tests

User Acceptance Tests Release

Check inTrigger

TriggerApproval

Continuous Delivery release pipeline

Dev / Integration Staging Production

Check inTrigger

TriggerApproval

Check inTrigger

TriggerApproval

Check inTrigger

TriggerApproval

Check inTrigger

TriggerApproval

Approval

Check inTrigger

TriggerApproval

Feedback

Dev Team Version Control Build & Unit Tests

Automated Acceptance Tests

User Acceptance Tests Release

Check inTrigger

TriggerApproval

Continuous Delivery release pipeline

Check inTrigger

TriggerApproval

Check inTrigger

TriggerApproval

Check inTrigger

TriggerApproval

Check inTrigger

TriggerApproval

Approval

Check inTrigger

TriggerApproval

Feedback

"Because you’re integrating so frequently, there is significantly less back-tracking to discover where things went wrong , so you can spend more time building features.” !

—ThoughtWorks !

!http://www.thoughtworks.com/continuous-integration

Dev / Integration Staging Production

Dev Team Version Control Build & Unit Tests

Automated Acceptance Tests

User Acceptance Tests Release

Check inTrigger

TriggerApproval

Continuous Delivery release pipeline

Check inTrigger

TriggerApproval

Check inTrigger

TriggerApproval

Check inTrigger

TriggerApproval

Check inTrigger

TriggerApproval

Approval

Check inTrigger

TriggerApproval

Feedback

Dev / Integration Staging Production

Where’s the bug? !

In one of the numerous check-ins? Missing unit tests? Missing automated UA tests? Missing manual UA tests?

Dev Team Version Control Build & Unit Tests

Automated Acceptance Tests

User Acceptance Tests Release

Check inTrigger

TriggerApproval

Continuous Delivery release pipeline

Check inTrigger

TriggerApproval

Check inTrigger

TriggerApproval

Check inTrigger

TriggerApproval

Check inTrigger

TriggerApproval

Approval

Check inTrigger

TriggerApproval

Feedback

Dev / Integration Staging Production

Where’s the bug? !

In one of the numerous check-ins? Missing unit tests? Missing automated UA tests? Missing manual UA tests? !Data out of sync?

Server configurations out of sync?

Capacity vs. current load?

Deployment script?

Dev Team Version Control Build & Unit Tests

Automated Acceptance Tests

User Acceptance Tests Release

Check inTrigger

TriggerApproval

Continuous Delivery release pipeline

Check inTrigger

TriggerApproval

Check inTrigger

TriggerApproval

Check inTrigger

TriggerApproval

Check inTrigger

TriggerApproval

Approval

Check inTrigger

TriggerApproval

Feedback

Dev / Integration Staging Production

How will we know when something is wrong in production? !

How long will it take to resolve the issue?

Check inTrigger

We aim to reduce fundamental surprise in every release.

Furthermore, we optimize for detecting and recovering from failures quickly.

Pre-production validationCode deployed to de-pooled application (web) servers touching prod services and databases.

Smoke tests

Integration tests

Functional tests

User-Acceptance (ad hoc)

Production validationExactly the same server configs, services and data as pre-prod, but this is where we introduce application code to live traffic.

Production validationExactly the same server configs, services and data as pre-prod, but this is where we introduce application code to live traffic.

Smoke tests (esp. over public hostnames)

User-Acceptance testing behind config flags

Gratuitous monitoring

Customer support and forums

Single release

Many releases

50K LOC/month

Few opportunities for failureWide surface area (50,000 LOC) High MTTR !

All of the bugs we’ve written

More opportunities for failure Narrow surface area (< 100 LOC)

Low MTTR !

A fraction of the bugs we’vewritten per release

Imagine that we’ll write

Monitoring

Monitoring

PHP Warnings Bug Reports and Help Requests

Deploy logs

Post-Mortems

Check inTrigger

Feedback Trigger

Feedback Approval

Approval

Smoke Tests User Acceptance!Tests ReleaseDev Team Version Control Build & Unit

TestsAutomated

Acceptance TestsUser Acceptance

Tests Deploy (Prod) Monitoring and Automated Alerts

Continuous Deployment release pipeline

Feedback

Dev Pre-Production (“Princess”)

Production

Check inTrigger

Feedback Trigger

Feedback Approval

ApprovalFeedback

Smoke Tests User Acceptance!Tests ReleaseDev Team Version Control Build & Unit

TestsAutomated

Acceptance TestsUser Acceptance

Tests Deploy (Prod) Monitoring and Automated Alerts

Continuous Deployment release pipeline

CI

Dev Pre-Production (“Princess”)

Production

Check inTrigger

Feedback Trigger

Feedback Approval

ApprovalFeedback

Smoke Tests User Acceptance!Tests ReleaseDev Team Version Control Build & Unit

TestsAutomated

Acceptance TestsUser Acceptance

Tests Deploy (Prod) Monitoring and Automated Alerts

Continuous Deployment release pipeline

CI

ApprovalApproval

FeedbackFeedback

Feedback

Dev Pre-Production (“Princess”)

Production

Check inTrigger

Feedback Trigger

Feedback Approval

ApprovalFeedback

Smoke Tests User Acceptance!Tests ReleaseDev Team Version Control Build & Unit

TestsAutomated

Acceptance TestsUser Acceptance

Tests Deploy (Prod) Monitoring and Automated Alerts

Continuous Deployment release pipeline

CI

ApprovalApproval

FeedbackFeedback

Feedback ApprovalFeedback Approval

Feedback

“Allow buttons properly to inherit color from their parent node.”

Five years ago…2-3 weeks of code changes

Release and rollback plans

Traffic and infrastructure management (Ops)

6-14 hours

Five years ago…“Deployment Army”

Stressful, especially when things go wrong

Long days and late nights

Scheduled downtime

Why do we do this?

Innovate or die.

Resolve scaling hurdles.

Mean-time-to-recovery.

Healthy and talented engineering team.

Stop stressing about releases.

Innovate or die.

Resolve scaling hurdles.

Mean-time-to-recovery.

Healthy and talented engineering team.

Stop stressing about releases.

Admin-launch and whitelist

Ramp-up public traffic

mainline

header_redesign

search_filter_custom_orders

checkout_blue_button

listing_css_refactor

www.etsy.com

beta01.etsy.com

beta02.etsy.com

beta03.etsy.com

www.etsy.com

beta01.etsy.com

beta02.etsy.com

beta03.etsy.com

date":"30\/Mar\/2014:12:49:48","locale_currency_code":"USD","pref_language":"en-

US","region":"US","detected_currency_code":"USD","detected_language":"en-

US","detected_region":"US","accept-languages":"en-US","cdn-

provider":"","isMobileDevice":"0","isMobileSupported":"0","isMobileRequestIgnoreCookie":"0"

,"isTabletSupported":"0","isTouch":"0","isEtsyApp":"0","isPreviewRequest":"0","isChromeInst

antRequest":"0","isMozPrefetchRequest":"0","listing_ids":

[104073511,130604774,159651433,155451607,160523743,124025232,95186610,82967340,114692884,11

4767467,117266897,157579748],"scheduled_modules_content_ids":

[10808052776,10256029946],"primary_event":"1",".event_source":"web",".event_logger":"fronte

nd","php_ab_test_names":"translation_profiler.profiling;translation_profiler.logging;transl

ation_profiler.backend_event_logging;footer_redesign_20131201;international.languages.el;in

ternational.languages.ja;international.languages.no;international.languages.pl;internationa

l.languages.ro;international.languages.tr;simplified_locale_experience;full_site_ssl;admin_

toolbar;enabled_locale_subdirectories;affiliates.publishing.user_publishers;buyer_invites_r

ecipients;home_improvement;home_improvement.new_homepage;authoritative_items;refactored_foo

ter;conversations.rejuvination;contextual_homepage_recs.global;css_from_www;shrinkray.css;c

srf_nonce_refactor.allow_colon;csrf_nonce_refactor.reverse_order;csrf_nonce_refactor.no_encAnalytics connected to config names

Catapult

Observed impact

Time series data for duration of the experiment

Observed impact

Time series data for duration of the experiment

FrankProduct Manager

“I want to find out whether buyers will favor a single price for the product that includes shipping.”

https://www.etsy.com/shop/lucra

Eligibility requirements:

- Must be first page of visit

- Buyer & seller in same region

- etc…

Time: < 8 hours Staff: One !

Design, config flag (disabled), eligibility code in controller, template code, CSS, code review, automated tests, deployed code, config flag enabled.

We do not bundle the item price and shipping cost together today. !

https://www.etsy.com/shop/lucra

Ambitious Product Goal

Ambitious Product Goal

Monolithic Building and measuring many things at once.

Ambitious Product Goal

Monolithic Building and measuring many things at once.

Iterative One thing at a time, our design goal is always in sight.

Time: < 8 hours Staff: One !

Design, config flag (disabled), eligibility code in controller, template code, CSS, code review, automated tests, deployed code, config flag enabled.

Deployed Deployed

Is this for me?

http://timothyfitz.com/2009/02/10/continuous-deployment-at-imvu-doing-the-impossible-fifty-times-a-day/

“Maybe this is just viable for a single developer … your site will be down. A lot.”

etsystatus.com

@mikebrittain

Very end of 2009 Today

DEPLOYMENTS PER DAYAPP CODE CONFIG FILES

$1.35 Billion Goods sold in 2013 60+ Million Unique visitors per month !

175+ Committers, everyone deploys

http://www.etsy.com/blog/news/2013/etsy-statistics-december-2012-weather-report/Items by anjaysdesigns, betwixxt, OneStarLeatherGoods, mediumcontrol, TheDesignPallet

Thank you.

Mike Brittain

Engineering Director, Etsy

@mikebrittain mikebrittain.com/talks

top related