Scalable Deployments How we deploy Rails app to 150+ hosts in a minute

AD ✮鼅 㕼 傈 ˋؔٝٓ؎ٝ ו׍׵ַ傈 劤䨌 㕼 !-*/&吳䒭⠓爡幪靼ؼؕٔؒؔؿ؍أ ! ˑֶ겗הז׷8FC؟٦ؽأ׾寸׭׵׸׋ؘُٖٖ٦ءّٝך⚥ד ꣲ歲תד넝鸞⻉׾㔳׷ثُ٦صؚٝغزٕ˒ ! ⚺⪵-*/&吳䒭⠓爡 ㉏겗⡲䧭涯ꆃ⹛暟㕦 $PPLQBE http://isucon.net/

OK

Scalable Deployments How we deploy Rails app to 150+ hosts in a minute

TODAY I TALK ABOUT How Cookpad performs deployments

TOPICS NOT INCLUDED Rails Continuous Delivery App servers’ auto scaling

LINKS Cookpad's deployment and auto scaling Continuous Delivery in Cookpad https://speakerdeck.com/mirakui/cookpads-deployment-and-auto-scaling https://speakerdeck.com/takai/continuous-delivery-in-cookpad

Cookpad Inc. Dev-Infra group (Ꟛ涪㛇湍) Ruby committer ׉׵כ TPSBIFS Shota Fukumori @sora_h ! sorah

DEPLOYMENTS SO,

quoted from New Oxford American Dictionary 3rd edition © 2010, 2012 by Oxford University Press photo: https://www.flickr.com/photos/thenationalguard/4401592829 deploy |diˈploi| verb [ with obj. ] ! move (troops) into position for military

Software deployment is all of the activities that make a software system available for use. ! The general deployment process consists of several interrelated activities with possible transitions between them. These activities can occur at the producer side or at the consumer side or both. Because every software system is unique, the precise processes or procedures http://en.wikipedia.org/wiki/Software_deployment

DEPLOYMENTS

We perform deployment of the following Rails app: +----------------------+-------+-------+---------+---------+-----+-------+ | Name | Lines | LOC | Classes | Methods | M/C | LOC/M | +----------------------+-------+-------+---------+---------+-----+-------+ | Controllers | 41964 | 33824 | 436 | 3397 | 7 | 7 | | Helpers | 13296 | 10950 | 10 | 1289 | 128 | 6 | | Models | 87626 | 69239 | 1530 | 7604 | 4 | 7 | | Mailers | 300 | 240 | 11 | 26 | 2 | 7 | | Javascripts | 38740 | 33240 | 34 | 4789 | 140 | 4 | | Libraries | 56189 | 46375 | 532 | 4371 | 8 | 8 | | Async_view specs | 247 | 212 | 0 | 0 | 0 | 0 | | Controller specs | 55098 | 45557 | 7 | 117 | 16 | 387 | | Feature specs | 36807 | 30226 | 0 | 165 | 0 | 181 | | Helper specs | 3598 | 2956 | 0 | 7 | 0 | 420 | | Lib specs | 21636 | 18095 | 27 | 124 | 4 | 143 | | Mailer specs | 306 | 251 | 0 | 0 | 0 | 0 | | Policy specs | 1594 | 1302 | 0 | 0 | 0 | 0 | | Request specs | 28698 | 24526 | 0 | 11 | 0 | 2227 | | Routing specs | 648 | 523 | 0 | 0 | 0 | 0 | | View specs | 619 | 508 | 0 | 2 | 0 | 252 | | Worker specs | 862 | 715 | 0 | 1 | 0 | 713 | +----------------------+-------+-------+---------+---------+-----+-------+ | Total | 388228 | 318739 | 2587 | 21903 | 8 | 12 | +----------------------+-------+-------+---------+---------+-----+-------+

+----------------------+-------+-------+---------+ | Name | Lines | LOC | Classes | +----------------------+-------+-------+---------+ | Controllers | 41964 | 33824 | 436 | | Helpers | 13296 | 10950 | 10 | | Models | 87626 | 69239 | 1530 | | Mailers | 300 | 240 | 11 | | Javascripts | 38740 | 33240 | 34 | | Libraries | 56189 | 46375 | 532 | | Controller specs | 55098 | 45557 | 7 | | Feature specs | 36807 | 30226 | 0 | | Helper specs | 3598 | 2956 | 0 | | Lib specs | 21636 | 18095 | 27 |

| Controller specs | 55098 | 45557 | 7 | | Feature specs | 36807 | 30226 | 0 | | Helper specs | 3598 | 2956 | 0 | | Lib specs | 21636 | 18095 | 27 | | Mailer specs | 306 | 251 | 0 | | Policy specs | 1594 | 1302 | 0 | | Request specs | 28698 | 24526 | 0 | | Routing specs | 648 | 523 | 0 | | View specs | 619 | 508 | 0 | | Worker specs | 862 | 715 | 0 | +----------------------+-------+-------+---------+ | Total | 388228 | 318739 | 2587 +----------------------+-------+-------+---------+

140 servers 10 times / day ! " (peak)

RULES ON DEPLOYMENT Deploy revisions which CI build passes Only during working time After deployment, monitor errors for an hour Rollback if error rate increase, or any trouble

Our deployment was: CI git repo developer # $ " % pass tag check deploy & merge ' build

Our deployment was: App App App : Deploy Server ssh+rsync capistrano 2 ( )

Deploy via Chat # $ ( check " deploy!

How long time spent for deployment? CI git repo developer # $ " % & ' 10 min 1..5 min 10 min

How long time spent for deployment? 15…20 min

PROBLEMS

PROBLEMS Capistrano 2 with complicated super historical deploy script ! Seemed time to re-new

$ tree config config ├── cutty_deploy.rb ├── deploy │ ├── ***.rb │ ├── production.rb │ ├── production_test.rb │ ├── production_***.rb │ ├── production_***_test.rb │ ├── (snip) │ ├── rails41.rb │ ├── ruby210.rb │ ├── staging.rb │ ├── staging_***.rb │ └── ***_test.rb ├── deploy_support │ ├── bundler_capistrano.rb │ ├── chat_notification.rb │ ├── deploy_utils.rb │ └── rsync_with_remote_cache.rb :

$ wc -l config/cutty_deploy.rb \ config/deploy/* \ config/deploy_support/* ! 2595 total

PROBLEMS SSH is slow High CPU usage on deployment Sometime Fails

PROBLEM: Sometime Fails

How long time spent for deployment? CI git repo developer # $ " % & ' 10 min 1..5 min 10 min

How long time spent for deployment? CI git repo developer "

How long time spent for deployment? CI git repo developer x 3 min " retry

How long time spent for deployment? CI git repo developer x 3 min " retry 3 min … x

How long time spent for deployment? CI git repo developer # " % ' 10 min 1..5 min 10 min ? $ &

How long time spent for deployment? CI git repo developer # $ " % & ' 10 min 1..5 min 10..20 min+

My team “dev-infra” aims to: Improve developers’ productivity ! Keep development fast Maintain & improve test environment etc

WE HAVE TO IMPROVE of course,

IMPROVEMENT PLANS Upgrade to Capistrano 3? ! It has better SSH handling, but still depends on SSH. SSH is slow.

CREATE, NEW TOOL! Create new tool that uses another way for deployments!

INTRODUCING * sorah/mamiya (pronounce like mar-me-ya)

MAMIYA use Serf for orchestration use Amazon S3 for file distribution (by default) compatible directory structure with Capistrano

SERF? * hashicorp/serf Orchestration tool ! Decentralized, fault-tolerant, highly available Uses Gossip protocol (SWIM)

GOSSIP PROTOCOL? A gossip protocol is a style of computer-to- computer communication protocol inspired by the form of gossip seen in social networks. http://en.wikipedia.org/wiki/Gossip_protocol

GOSSIP PROTOCOL: interval: 200ms, total nodes: 8, fanout: 2 = event node node node node node node node node e

node node node node node GOSSIP PROTOCOL: node node e Receives Event. node 0ms 200ms 400ms 600ms 800ms +

node node node node node GOSSIP PROTOCOL: node node e Receives Event. Choose nodes to gossip node 0ms 200ms 400ms 600ms 800ms +

node node node node node GOSSIP PROTOCOL: node node e Receives Event. Choose nodes to gossip node 0ms 200ms 400ms 600ms 800ms +

node node node node node GOSSIP PROTOCOL: node node e Receives Event. Choose nodes to gossip Fan out. node 0ms 200ms 400ms 600ms 800ms +

node node node node node GOSSIP PROTOCOL: node node e e e Receives Event. Choose nodes to gossip Fan out. node 0ms 200ms 400ms 600ms 800ms +

node node node node node GOSSIP PROTOCOL: node node e e e Receives Event. Choose nodes to gossip Fan out. node 0ms 200ms 400ms 600ms 800ms +

node node node node node GOSSIP PROTOCOL: node node e e e Receives Event. Choose nodes to gossip Fan out. node 0ms 200ms 400ms 600ms 800ms +

node node node node node GOSSIP PROTOCOL: node node e e e Receives Event. Choose nodes to gossip Fan out. node 0ms 200ms 400ms 600ms 800ms +

node node node node node GOSSIP PROTOCOL: node node e e e e e e e Receives Event. Choose nodes to gossip Fan out. node 0ms 200ms 400ms 600ms 800ms +

node node node node node GOSSIP PROTOCOL: node node e e e e e e e Receives Event. Choose nodes to gossip Fan out. node 0ms 200ms 400ms 600ms 800ms +

node node node node node GOSSIP PROTOCOL: node node e e e e e e e Receives Event. Choose nodes to gossip Fan out. node 0ms 200ms 400ms 600ms 800ms +

node node node node node GOSSIP PROTOCOL: node node e e e e e e e Receives Event. Choose nodes to gossip Fan out. node 0ms 200ms 400ms 600ms 800ms +

node node node node node GOSSIP PROTOCOL: node node e e e e e e e Receives Event. Choose nodes to gossip Fan out. node e e e e e e e e 0ms 200ms 400ms 600ms 800ms +

node node node node node GOSSIP PROTOCOL: node node e e e e e e e Receives Event. Choose nodes to gossip Fan out. node e e e e e e e e Drop received event if it’s known 0ms 200ms 400ms 600ms 800ms +

node node node node node GOSSIP PROTOCOL: node node e e e e e e e Receives Event. Choose nodes to gossip Fan out. node e Drop received event if it’s known 0ms 200ms 400ms 600ms 800ms +

node node node node node GOSSIP PROTOCOL: node node e e e e e e e Receives Event. Choose nodes to gossip Fan out. node e Drop received event if it’s known 0ms 200ms 400ms 600ms 800ms +

node node node node node GOSSIP PROTOCOL: node node e e e e e e e Receives Event. Choose nodes to gossip Fan out. node e Drop received event if it’s known 0ms 200ms 400ms 600ms 800ms +

node node node node node GOSSIP PROTOCOL: node node e e e e e e e Receives Event. Choose nodes to gossip Fan out. node e Drop received event if it’s known 0ms 200ms 400ms 600ms 800ms +

node node node node node GOSSIP PROTOCOL: node node e e e e e e e Receives Event. Choose nodes to gossip Fan out. node e Drop received event if it’s known e e 0ms 200ms 400ms 600ms 800ms +

node node node node node GOSSIP PROTOCOL: node node e e e e e e e Receives Event. Choose nodes to gossip Fan out. node e Drop received event if it’s known e e 0ms 200ms 400ms 600ms 800ms +

node node node node node GOSSIP PROTOCOL: node node e e e e e e e Receives Event. Choose nodes to gossip Fan out. node e Drop received event if it’s known e e 0ms 200ms 400ms 600ms 800ms +

node node node node node GOSSIP PROTOCOL: node node e e e e e e e Receives Event. Choose nodes to gossip Fan out. node e Drop received event if it’s known e e 0ms 200ms 400ms 600ms 800ms +

node node node node node GOSSIP PROTOCOL: node node e e e e e e e Receives Event. Choose nodes to gossip Fan out. node e Drop received event if it’s known 0ms 200ms 400ms 600ms 800ms +

node node node node node GOSSIP PROTOCOL: node node e e e e e e e Receives Event. Choose nodes to gossip Fan out. node e Drop received event if it’s known 0ms 200ms 400ms 600ms 800ms +

node node node node node GOSSIP PROTOCOL: node node e e e e e e e Receives Event. Choose nodes to gossip Fan out. node e Drop received event if it’s known 0ms 200ms 400ms 600ms 800ms + Now all nodes has the event.

SERF Consumes UDP bandwidth proportional to cluster size, but better than SSH. * hashicorp/serf

VILLEIN Side-effect gem for Mamiya Simple gem to control `serf` from Ruby * sorah/villein

HOW IT WORKS

Terminologies & Concept Master node Agent node Package Storage Steps

Terminologies & Concept Master node has HTTP API to control cluster sends requests to agents (via serf) watches agents’ status

Terminologies & Concept Agent node accepts requests from master node runs deploy tasks

Terminologies & Concept Deploy Script writes how to “build”, “prepare”, and “release”

Terminologies & Concept Package is a tarball of files to deploy can be pushed to storage contains Deploy Script

Terminologies & Concept Storage can store Packages used from Agent nodes

Terminologies & Concept Step is part of deployment can be run separately called remotely

Steps Fetch package from storage Prepare fetched package (bundle install, etc) Switch to prepared package (reload, graceful)

1. CI builds package when passed 2. CI pushes the package to storage 3. — Deployment starts — 4. Master sends “prepare” request to Agents 5. Agents fetch package, then prepare Mamiya’s Deploy flow

6. Master confirms all agents have prepared 7. Master sends “switch” request to Agents 8. Agent switches symlinks, then reload app process Mamiya’s Deploy flow

mamiya’s deploy flow storage CI developer , build+push check deploy pass prepare app merge " & # - % reload .

Result Removed dependency to slow SSH ! but, more…?

Terminologies & Concept Step is part of deployment can be run separately called remotely

ANOTHER GOAL Do preparation before developer say “DEPLOY!”

mamiya’s deploy flow (prepare earlier) storage CI developer , build+push check deploy pass prepare app merge " & # . - % reload app

Result (cap→mamiya) Before: 8.4 minutes After: 45 seconds ! for 110 servers, 11.2x faster!

DEMO

Future Plans Better documentation (soon) Auto-deploy when joining cluster Web UI Better error tracking, handling Incremental Packages “master is always deployed”

THANKS * sorah/mamiya questions? @sora_h