"Fast but not Loose" -- Typesafe client/server architecture at Gilt Groupe

Fast but Not Loose:Typesafe Clients in a

Distributed Service Architecture, a retrospective

Eric BowmanVP Architecture @ Gilt Groupe

@[email protected]

#gotoaar #gilttech#gotocon

Monday, September 30, 13

mailto:[email protected]

mailto:[email protected]

•Scala

•Play

•PostgreSQL

•MongoDB

•Voldemort

•Kafka

•Aster Data

•Mahout

•Jersey

•SBT

•Docker

•Continuous Delivery

Microservices


http://upload.wikimedia.org/wikipedia/commons/1/16/Ruby_on_Rails-logo.pnghttp://wiki.postgresql.org/wiki/File:PostgreSQL_logo.3colors.svgMonday, September 30, 13

http://upload.wikimedia.org/wikipedia/commons/1/16/Ruby_on_Rails-logo.png

http://upload.wikimedia.org/wikipedia/commons/1/16/Ruby_on_Rails-logo.png

http://wiki.postgresql.org/wiki/File:PostgreSQL_logo.3colors.svg

http://wiki.postgresql.org/wiki/File:PostgreSQL_logo.3colors.svg

http://logonoid.com/images/thumbs/christian-louboutin-logo.jpgMonday, September 30, 13

http://logonoid.com/christian-louboutin-logo/%5D%5Bimg%5Dhttp://logonoid.com/images/thumbs/christian-louboutin-logo.jpg%5B/img%5D%5B/url%5D

http://logonoid.com/christian-louboutin-logo/%5D%5Bimg%5Dhttp://logonoid.com/images/thumbs/christian-louboutin-logo.jpg%5B/img%5D%5B/url%5D

http://data.iluxdb.com/data/christian-louboutin-daffodile-160mm-python-masai-1130127cm09_001.jpg?dd80c0Monday, September 30, 13

http://data.iluxdb.com/data/christian-louboutin-daffodile-160mm-python-masai-1130127cm09_001.jpg?dd80c0

http://data.iluxdb.com/data/christian-louboutin-daffodile-160mm-python-masai-1130127cm09_001.jpg?dd80c0

http://upload.wikimedia.org/wikipedia/commons/7/79/Operation_Upshot-Knothole_-_Badger_001.jpg




http://blog.verwilst.be/wp-content/uploads/2008/12/java.gif




user_service order_service payment_service product_service cart_service

swift

pagegen


http://megmurph.com/wp-content/uploads/2013/03/success-1.jpg




Front End

Service X

Service Y

Service Z


Front End

Service X

Service Y

Service Z

Content from Service X

Content from Service Y

Content from Z


Front-End Tier

Service Tier

Data Tier


Front-End Tier

Service Tier

Data Tier

CachingLight Computation

Orchestration


Front-End Tier

Service Tier

Data Tier


Orchestration

CachingHeavier Computation

Separation of Concerns


Front-End Tier

Service Tier

Data Tier


Orchestration



Data Access(Disk/SSD/RAM)


•Runtime Temperature•Development Temperature

Front-End Tier

Service Tier

Data Tier


Orchestration



Data Access(Disk/SSD/RAM)


Data Source

Service

JDBC ➾ HashMap ➾ JSON

http://gilt.com

Javascript Application

Front End

JSON ➾ HashMap ➾ JSON


Chaos Grows Quickly

Data Source

Service

JDBC ➾ HashMap ➾ JSON

http://gilt.com


Front End



http://gilt.com


Front End




Service

JDBC ➾ HashMap ➾ JSONData Source

Service


Service



http://gilt.com


Front End




Service


Service


Service


Service




http://www.recruitmenttakeout.com/wp-content/uploads/2013/04/brian.png




•Fast

•Service Decomposition

•Implicit core model was good


•Org Scaling

•APIs

•Implicit core model was ... implicit


A data model and APIs for services


A data model and APIs for services

(aka, RPC)


http://git-scm.com/images/logos/downloads/Git-Logo-2Color.png

core service

core service

core service

core service

Legacy





core service

core service

core service

core service

Legacy

commons.jar

core clients

core data model

async client framework

Greenfield





core service

core service

core service

core service

Legacy

commons.jar

core clients

core data model

async client framework

Greenfield

•Users•Sales•Products•Skus•Assets•Targeting•Auth




service

client

core


service

client

core

•RESTful•Scala clients•All APIs futures-based•Case class schema


service

client

core

client

core

service

Compile/RuntimeDependency


ConsumerServiceClient

Core Core

“Embassy Soil”



Core Core

•Easy functional testing•Service response capture•Test linking•Upgradable•Emergent Regression•Automated upgrades•Compile farmers



Core Core


•Environment ⊕ Config•Live updates•Indirection•Circuit Breaker


Core Core


“All of this is completely wrong.”*

*Not an actual quoteMonday, September 30, 13

1.The network is reliable2.Latency is zero3.Bandwidth is infinite4.The network is secure5.Topology doesn't change6.There is one administrator7.Transport cost is zero8.The network is homogeneous

https://blogs.oracle.com/jag/resource/Fallacies.html




1.The network is reliable





1.The network is reliable


Resolved




2.Latency is zero





2.Latency is zero


Resolved




3.Bandwidth is infinite





3.Bandwidth is infinite


Resolved




4.The network is secure





4.The network is secure


Resolved




5.Topology doesn't change





5.Topology doesn't change


Resolved




6.There is one administrator





6.There is one administrator


Resolved




7.Transport cost is zero





7.Transport cost is zero


Resolved




8.The network is homogeneous





8.The network is homogeneous


Resolved




“Still Wrong.”*

*Possibly an actual quoteMonday, September 30, 13

Convenience Over Correctness

http://steve.vinoski.net/blog/2008/07/01/convenience-over-correctness/Monday, September 30, 13

http://steve.vinoski.net/blog/2008/07/01/convenience-over-correctness/

http://steve.vinoski.net/blog/2008/07/01/convenience-over-correctness/

Database

Database Model

Service Backend

Service Model

Service Frontend

Client/Server Serialization Model

Client Backend

Client Data Model

Client Layer

Public Data Model


Database

Database Model

Service Backend

Service Model

Service Frontend


Client Backend

Client Data Model

Client Layer

Public Data Model

•So many models•Corners are cut•Typesafe helps•Conflation?•Just the data


“Works in Practice for some use cases”

Database

Database Model

Service Backend

Service Model

Service Frontend


Client Backend

Client Data Model

Client Layer

Public Data Model



“Works in Practice for some use cases”“No free silver bullet lunches.”

Database

Database Model

Service Backend

Service Model

Service Frontend


Client Backend

Client Data Model

Client Layer

Public Data Model



However...


•No machine generated stubs•Embassy-Oriented Programming•Lots of indirection•Type-system support for failures

However...


What Actually Sucks about RPC:


What Actually Sucks about RPC:•Remote objects •Failures•Idempotency


Ignoring all that was Too Easy.


Batch Jobs


1.The network is reliable2.Latency is zero3.Bandwidth is infinite4.The network is secure5.Topology doesn't change6.There is one administrator7.Transport cost is zero8.The network is homogeneous


Law of Instrument


Pulling data.


Pushing code.


SOA, Reloaded


SOA, Reloaded•CRUD•Event Streams•Batch Processing•Lambda Architecture•CQRS


Client

Service

DB


Client

Fast Storage

Service

Kafka

Data Warehouse

Batch Jobs

DB

CEP

Clickstream data


http://tech.gilt.com

join us.new york & dublin




"Fast but not Loose" -- Typesafe client/server architecture at Gilt Groupe

Technology

Transcript of "Fast but not Loose" -- Typesafe client/server architecture at Gilt Groupe