Aviran Mordo Head of linkedin.com/in/aviran aviransplace.com Scaling with Microservices Architecture and Multi-cloud platforms

Wix in Numbers Over 72M users (website builders) Static storage is >2PB of data 3 data centers + 3 clouds (Google, Amazon, Azure) 2B HTTP requests/day 1000 people work at Wix

Initial Architecture Built for fast development Stateful login (Tomcat session), Ehcache, file uploads No consideration for performance, scalability and testing Intended for short-term use Tomcat, Hibernate, custom web framework Lighttpd (file serving) MySQL DB Wix (Tomcat)

The Monolithic Giant One monolithic server that handled everything Dependency between features Changes in unrelated areas of the system caused deployment of the whole system Failure in unrelated areas will cause system wide downtime

Breaking the System Apart

Concerns and SLA Data Validation Security / Authentication Data consistency Lots of data Edit websites High availability High performance Lots of static files Very high traffic volume Viewport optimization Long tail (immutable) Serving Media High availability High performance High traffic volume Long tail (mutable) View sites, created by Wix editor

Wix Segmentation 1. Editor Segment3. Public Segment 2. Media Segment Networking

HTML Editor Flash Editor MSM Private Media Public Media Editor Segment Public Segment Premium Services eCommerse List DB App Builder App Store App Market Dashboard Statics/me dia Mailer TimeZone Public HTML API Public API (Flash) MSP Public Server HTML Renderer HTML SEO Renderer Flash Renderer Flash SEO Renderer Sitemap Renderer Robots.txt Renderer User Server Template Viewer Contacts HUB Activit y Site Members Provided Mailing Service Comments Snapshoter User Pref Feed Me Shout-outHotels PETRI Site Pref Dist Logger Slicer eCom Renderer eCom Cart eCom Checkout eCom Catalog eCom Orders Payment Facade Account Info HTML API HTML Embeder Blog Mobile

It is all about

Microservices Guidelines Each service has its own DB schema (if one is needed) Only one service should write to a specific DB table(s) There may be additional read-only services that directly accesses the DB (for performance reasons) Services are stateless No DB transactions Cache is not a building block, but an optimization

Microservices Tradeoffs Each service has its own DB schema (if one is needed) Gain - Easy to scale microservices based on service level concerns Tradeoff – system complexity, performance Only one service should write to a specific DB table(s) Gain - Decoupling architecture – faster development Tradeoff – system complexity / performance May have additional read-only services that accesses the DB Gain - Performance gain Tradeoff - coupling Services are stateless Gain - Easy to scale out (just add more servers) Tradeoff - performance / consistency No DB transactions Gain - Better DB performance, easier to scale Tradeoff - system complexity

1. Editor Segment

Editor Server Immutable JSON pages (~3M / day) Site revisions Active – standby MySQL cross datacenters Editor Server MySQL Active Sites MySQL Archive

Protect The Data DB outage with fast recovery = replication Data poisoning/corruption = revisions / backup Make the data available at all times = data distribution to multiple locations / providers

Browser Editor Server GCS MySQL Active Sites MySQL Archive Saving Editor Data WixMedia (Amazon) WixMedia (Google) Save Page(s) 200 OK Upload Save Page DC replication Notify MySQL Archive MySQL Active Sites S3 WixMedia (DC-1)

Browser Editor Server GCS MySQL Active Sites MySQL Archive WixMedia (Amazon) WixMedia (Google) Save Page(s) 200 OK Upload Save Page DC replication Notify MySQL Archive MySQL Active Sites S3 WixMedia (DC-1) Self Healing Process

No DB Transactions Save each page (JSON) as an atomic operation Page ID is a content based hash (immutable/idempotent) Finalize transaction by sending site header (list of pages) Can generate orphaned pages, not a problem in practice

2. Media Segment (WixMP)

Wix Media Platform (WixMP) Eventual consistent distributed file system (2PB user media files) Dynamic media processing Multi datacenter aware Automatic fallback cross DC Run on commodity servers & cloud

T Google Cloud Prospero – Wix Media Manager get image.jpg First fallback Second fallback If not in CDN Amazon x36 T T x32 Austin CDN

3. Public Segment

Public Segment Roles Routing (resolve URLs) Dispatching (to a renderer) Rendering (HTML,XML,TXT) Public Server HTML Renderer HTML SEO Renderer Flash Renderer Sitemap Renderer Robots.txt Renderer Flash SEO Renderer

Public SLA Our goal: 99% response time <100ms at peak traffic

Publish Site Publish site header (a map of pages for a site) Publish routing table Publish site header / routes (CQRS) Editor SegmentPublic Segment

Built For Speed Minimize out-of-process hops (2 DB, 1 RPC) Lookup tables are cached in memory, updated every few minutes Denormalized data – optimize for read by primary key (MySQL) Minimize business logic

How a Page Gets Rendered Bootstrap HTML template that contains only data Only JavaScript imports JSON data (site-header + dynamic data) No “real” HTML view

Offload rendering work to the browser

The average Intel Core i750 can push up to 7 GFLOPS without overclocking

Why JSON? Easy to parse in JavaScript and Java/Scala Fairly compact text format Highly compressible (5:1 even for small payloads) Easy to fix rendering bugs and cross browsers issues (just deploy a new code)

Minimum Number of Public Servers Needed to Serve 66M Sites 4

Public SLA Be Available %

Serving a Site – Sunny Day Archive CDN WixMP Browser Store HTML to cache HTTP Request Notify site view LB Public Renderer HTML Resources / Media HTTP Request

Serving a Site – DC Lost Archive CDN WixMP Browser LB Public Renderer LB Public Renderer Change DNS HTTP Request

Serving a Site – Public Lost Archive Browser LB Public Renderer Get Cached HTML Version HTML HTTP Request LB Public Renderer Fallback to 2 nd DC

Living in the Browser CDN WixMP Browser LB Public Rendere r Editor Pages Fallback JSON / Media HTML HTTP Request Fallback

Summary Identify concerns and SLA for different parts of the system Build redundancy in critical path (for availability) De-normalize data (for performance) Minimize out-of-process hops (for performance) Take advantage of client’s CPU power

Aviran Mordo Head of linkedin.com/in/aviran aviransplace.com