web-scalability-presentation

View Github Repository Open presentation in a new window

dwurf

See all presentation from dwurf

web-scalability-presentation

0 0

web-scalability-presentation

Presentation on web scalability, using LatLonGO as a case study

On Github dwurf / web-scalability-presentation

Web Scalability

Review of a scaling web architecture

Darren Wurf

Scale-up vs scale-out

Scale-up: Buy bigger machines

Scale-out: Buy more machines

Scale-up is easier if you are lucky enough to be able to do it resource sliders (CPU/memory) Scale-up not always possible (resource contention, HA, c20k) Scale-out must be designed into the product

HTTP: the stateless protocol

  • A stateless architecture is easily scalable
  • Cookies are used to maintain state between requests
  • Sites with a login/password are typically stateful (session credentials)"stateless": client makes request web server responds and forgets* Examples of state: shopping cart, member area, facebook wall

HTTP: the stateless protocol

Requests could go to any web server Client doesn't know or care which server handled the request * Next request could go to a different server

Sessions and caching

  • Multiple web application servers need to share session data
  • Store authentication data in a central server (eg LDAP/DB)
  • Cache session data on the web server (memcached)

Other caching

  • Client-side response caching: Last-Modified / E-Tags
  • Server-side response caching: memcached

High availability

  • Stateless architectures can also be made highly available
  • Remove all single points of failure

Reference architecture

Rackspace Cloud reference architecture, from Rackspace's documentation (CC by/nc/nd)

Web appliation tier can scale out, failures do not take down system Highly available load balancer and database are provided by rackspace

Case Study: LatLonGO

Product developed by my employer Utility GIS data on your tablet * For field construction / repair crews

LatLonGO: Overview

  • Render layers in corporate GIS
  • Serve data (usually as deltas) to tablets in the field
  • Office workers use web interface
Utilities want to be able to find assets in the field Solves the problem of "secure, disconnected access" * I don't care about the tablet aspect at all

LatLonGO: Web interface

  • Started as a "preview" for uploaded layers
  • Clients liked it
  • Web interface must scale to 100s of concurrent users
Must be highly available, 100s of concurrent users == 1000s of requests/sec

LatLonGO: Technology stack

  • Nginx
  • Django
  • Some C++ magicExposed via ctypes

LatLonGO: Bottleneck

Django

tiles Django can be easily scaled to 100s of req/s

LatLonGO: Scaling up

  • Expose C++ code as FastCGI
  • Cache sessions in memcached
  • Cache tiles in memcached (optional)Problem: tile access is not authenticated nginx module called auth_request - delegate authentication configuration trickery: can delegate authentication to memcached

LatLonGO: Scaling out

  • HTTP is stateless
  • Django stores session details centrally (DB)
  • FastGCI/C++ is stateless
  • memcached is stateless
  • Copy-paste servers!we can use a load balancer or round-robin DNS to distribute load For consistent updates, use an ip-hashing load balancer

LatLonGO: High availability

  • Load balancer
  • Django database
  • LDAP
  • Layer/tile storage

THE END

Darren Wurf

http://www.latlongo.com/

Created using reveal.js