Hundreds of HVs Thousands of HVs Tens of Thousand HVs Hundreds of Thousand HVs Global Cloud Region Region Cell Cell Cell HV HV HV HV HV HV Cell Cell Region 5 What does “At Scale” Mean? Learning to Scale OpenStack
• Could use LB & node rota?on • Some stacks handle this well • Challenge for infrastructure Deploy Strategy 0110010001100101011100100111000001101100011011110111100100100001 LB Web01-‐A Web02 –A Web01-‐B Web02-‐B
• Debs + freight + apache • Configura?on with Puppet w/ PuppetMasters • Bash + ssh will take you a long way. “Agile” deploy process 011010000110000100101100001000000110000101100111011010010110110001100101
• Is 99.9% system online good or bad? – 10,000*.001 = 10 down nodes – Constantly fixing • Build your system to deal with breakage – Autohealing will help you stay sane Deal with Breakage 011000100111010101110011011101000110010101100100001011100010111000101110
11 Improving the Deploy Mechanism 01101000011001010110110001110000001000000110110101100101 • Easy enough with 100 nodes • Now scale to 1,000… then 10,000 • OS independent • Build for 10x more
• Switched from Debian packages to virtual environments Package • Torrent for package, pssh for fact files, and mcollec?ve for ac?ons Distribute • From centralized puppet master to decentralized masterless puppet Execute 12 Improving the Deploy Mechanism Deploying from OpenStack Trunk
• Torrents save bandwidth • Pssh for quick interac?ons • Mcollec?ve for lots of ac?ons Pushing bits 011001110110111101100111011011110110011101101111 • Torrent for package, pssh for fact files, and mcollec?ve for ac?ons Distribute
• S?ll need configura?on management • Configura?ons reside on the nodes • Less infrastructure to maintain Configure and Execute 0111001101100001011101100110010100100000011001100110010101110010011100100110100101110011 • From centralized puppet master to decentralized masterless puppet Execute
• We build Infrastructure as a Service • Built like an applica?on – Transac?onal – “S?cky” data • Deploy it like a website SHll a ways to go… 0110110101101111011100100110010100111111
• Reduced deploy ?me from hours to minutes • Most ?me spent tes?ng • Deployment tools are part of the product, not an aoerthought End Results? 011100110110110001100101011001010111000000100001