Giant Scale Services
Summary
Generic Service Model of Giant Scale Services
Load balancer in front of server, each with data store
DQ Principle
yield = completed / offered
harvest = viewed / full
DQ means Data Rate per Query * Data Served
- DQ is constant
- Increase D by Decreasing Q
DQ is a better metric than uptime
Online Evaluation and Growth
How to do you upgrade servers without bringing down the cluster?
Fast
- offpeak (e.g. use
diurnal
server property)
Rolling upgrade
- one by one
Big Flip
- Half the nodes at once
DQ Loss is same for all
- DQ * upgrade time per node * number of nodes