Theo Schlossnagle gave a talk to the Ny Cto Club about Scalabil Ity

models

• Enter Prise class: 5-9s for 100% of population: reality everyone disappointed

• Carrier class: 5-9s for 5-9s of population: few miserable, most happy

  • example (amazon): block some portion of new sessions at the [Load Balancer] (need Load Balancing hardware with that feature) (could also do in software layer, but that uses the processor that's already the bottleneck)

Software architecture

• db federation: don't throw away data integrity universally (because some app will screw up, and you'll get dirty data), pick key things to sacrifice

  • pick doc-like fields, put them in CouchDb (or File System...) (Keep db as small as possible)

• Typically run at 30% capacity, plus have other hardware available

• For each page or action, count number of [IO] ops (incl log writing, etc)

• Logging: learn from your own history (pick right metrics)

• Identify component that needs to scale, build from the start on 3 boxes: going from 3 to bigger is much easier

Cloud

• no real guarantees

• [Euca Lyptus] (xen): Open Source Amazon Ec2 clone to run yourself

• Amazon Ec2 great for math (video transcoding)

• Scaling not fast enough for digg: takes 10min to get new machine up and tested - by that time it's too late

Chat afterwards

• don't automate failover if large cost for false alarm (eg in MySQL you may lose some transactions. If that's ok, then is it always ok, so you can architect for that?)

• MySQL is their most prevalent db, but they won't use it for important data. They'll use Postgre S Q L, Ms Sql Server, or other (even Oracle)

User Options Recent Changes Help Page