Week 4, Day 1 (Monday Jan 27)

In-class exercise

List the biggest unknowns or risks you see for yourself or your team in Assignment 2. Items to think about:

In Part 1:

• How will you format the messages from the server to the worker?

In Part 2:

• How will you set up an EC2 instance so that it will start the worker process when CloudWatch starts a new instance?
• How will you force the load high enough to require that new instances be started?

Reducing latency of service requests

• The SLA for our service might be legally binding or a market requirement.
• Variable latencies are inevitable.
• How to keep our latencies within our (binding or best-effort) SLA?

• Ways to reduce latency of a single request:

  • Hedge the request: Wait a bit, request a duplicate, cancel second on first reply
  • Tie the request: Issue two requests, cancel second once first begins

Reading guide to next class

Read There Is No Getting Around It: You Are Building a Distributed System, up to but not including “Messaging”, p. 67.

Note that (as he says on p. 85), his use of “distributed system” is synonymous with how we use “cloud”.

Key point:

While much of the discussion, literature, and presentations focus on the en vogue technologies such as Paxos, Dynamo, or MapReduce … the reality of building a large distributed system is that many “traditional” problems are left unsolved by off-the-shelf solutions, whether commercial or open source. (p. 64)

You don’t need to pay close attention to:

• The opening pages (pp. 63–64), giving two reasons for using a distributed system, are background.