Lecture 19: Cloud Computing
Notes By: Connor Fong and Rohit Seth

• One unit of Computation = one container
• Power is a problem
  • Low power nodes are a plus
  • Use "sleep mode" when not in use.

• Short-term commitment
• As needed payment
• Can grow "instantly"
• Unknown demand (eg. startups)
• Varying demand (eg. The Superbowl)

• Who pays? Causes accounting problems!
• Who controls it?
  • Can it be subleased?
  • How to manage resources?
  • Security issues
• Software Licensing (eg. charge per core)
• Data confidentiality. Requires a minimum level of trust with your vendor.
• Data transfer bottlenecks
  • Sometimes the "Sneakernet" alternative is faster
  • Archiving data (Data mitigation problem)

• Overload risk. What if your network provider is flakey? Multiple providers increases complexity.
• Global meltdown via DDoS attacks. This is a huge problem!
• Lock-in: EC2 APIs -> AWS is not very portable
• Scaling-up risks. Bugs often occur after scaling up in production.

• Consistency - All parts of cloud agree on user-visible state
• Availability - Questions are resolved quickly
• Partition Tolerance - Cloud still works when partitioned

You can chose any two, but you can't get all three.

SQL:

• Atomicity
• Consistency
• Isolation (Serializability)
• Durability

No SQL

• Basically
• Available (Works fine if only reading)
• Soft state - Eventually stabilizes
• Eventually Consistent

Latency Issues often predominate. What can you do?

• Retry the operation.
• Cancel the operation.
• Wait for the operation to finish.

• Merge changes
• How do you deal with conflict? Prefer commutative operations
• Some problems can't be automatically fixed
  • Make manual repairs
  • Compensate users for downtime

All programs aren't trusted! Therefore, all operations are checked by the underlying system.
Some programs have special rights:

• The kernel itself
• uid programs

This is the part of hardware and software that you must trust. Keep this small! Only trust:

• The kernel
• Setuid Programs