Intro

Choosing proper boundaries between functions is the primary activity of system designer
discusses class of function placement that is used without explicit recognition
system w/ communication has a firm interface between communication subsystem and the rest of the system
function in an application can completely and correctly be implemented with only the knowledge and help of application standing at the ends of the communication system.
providing a function as a feature of the communication system is not possible (since both ends need it to communicate properly)

E2E Caretaking

consider a file transfer program from host A to B
- threats:
  - reading file from original disk could be corrupted
  - errors in copying or buffering data from A or B
  - CPU or memory transient error during buffering/processing
  - communication system drops or changes bits
  - one or both hosts may crash
is adding network features to the comm system useful to the application? (redundancy, packet checksum, etc)
- argument that while it eliminates one of the threats, in the end it is really only reducing the number of retries.
- communication system adding reliability does not decrease the work of the application developer
bug in reliable system caused headache for users in MIT networks
- gateway was flipping 1/1M bytes

reliability of networks is for performance rather than correctness
- large file transfer with unreliable packets – small probability to succeed with many retries - reliable transfer increases successful network transfer but doesn’t eliminate other issues
applications pay, even when they don’t want/need it

Delivery Guarantees
- ARPANet used RFNM (request for next message) which isn’t really useful to applications because the apps want to know if the receiving end acted on the message, rather than if it is just requesting the next
Secure data transmission
- authenticity of messages still needs to be checked, even if the data coming over the wire is encrypted.
Duplicate message suppression
- Applications may duplicate messages anyways during failure/retry scenarios, so why should the network protocol suppress them?
Guaranteeing FIFO Message Delivery
transaction mgmt