Keynote - Systems at Facebook Scale

Here is the link.

Tech-lead of web foundation team

Facebook in 30 seconds

Rapid change

- Code released twice a day
- Rapid feature development - e.g. Lookback videos
 - 450 Gbps of egress
 - 720 million videos rendered ( 9 million/ hour)


Proactive design for scaling

Solve scaling problems only once
All Facebook projects in a single source control repo - easy code reuse

Folly: base C++ library
Thrift: RPC
Proxygen: HTT(s) server
RocksDB: persistent key-value store

Efficient synchronization

Producer/Consumer queues

How to implement Producer/Consumer

Multiple wakeups/ Deque

Potential context switches

Word thread ordering
FIFO
LIFO

Good queueing
Deal with a burst of load
Processing speed > arrival speed
Increase reliability

Bad queueing
Server is overloaded
Processing speed < arrival speed
Causes latency

Overload handling philosophy
If you're going to fail, fail quickly - prevents delaying the overall request
Servers should signal overload - OK to say "I can't help you right now"
Order doesn't matter - servers need not first-com-first-serve
Clients should defend themselves - don't rely on the server
complex knobs are tuned poorly - design parameter-free abstractions

Controlled delay