Fast Data at Massive Scale Lessons Learned at Facebook Bobby Johnson.

Fast Data at Massive Scale

Lessons Learned at Facebook

Bobby Johnson

Me

• Director of Engineering– Scaling and Performance– Site Security– Site Reliability– Distributed Systems– Development tools– Customer Service Tools

• Took Facebook from 7M users to 120M.

ArchitectureLoad Balancer

(assigns a web server)

Web Server (PHP assembles data)

Memcache(fast)

Database(slow, persistent)

Other servicesSearch, Feed, etc(ignore for now)

- 1/2 the time is in PHP

- 1/4 is in memcache

- 1/8 is in database

One year ago, almost half the time was memcache

Network Incast

PHP Client

Switch

memcache memcache memcache memcache

Many Small Get Requests

Network Incast

PHP Client

Switch

memcache memcache memcache memcache

Many big data packets

Clustering

PHP Client

memcache

10 objects

1 round trip for 10 objects

Clustering

PHP Client

memcache memcache

5 objects5 objects

- 2 round trips total

- 1 round trip per server

- longest request is 5

Clustering

PHP Client

memcache memcache

3 objects4 objects

- 3 round trips total

- 1 round trip per server

- longest request is 4

memcache

3 objects

Clustering

• If objects are small, round trips dominate so you want objects clustered

• If objects are large, transfer time dominates so you want objects distributed

• In a web application you will almost always be dealing with small objects

Caching

- Basic tools are parallelism and clustering

- Clustering is a latency/throughput tradeoff

- Application code must be aware

- Networking is a burst problem

- Dropped packets kill you

- TCP quick ack

PHP CPU

Application Improvements

know what your libraries do

$results = get_search_results( $needle );

foreach ( $results as $result ) {

if ( is_pending_friend( $result[‘id’] ) ) {

// we’ll change the links based on this

$result[‘pending’] = true;

}

}

know what your libraries do

function is_pending_friend( $id ) {

// this is short-lived, so don’t cache

expensive_db_query( $id …)

Databases

- Tend to be slower than lighter weight alternatives, so avoid using them

- If you do use them partition them right from the start

- If a query is _really_ slow, like a few seconds or a few minutes, you probably have a bug where you’re scanning a table

- The db should have a command to tell you what index it’s using for a query, and how many rows it’s examining

General Lessons

• Your best tool is parallelism

• Look at your data

• Build tools to look at your data

• Don’t make assumptions about what components are doing

• Algorithmic and system improvements are almost always better than micro-optimization