Why is it hard? (Big store system)

Performance -> Sharding -> constant faults

Faults -> fault tolerant systems

Tolerance -> Replication

Replication -> In consistency

Consistency -> Low Performance

Consistency: like talking to one server

STRONG CONSISTENCY

Master - mapping : filename, and position

Chunk server 0

Chunk server 1

…

64MB per chunk

Master Data:

2 main tables: (On ram)

  • filename -> array of chunk handles (nv)

  • chunk handles ->list of chunk servers (v)

    ​ version number (nv)

    ​ primary (v)

    ​ lease expiration (v)

Will store some on disk

LOG, checkpoint(snapshot) - disk

n - not v - volatile

restart: recover to snapshot and replay LOG.

READ

  • client send filename and offset to mcoinaster

  • master sends handle, list of servers

    (cached. Next time, won’t ask master)

  • client -> chunk servers: find the file

    • <- return the data

WRITES

if NO PRIMARY? - ON MASTER

find up to date replica (with version number)

pick primary and secondary

increment version number

tells primary, secondary the version number - lease

master write version number to disk

primary picks an offset

all replicas told to write at offset

if all “yes”, primary “success” -> client

else “no” -> client, and client reissue