DISTRIBUTED COORDINATIONWITH PYTHON

Ben Bangertmozilla

Tools of the Trade

DISTRIBUTED COORDINATION IS NOT...

• Distributed Databases (Cassandra, Riak)

• Distributed Computing (Hadoop, etc.)

• Distributed Event Analysis (Storm)

The Common Element

Apache Zookeeper

ZooKeeper is a centralized service for maintaining configuration information,

naming, providing distributed synchronization, and providing group services.

ZOOKEEPER

WHY NOT USE...

• Memcached?

• MongoDB?

• Postgres/MySQL?

Hierarchical data structure in znodes

• Session Based

• Znode watches

• Ephemeral and Sequential Znodes

• Last for duration of client session

• Session dies when connection is closed or expires

• Can’t have children znodes

EPHEMERAL ZNODES

SEQUENTIAL ZNODES

• Supply a node name (or not), get node name back with a trailing sequence number (0001, 0002, 0003, etc.)

• Can be combined with ephemeral flag

BASIC COMMANDS

• create(PATH, DATA...)

• get(PATH...)

• get_children(PATH...)

• set(PATH, DATA...)

• delete(PATH...)

PYTHON CLIENTS

• txzookeeper

• kazoo

• unified client that works with gevent

• implements wire protocol in pure Python

USE KAZOO

EASY TO USE

from kazoo.client import KazooClient

client = KazooClient()client.start()

USE CASES

CONFIGURATION

• Store settings in node data

• Organize node structure

• Set watches on nodes of interest

PARTY MEMBERSHIP

• Join a party, find out who else is around

• Elect a leader if desired

• Recipe in Kazoo

LOCKS

• Lock a resource for a single client

• Lock a resource for multiple clients (Semaphore)

• Hard to write properly

• Recipe in Kazoo

BUILDING HIGHER LEVELABSTRACTIONS

ONZOOKEEPER

CAVEAT

DO NOT IMPLEMENT YOURSELFUSE THE RECIPE

BASIC STEPS

• Create lock parent node if needed

• Create ephemeral+sequence node under parent, store node name returned

• Get children of lock node

• Sort children list by sequence number

• First child in the list has the lock!

THINGS TO WATCH OUT FOR

• Avoid the thundering herd, use watches only when needed

• When our node isn’t the lowest, watch the one in front of us

• Only one client wanting a lock is ‘woken’ when the lock is released by a different client

HANDLING FAILURE

ROBUST CODE TAKES EFFORT

• What happens when a server fails?

• What happens when the client fails?

• What happens when we don’t know if the server has failed?

STOPPING WHEN UNCERTAIN

A BIT BETTER VERSION...

EVEN BETTER

FAILURE WILL HAPPEN

• Fail fast, fail completely.

• Session expiration is a good time to sys.exit

• Always include jitter (kazoo includes jitter on its connection and command retry operations)

• Consider what exceptions can occur in any code relying on a distributed system

• Distributed systems are hard

• Use existing battle-proven tools (Zookeeper, Kazoo)

• Always consider everything that can fail, and how

• Be wary of tools that don’t tell you how they fail

• Read Kyle Kingsbury’s Jepsen posts to see examples of systems failing: http://aphyr.com/tags/jepsen

FIN

QUESTIONS?