Apache Cassandra in Action

Jonathan Ellis
[email protected] / @spyced

Why Cassandra?

Relational databases are not designed to scale B-trees are slow

and require read-before-write

(The eBay Architecture, Randy Shoup and Dan Pritchett)

Writer

Memtable

Reader

Commitlog

The Log-Structured Merge-Tree, Bigtable: A Distributed Storage System for Structured Data

Bigtable, 2006

Dynamo, 2007

OSS, 2008

Incubator, 2009

TLP, 2010

Cassandra in production

Digital Reasoning: NLP + entity analytics OpenWave: enterprise messaging OpenX: largest publisher-side ad network in the world Cloudkick: performance data & aggregation SimpleGEO: location-as-API Ooyala: video analytics and business intelligence ngmoco: massively multiplayer game worlds

FUD?

Cassandra is only appropriate for unimportant data.

Durabilty

Write to commitlog

fsync is cheap since its append-only

Write to memtable [amortized] flush memtable to sstable

SSTable format, briefly

<key 127> <key 255> ...

<row data 0> <row data 1> ... <row data 127> ... <row data 255> ... Sorted [clustered] by row key

Scaling

T L

T L

F
(A-L]

T L

(A-F]

T
(F-L]

Key C

T L

Reliability

No single points of failure Multiple datacenters Monitorable

Some headlines

Resyncing Broken MySQL Replication How To Repair MySQL Replication Fixing Broken MySQL Database Replication Replication on Linux broken after db restore MySQL :: Repairing broken replication

Good architecture solves multiple problems at once

Availability in single datacenter Availability in multiple datacenters

Y W A
Key C

T P

Y W A
Key C

hint P

X
L

Y W A

T P

Y W A
Key C

T P

Y W A
Key C

T L P

Tuneable consistency

ONE, QUORUM, ALL R+W>N Choose availability vs consistency (and latency)

Monitorable

JMX

OpsCenter

When do you need Cassandra?

Ian Eure: If youre deploying memcache on top of your database, youre inventing your own ad-hoc, difficult to maintain NoSQL data store

Not Only SQL

Curt Monash: ACID-compliant transaction integrity commonly costs more in terms of DBMS licenses and many other components of TCO (Total Cost of Ownership) than [scalable NoSQL]. Worse, it can actually hurt application uptime, by forcing your system to pull in its horns and stop functioning in the face of failures that a non-transactional system might smoothly work around. Other flavors of complexity can be a bad thing apply as well. Thus, transaction integrity can be more trouble than its worth. [Curts emphasis]

Keyspaces & ColumnFamilies

Conceptually, like schemas and tables

Inside CFs, columns are dynamic

Twitter: Fifteen months ago, it took two weeks to perform ALTER TABLE on the statuses [tweets] table.

ColumnFamilies

Static

Object data

Dynamic

Precalculated query results

static columnfamilies
Users
zznate driftx thobbs jbellis Password: * Password: * Password: * Password: * Name: Nate Name: Brandon Name: Tyler
Name: Jonathan

Site: riptano.com

dynamic columnfamilies
Following
zznate driftx thobbs jbellis zznate: driftx:
mdennis: pcmanus

driftx:

thobbs:

thobbs:

xedin:

zznate

Inserting

Really insert or update Not a key/value store update as much of the row as you want

Example: twissandra

http://twissandra.com

CREATE TABLE users ( id INTEGER PRIMARY KEY, username VARCHAR(64), password VARCHAR(64) ); CREATE TABLE following ( user INTEGER REFERENCES user(id), followed INTEGER REFERENCES user(id) ); CREATE TABLE tweets ( id INTEGER, user INTEGER REFERENCES user(id), body VARCHAR(140), timestamp TIMESTAMP );

Cassandrified
create column family users with comparator = UTF8Type and column_metadata = [{column_name: password, validation_class: UTF8Type}] create column family tweets with comparator = UTF8Type and column_metadata = [{column_name: body, validation_class: UTF8Type}, {column_name: username, validation_class: UTF8Type}] create column family friends with comparator = UTF8Type create column family followers with comparator = UTF8Type create column family userline with comparator = LongType and default_validation_class = UUIDType create column family timeline with comparator = LongType and default_validation_class = UUIDType

Connecting
CLIENT = pycassa.connect_thread_local('Twissandra') USER = pycassa.ColumnFamily(CLIENT, 'User')

User
RowKey: ericflo => (column=password, value=****, timestamp=1289446382541473) ------------------RowKey: jbellis => (column=password, value=****, timestamp=1289446438490709) uname = 'jericevans' password = '**********' columns = {'password': password} USER.insert(uname, columns)

Natural keys vs surrogate

Friends and Followers

RowKey: ericflo => (column=jbellis, value=1289446467611029, timestamp=1289446467611064) => (column=b6n, value=1289446467611031, timestamp=1289446467611080) to_uname = 'ericflo' FRIENDS.insert(uname, {to_uname: time.time()}) FOLLOWERS.insert(to_uname, {uname: time.time()})

zznate driftx thobbs jbellis

driftx:

thobbs:

zznate: driftx: mdenni s: pcmanu s: thobbs: xedin: zznat e:

Tweets
RowKey: 92dbeb50-ed45-11df-a6d0-000c29864c4f => (column=body, value=Four score and seven years ago, timestamp=1289446891681799) => (column=username, value=alincoln, timestamp=1289446891681799) ------------------RowKey: d418a66e-edc5-11df-ae6c-000c29864c4f => (column=body, value=Do geese see God?, timestamp=1289501976713199) => (column=username, value=pdrome, timestamp=1289501976713199)

Userline
RowKey: ericflo => (column=1289446393708810, value=6a0b4834-ed44-11dfbc31-000c29864c4f, timestamp=1289446393710212) => (column=1289446397693831, value=6c6b5916-ed44-11dfbc31-000c29864c4f, timestamp=1289446397694646) => (column=1289446891681780, value=92dbeb50-ed45-11dfa6d0-000c29864c4f, timestamp=1289446891685065) => (column=1289446897315887, value=96379f92-ed45-11dfa6d0-000c29864c4f, timestamp=1289446897317676)

Userline

zznate driftx thobbs jbellis

1289847840615: 3f19757a-c89d...

1289847887086: a20fcf52-595c...

1289847887086: a20fcf52-595c... 1289847840615: 3f19757a-c89d... 128984784425: 844e75e2-b546...

Timeline
RowKey: ericflo => (column=1289446393708810, value=6a0b4834-ed44-11dfbc31-000c29864c4f, timestamp=1289446393710212) => (column=1289446397693831, value=6c6b5916-ed44-11dfbc31-000c29864c4f, timestamp=1289446397694646) => (column=1289446891681780, value=92dbeb50-ed45-11dfa6d0-000c29864c4f, timestamp=1289446891685065) => (column=1289446897315887, value=96379f92-ed45-11dfa6d0-000c29864c4f, timestamp=1289446897317676)

Adding a tweet
tweet_id = str(uuid()) body = '@ericflo thanks for Twissandra, it helps!' timestamp = long(time.time() * 1e6) columns = {'uname': useruuid, 'body': body} TWEET.insert(tweet_id, columns) columns = {ts: tweet_id} USERLINE.insert(uname, columns) TIMELINE.insert(uname, columns) for follower_uname in FOLLOWERS.get(uname, 5000): TIMELINE.insert(follower_uname, columns)

Reads
timeline = USERLINE.get(uname, column_reversed=True) tweets = TWEET.multiget(timeline.values()) start = request.GET.get('start') limit = NUM_PER_PAGE timeline = TIMELINE.get(uname, column_start=start, column_count=limit, column_reversed=True) tweets = TWEET.multiget(timeline.values())

Programatically

Don't use thrift directly Higher level clients have a lot of features you want

Knowledge about data types Connection pooling Automatic retries Logging

Raw thrift API: Connecting

def get_client(host='127.0.0.1', port=9170): socket = TSocket.TSocket(host, port) transport = TTransport.TBufferedTransport(socket) transport.open() protocol = TBinaryProtocol.TBinaryProtocolAccelerated(transport) client = Cassandra.Client(protocol) return client

Raw thrift API: Inserting

data = {'id': useruuid, ...} columns = [Column(k, v, time.time()) for (k, v) in data.items()] mutations = [Mutation(ColumnOrSuperColumn(column=c)) for c in columns] rows = {useruuid: {'User': mutations}} client.batch_mutate('Twissandra', rows, ConsistencyLevel.ONE)

API layers

libpq JDBC JPA

Thrift Hector Hector objectmapper

Running twissandra

Login: notroot/notroot

(root/riptano)

cd twissandra python manage.py runserver & Navigate to http://127.0.0.1:8000 Login as jim/jim, tom/tom, or create your own

One more thing

!PUBLIC! userline

Exercise 1

$ cassandra-cli --host localhost ] use twissandra; ] help; ] help list; ] help get; ] help del; Delete the most recent tweet

How would you find this w/o looking at the UI?

Exercise 2

User jim is following user tom, but twissandra doesn't populate Timeline with tweets from before the follow action. Insert a tweet from tom before the follow action into jim's timeline

Secondary (column) indexes

Exercise 3

Add a state column to the Tweet column family definition, with an index (index_type KEYS).

Hint: a no-op update column family on Tweet would be update column family Tweet with column_metadata=[{column_name:body, validation_class:UTF8Type}, {column_name:username, validation_class:UTF8Type}]

Set the state column on several tweets to TX. Select them using get where.

Language support

Python

pycassa telephus

Ruby

Speed is a negative

Java

Hector

PHP

phpcassa

Done yet?

Still doing 1+N queries per page Solution: Supercolumns

Applying SuperColumns to Twissandra

jbellis 1289847840615 1289847844275
Id: Id: 3f19757a-c89d... 3f19757a-

1289847844275
Id: Id: 844e75e2-b546... 3f19757a-

1289847887086 1289847844275
Id: Id: a20fcf52-595c... 3f19757ac89d... uname: uname: zznate zznate

c89d... uname: uname: zznate zznate

body: body: O Do geese see stone be not so

c89d... uname: uname: driftx zznate body: body: Rise geese see Do to vote sir ...

...

body: body: Do Igeese see prefer pi ...

Supercolumns: limitations

Requires reading an entire SC (not the entire row) from disk even if you just want one subcolumn

UUIDs

Column names should be uuids, not longs, to avoid collisions Version 1 UUIDs can be sorted by time (TimeUUID) Any UUID can be sorted by its raw bytes (LexicalUUID)

Usually Version 4 Slightly less overhead

Lucandra

What documents contain term X?

and term Y? or start with Z?

Fields and Terms

<doc> <field name=title>apache talk</field> <field name=date>20110201</field> </doc>

feld
title title date

term
apache talk 20110201

freq
1 1 1

position
0 1 0

Lucandra ColumnFamilies
create column family documents with comparator = BytesType; Create column family terminfo with column_type = Super and comparator = BytesType and subcomparator = BytesType;

Lucandra data
Document Key col name value "documentId" => { fieldName , value } Term Key col name value "field/term" => { documentId , position vector }

Lucandra queries

get_slice get_range_slices No silver bullet

FAQ: counting

UUIDs + batch process column-per-app-server counter API (after 1.0 is out)

Locking

Zookeeper Cages: http://code.google.com/p/cages/ Not suitable for multi-DC

UUIDs
counter1 counter2 672e34a2-ba33... 3f19757a-c89d... b681a0b1-58f2... 844e75e2-b546... a20fcf52-595c...

counter1 counter2

aggregated: 27 aggregated: 42

Column per appserver

counter1 counter2 672e34a2-ba33: 12 3f19757a-c89d: 7 b681a0b1-58f2: 4 844e75e2-b546: 11 1872c1c2-38f1: 9

Counter API
key counter1: (14, 13, 9) counter2: (11, 15, 17)

General Tips

Start with queries, work backwards Avoid storing extra timestamp columns Insert instead of check-then-insert Use client-side clock to your advantage use TTL Learn to love wide rows