Apache Cassandra

An eventually consistent database,
that scales with ease

by Sabine Maennel
at pydata Zürich 25.1.2017

Overview

• Understanding Cassandra
• Cassandra in Action: Twissandra

Some characteristics

• A NoSQL database
• for big distributed data
• no single point of failure
• quick response times
• runs on cheap hardware spread to datacenters around the globe

History

1. Facebook invented Cassandra
2. Influenced by Google and Amazon
3. Today its driven by Apache as an Open Source Project

How I met Cassandra

• exercise of analysing logfiles
• on second thought, it was a solution for scaling

coming from an RDBMS background ...

• the terminology is mostly the same

But Casssandra is different ...

Cassandra is a mulitlevel-map rather then a structure

• rows are a map!
• columns are a map!
• instead of transactions Cassandra's reads and writes are more like a multistep operations of copying and retrieving data

How to think of Cassandra

1. Cassandra is mostly hosted on servers, that form rings: "clusters"
2. there is no manager node
3. they talk via a Gossip-Protocol

A table is distributed

1. Partitions of the table are mapped to different serves in the ring
2. the mapping is done by a hashing algorithm
3. there are replications of each row -> replication-factor

Why replications of data?

• Availability: quick responses
• Partition tolerance: works even when some servers are down
• Conistency: all stored data are up to date

Now we understand why a table is distributed

• Partition tolerance

Imagine a query in this distributed system

1. it does not work!
2. but, some rows are closer then others ...

partitions are cluster of rows

the primary key has two parts:

• partions are determined by the partition key
• rows in a partion are sorted by the clustering key, which may be queried by range

rows contain maps rather than columns

• tables are also called column familys
• they consist of key-value pairs
• in an RDBMS these would be sparse tables

Columns are maps

1. they consist of key-value pairs
2. they come with a timestamp
3. they may even expire

remember ...

• "map of maps"
• "map of maps of maps"
• "containers that hold collections of column objects"
• "columns ... as 3-tuples"

What does eventually consistent mean?

-> look at how Cassandra reads and writes

• write consistency: how many writes must be okay for write success?

Cassandra is consistent if ...

• write consistency + read consistency > replication factor

Lets look at an example Twissandra

        cqlsh> CREATE KEYSPACE IF NOT EXISTS twissandra 
               WITH REPLICATION = {'class': 'SimpleStrategy', 'replication_factor': 1};
        

        cqlsh> DESCRIBE twissandra;
        CREATE KEYSPACE twissandra WITH replication = {'class': 'SimpleStrategy', 'replication_factor': '1'}  AND durable_writes = true;
        

Now we have to use that keyspace

        cqlsh> USE twissandra; 
        

We are ready to create the first table:

        
        cqlsh> CREATE TABLE users (
               username text PRIMARY KEY,
               password text);
        

• the partion key is username

Look at your table

        cqlsh> DESCRIBE users; 
        CREATE TABLE twissandra.users (
            username text PRIMARY KEY,
            password text
        ) WITH bloom_filter_fp_chance = 0.01
            AND caching = {'keys': 'ALL', 'rows_per_partition': 'NONE'}
            AND comment = ''
            AND compaction = {'class': 'org.apache.cassandra.db.compaction.SizeTieredCompactionStrategy', 'max_threshold': '32', 'min_threshold': '4'}
            AND compression = {'chunk_length_in_kb': '64', 'class': 'org.apache.cassandra.io.compress.LZ4Compressor'}
            AND crc_check_chance = 1.0
            AND dclocal_read_repair_chance = 0.1
            AND default_time_to_live = 0
            AND gc_grace_seconds = 864000
            AND max_index_interval = 2048
            AND memtable_flush_period_in_ms = 0
            AND min_index_interval = 128
            AND read_repair_chance = 0.0
            AND speculative_retry = '99PERCENTILE';        
        

There are a lots of defaults in place ...

• bloomfilters
• memtable_flush_period
• gc_grace_seconds
• read_repair_chance

Datamodelling in Cassandra: thinking in queries

Following and Followers

        
        cqlsh> 
            -- "username" follows "followed"
            CREATE TABLE following (
                username text,
                followed text,
                PRIMARY KEY(username, followed)
            );
            -- "username" is followed by "following"
            CREATE TABLE followers (
                username  text,
                following text,
                PRIMARY KEY(username, following)
            );
        

• in a RDBMS there would be only one table
• partition key username
• clustering key followed / following

Tweets and Userline

        
             CREATE TABLE tweets (
                tweetid uuid PRIMARY KEY, 
                username text, 
                body text
             );
            CREATE TABLE userline (
                tweetid  timeuuid,
                username text,
                body     text,
                PRIMARY KEY(username, tweetid)
            );
        

• two times the same table
• tweets has partition key tweetid
• userline has partition key username

Timeline

        
            CREATE TABLE timeline (
                username  text,
                tweetid   timeuuid,
                posted_by text,
                body      text,
                PRIMARY KEY(username, tweetid)
            );
        

• almost the same table
• posted_by was added

Live Demo on a virtual machine

• you need Java
• Cassandra
• ccm (Cassandra Cluster manager)

Twissandra sample data

• When Tom tweets: Marys and Susans and Johns timelines must be updated in addition to his!

Useful Links and some companies, that use Cassandra

• Tutorialspoint
• CCM (Cassandra Cluster Manager)
• Apache Cassandra