Presented at Data Day Texas 1/27/2018

1. Cassandra
Architecture
FTW!

3. Cassandra
Architecture
FTW!

4. Cassandra won because it solved a
problem that hadn’t yet been solved

5. What’s wrong with RDBMS?
• Pros
• Relational data modeling is well
understood
• SQL is easy to use and
ubiquitous
• ACID transactions - data integrity
• Cons
• Scaling is hard
• Sharding and replication have
side-effects (performance,
reliability, cost)
• Denormalize to get performance
gains
For fun: Relational Scaling and the Temple of Gloom

6. The NoSQL Revolution (~2009)
• Key-value – Dynamo, Riak, Voldemort, Redis,
Memcached
• Column-oriented – BigTable, Hbase
• Document – Mongo DB, Document DB, CouchDB
• Graph – Neo4J, DSE Graph
• Multi-model – DataStax Enterprise, CosmosDB
• Relational counter-revolution – AWS Aurora, Google
Spanner

7. Apache Cassandra Architecture Qualities
• Distributed,
decentralized
• Elastic scalability
• Commodity hardware?
• High performance
• High availability / fault
tolerant
• Tuneable consistency Apache Cassandra ® Apache Software Foundation

8. Two lies and a truth
• Cassandra is column-oriented
• Cassandra is a schemaless database
• Cassandra is eventually consistent
They’re all lies, in a way

9. Problems Cassandra is Especially Good At
• Large scale storage
• >10s of TB
• Lots of writes
• Time-series data, IoT
• Statistics and analytics
• For example, as a Spark data
source
• Geographic distribution
• Multiple data centers
Personalization
Customer
360
Recommendation
Fraud
Detection
Inventory
Management
Identity
Management
Security
Supply
Chain

10. A system must be designed for distribution
from the beginning in order to scale the most
effectively
© DataStax, All Rights Reserved.10

11. Cluster Topology
• Organization
– Nodes
– Racks
– Data Centers
• Goals
– Distribute copies (replicas) for
high availability
– Route queries to nearby nodes
for high performance
• Approaches
– Gossip
– Snitches

12. Clusters and Rings
• Organization
– Tokens
– Token Ring
• Goals
– Distribute data evenly across
nodes
• Approaches
– Partitioners
– Virtual nodes

13. Replication and Consistency
• Organization
– Client (with driver)
– Coordinator Node
– Replica nodes
• Goals
– Consistent data even when
nodes are down/unresponsive
• Approaches
– Replication Factor / Strategy
– Consistency Level

14. Cassandra works because it leverages
proven distributed system design patterns

17. Wait, who is this guy?

18. Proven Distributed System Techniques
• Gossip
• Failure detection (Phi)
• Partitioning
• Leaderless replication
• Hinted Handoff
• Bloom filters
• Consensus algorithms
(Paxos)
• Log Structured Merge Storage
Engine
– Memtables
– SSTables
– Commit log
– Compaction
– Tombstones
• Thrift API
• SEDA

19. Log Structured Merge Storage Engine

20. Cassandra’s data model is the key enabler
of its high performance and scalability

21. Terminology - Cassandra’s Data Model

22. Terminology – Partition and Clustering Keys

23. Partition and Clustering Key Example
Labeling:
K – partition key
C – clustering key
PRIMARY KEY ((customer_id), contact_time)

24. Cassandra survived because it didn’t
over-extend itself

25. Tuneable Consistency and Consistency Levels
• ONE, TWO, THREE
• Useful for speed
• ANY (Write only, use with care)
• ALL
• Number of nodes to respond =
RF
• Overly restrictive?
• QUORUM
• (RF / 2) + 1
• Frequently very useful
• LOCAL_ONE,
LOCAL_QUORUM
– Similar to above, but nodes must
be in local data center
• EACH_QUORUM
– Quorum of nodes must respond
in each data center

26. Strong Consistency vs. Eventual Consistency
• Eventual consistency
• i.e. W  ONE, R  QUORUM
• Use cases
• Write heavy
• Data not read immediately
• Strong Consistency
• i.e. W QUORUM, R 
QUORUM
• Use cases
• Read after write
• Data loading with validation
• Strong Consistency Formula
• R + W > RF = strong consistency
• R: read replica count required by
consistency level
• W: the write replica count
required by consistency level
• RF: replication factor
• Example: W QUORUM, R 
QUORUM, RF = 3
• 2 + 2 > 3
• Implication: all client reads will
see the most recent write

27. Inventory and Tuneable Consistency
© DataStax, All Rights Reserved.28
Approach Example Scope
QUORUM consistency
for reads and writes
Ensuring latest inventory counts
are always read
Data Tier
Lightweight Transaction Updating inventory counts Data Tier
Logged Batch
Writing to multiple denormalized
tables
Data Tier
Retrying failed calls
Data synchronization,
reservation processing
Service /
Application Tier
Compensating
processes
Verifying reservation processing System
Customer service
remediation
Rebooking, back order,
substitution
System
Eventual
consistency
Strong
consistency

28. Core
application
data
Microservices and Polyglot Persistence
© DataStax, All Rights Reserved.29
Servic
e A
Service
B
Tabular Key-value
Servic
e C
RelationalDocument Graph
Service
D
Service
E
Reference
data
Content
Customer
relationship
data
Legacy, low
volume data

29. Cassandra will remain viable for many years
because of its extensibility and pluggability

30. Snitch
Partitioner
Compaction
Replication
Secondary
Index
Storage
Engine?
Auth/Auth

31. Cassandra’s deployment flexibility can’t be
matched by cloud vendor database offerings

32. Consistency Level and Deployment
Active-Active, Multi-Region
– EACH_QUORUM may limit availability if connection to one region/data center is down
– LOCAL_QUORUM, relying on Cassandra to complete writes to remote data centers
– QUORUM is a reasonable middle-ground approach

33. Consistency Level and Deployment
Separate data center for analytics
– Writes in ”online DC” – LOCAL_QUORUM
• Writes to analytics DC in background
• Analytic DC availability/performance decoupled from online DC
• EACH_QUORUM, QUORUM overkill, unless “real-time” analytics required
– Reads in “analytics DC” – LOCAL_QUORUM
• Or even LOCAL_ONE

34. Cassandra is most powerful when combined
with complementary technologies to form a
data platform

35. Spark + Cassandra
• Access Cassandra from
Spark via DataStax connector
DataStax is a registered trademark
of DataStax, Inc. and its
subsidiaries in the United States
and/or other countries.
3
6
• Co-locate Spark and
Cassandra
Spa
rk
SQL
Spark
Streami
ng
MLi
b
Grap
hX
Spar
k R
Spark Core Engine
DataStax Spark-Cassandra Connector
Cassandra

36. DSE Analytics
DataStax is a registered trademark
of DataStax, Inc. and its
subsidiaries in the United States
and/or other countries.
Application
Real Time Operations
Cassandra
Analytics
Analytics
Queries
Your Analytics
Real Time Replication
Single DSE Custer
Streaming, ad-hoc, and batch
• High-performance
• Workload management
• SQL reporting
Compared to self-managed
Spark cluster:
• No ETL
• True HA without Zookeeper

37. DSE Core - Certified Apache Cassandra
• The best distribution of Apache Cassandra™
• Production certified Cassandra
• Performance improvements
• Advanced Security
• Multi-tenancy through row-level access control
• Advanced Replication
• Great for retail and IoT use cases
DataStax is a registered trademark of DataStax, Inc. and its subsidiaries in the United States
and/or other countries.

38. https://github.com/killrvideo
KillrVideo – a reference application

39. https://academy.datastax.com
DataStax Academy – a place to learn

40. 41
Contact
jeff.carpenter@
datastax.com
@jscarp
jeffreyscarpenter
medium.com/@jscarp