Cassandra Deep Diver & Data Modeling

C A S S A N D RA - A RC H I T E C T U R E & D ATA M O D E L I N G

CASSANDRA

AR

CH

I TE

CT

UR

E,

I NT

ER

NA

L S &

DA

T A M

OD

EL I N

G

1


REQUISITE SLIDE – WHO AM I?

- Brian Enochson- Home is the Jersey Shore- SW Engineer who has worked as designer / developer on Cassandra- Consultant – HBO, ACS, CIBER – - Specialize in SW Development, architecture and training

2


REQUISITE SLIDE # 2 – WHAT ARE WE TALKING ABOUT?• Cassandra Intro & Architecture• Why Cassandra• Architecture• Internals• Development

• Data Modeling Concepts• Old vs. New Way• Basics• Composite Types• Collections• Time Series Data• Counters

3


REQUISITE SLIDE #3 – THE NAME CASSANDRA

Cassandra was the most beautiful of the daughters of Priam ad Hecuba, the king and queen of Troy. She was given the gift of prophecy by Apollo who wished to seduce her; when she accepted his gift but refused his sexual advances, he deprived her prophecies of the power to persuade.**

Relate this to a database ->

Can see into the future. Good result for a database. This is good…

Can predict the future, but no one believed what was said!!! This is not so good….

** http://www.pantheon.org/articles/c/cassandra.html

Anyway, Cassandra is also universally known as C*

4

http://www.pantheon.org/articles/c/cassandra.html




HISTORY

• Developed At Facebook, based on Google Big Table and Amazon Dynamo **

• Open Sourced in mid 2008

• Apache Project March 2009

• Commercial Support through Datastax (originally known as Riptano, founded 2010)

• Used at Netflix, eBay and many more. Reportedly 300 TB on 400 machines largest installation

• Current verson is 1.2.x, 2.0 in RC1.

5


WHY EVEN CONSIDER C*

• Large data sets

• Require high availability

• Multi Data Center

• Require large scaling

• Write heavy applications

• Can design for queries

• Understand tunable consistency and implications (more to come)

• Willing to make the effort upfront for the reward

6


SOME BASICS

• ACID

• CAP Theorem

• BASE

7


ACID

Everyone has heard of ACID

• Atomic – All or None

• Consistency – What is written is valid

• Isolation – One operation at a time

• Durability – Once committed to the DB, it stays

This is the world we have lived in for a long time…

8


CAP THEOREM (BREWERS)

Many may have heard this one

CAP stands for Consistency, Availability and Partition Tolerance• Consistency –like the C in ACID. Operation is all or nothing,

• Availability – service is available.

• Partition Tolerance – No failure other than complete network failure causes system not to respond

So.. What does this mean?

** http://www.cs.berkeley.edu/~brewer/cs262b-2004/PODC-keynote.pdf

9

http://www.cs.berkeley.edu/~brewer/cs262b-2004/PODC-keynote.pdf

http://www.cs.berkeley.edu/~brewer/cs262b-2004/PODC-keynote.pdf


YOU CAN ONLY HAVE 2 OF THEM

10

Or better said in C* terms you can have Availability and Partition-Tolerant AND Eventual Consistency.

Means eventually all accesses will return the last updated value.


BASE

But maybe you have not heard this one…

Somewhat contrived but gives CAP Theorem an acronym to use against ACID… Also created by Eric Brewer.

Basically Available – system does guarantee availability, as much as possible.

Soft State – state may change even without input. Required because of eventual consistency

Eventually Consistent – it will become consistent over time.

** Also, as engineers we cannot believe in anything that isn’t an acronym!

11


C* - WHAT PROBLEM IS BEING SOLVED?

• Database for modern application requirements.

• Web Scale – massive amounts of data• Scale Out – commodity hardware• Flexible Schema (we will see this how this concept is evolving)• Online admin (add to cluster, load balancing). Simpler operations• CAP Theorem Aware

• Built based on• Amazon Dynamo – Took partition and replication from here **

• Google Bigtable – log structured column family from here ***

** http://www.allthingsdistributed.com/2007/10/amazons_dynamo.html*** http://research.google.com/archive/bigtable.html

12

http://www.allthingsdistributed.com/2007/10/amazons_dynamo.html

http://www.allthingsdistributed.com/2007/10/amazons_dynamo.html

http://research.google.com/archive/bigtable.html


C* BASICS

• No Single Point of Failure – highly available. • Peer to Peer – no master

• Data Center Aware – distributed architecture

• Linear Scaling – just add hardware

• Eventual Consistency, tunable tradeoff between latency and consistency

• Architecture is optimized for writes.

• Can have 2 billion columns!

• Data modeling for reads. Design starts with looking at your queries.

• With CQL became more SQL-Like, but no joins, no subqueries, limited ordering (but very useful)

• Column Names can part of data, e.g. Time Series

Don’t be afraid of denormalized and redundant data for read performance. In fact embrace it! Remember, writes are fast.

13


NOTE ABOUT EVENTUAL CONSISTENCY

** Important Term **

Quorum : Q = N / 2 + 1.

We get consistency in a BASE world by satisfying W + R > N

3 obvious ways:

1.W = 1, R = N

2.W = N, R = 1

3.W = Q, R = Q

(N is replication factor, R = read replica count, W = write replica count)

14


THE C* DATA MODEL

C* data model is made of these:Column – a name, a value and a timestamp. Applications can use the name

as the data and not use value. (RDBMS like a column).

Row – a collection of columns identified by a unique key. Key is called a partition key (RDBMS like a row).

Column Family – container for an ordered collection rows. Each row is an ordered collection of columns. Each column has a key and maybe a value. (RDBMS like a table). This is also known as a table now in C* terms.

Keyspace – administrative container for CF’s. It is a namespace. Also has a replication strategy – more late. (RDBMS like a DB or schema).

Super Column Family – say what?

15


SUPER COLUMN FAMILY

Not recommended, but they exist. Rarely discussed

It is a key, that contains to one or more nested row keys and then these each contain a collection of columns.

Can think of it as a hash table of hash tables that contain columns..

16


ARCHITECTURE (CONT.)

17

http://www.slideshare.net/gdusbabek/data-modeling-with-cassandra-column-families




OR CAN ALSO BE VIEWED AS…

18





TOKENS

Tokens – partitioner dependent element on the ring.

Each node has a single unique token assigned.

Each node claims a range of tokens that is from its token to token of the previous node on the ring.

Use this formula

Initial_Token= Zero_Indexed_Node_Number * ((2^127) / Number_Of_Nodes)

In cassandra.yaml

initial token=42535295865117307932921825928971026432

** http://blog.milford.io/cassandra-token-calculator/

19

http://blog.milford.io/cassandra-token-calculator/


C* PARTITIONER

RandomPartitioner – MD5 hash of key is token (128 bit number), gives you even distribution in cluster. Default <= version 1.1

OrderPreservingPartitioner – tokens are UTF-8 strings. Uneven distribution.

Murmur3Partitioner – same functionally as RandomPartitioner, but is 3 – 5 times faster. Uses Murmur3 algorithm. Default >= 1.2

Set in cassandra.yaml

partitioner: org.apache.cassandra.dht.Murmur3Partitioner

20


REPLICATION

• Replication is how many copies of each piece of data that should be stored. In C* terms it is Replication Factor or “RF”.

• In C* RF is set at the keyspace level:CREATE KEYSPACE drg_compare WITH replication = {'class':'SimpleStrategy', 'replication_factor':3};

• How the data is replicated is called the Replication Strategy• SimpleStrategy – returns nodes “next” to each other on ring,

Assumes single DC• NetworkTopologyStrategy – for configuring per data center.

Rack and DC’s aware.update keyspace UserProfile with strategy_options=[{DC1:3, DC2:3}];

21


SNITCH

• Snitch maps IP’s to racks and data centers.

• Several kinds that are configured in cassandra.yaml. Must be same across the cluster.

SimpleSnitch - does not recognize data center or rack information. Use it for single-data center deployments (or single-zone in public clouds)

PropertyFileSnitch - This snitch uses a user-defined description of the network details located in the property file cassandra-topology.properties. Use this snitch when your node IPs are not uniform or if you have complex replication grouping requirements.

RackInferringSnitch - The RackInferringSnitch infers (assumes) the topology of the network by the octet of the node's IP address.

EC2* - EC2Snitch, EC2MultiRegionSnitch

22


RING TOPOLOGY

When thinking of Cassandra best to think of nodes as part of ring topology, even for multiple DC.

23


SimpleStrategy

Using token generation values from before. 4 node cluster. Write value with token 32535295865117307932921825928971026432

24


SimpleStrategy #2

25


SimpleStrategy #3

With RF of 3 replication works like this:

26


NetworkTopologyStrategyUsing LOCAL_QUORUM, allows write to DC #2 to be asynchronous.

Marked as success when writes to 2 of 3 nodes (http://www.datastax.com/dev/blog/deploying-cassandra-across-multiple-data-centers)

27

http://www.datastax.com/dev/blog/deploying-cassandra-across-multiple-data-centers

http://www.datastax.com/dev/blog/deploying-cassandra-across-multiple-data-centers


COORDINATOR & CL

• When writing, Coordinator Node will be selected. Selected at write (or read) time. Not a SPF!

• Using Gossip Protocol nodes share information with each other. Who is up, who is down, who is taking which token ranges, etc. Every second, each node shares with 1 to 3 nodes.

• Consistency Level (CL) – says how many nodes must agree before an operation is a success. Set at read or write operation.• ONE – coordinator will wait for one node to ack write (also TWO, THREE). One is

default if none provided.• QUORUM – we saw that before. N / 2 + 1. LOCAL_QUORUM, EACH_QUORUM• ANY – waits for some replicate. If all down, still succeeds. Only for writes.

Doesn’t guarantee it can be read.• ALL– Blocks waiting for all replicas

28


ENSURING CONSISTENCY

3 important concepts:

Read Repair - At time of read, inconsistencies are noticed between nodes and replicas are updated. Direct and background. Direct is determined by CL.

Anti-Entropy Node Repair - For data that is not read frequently, or to update data on a node that has been down for a while, the nodetool repair process (also called anti-entropy repair). Builds Merkle trees, compares nodes and does repair.

Hinted Handoff - Writes are always sent to all replicas for the specified row regardless of the consistency level specified by the client. If a node happens to be down at the time of write, its corresponding replicas will save hints about the missed writes, and then handoff the affected rows once the node comes back online. This notification happens is via Gossip. Default 1 hour.

29


ENTER VNODES

• Virtual Nodes introduced with Cassandra 1.2

• Allow spreading of a node across physical servers

• Why• Distribution of load• No token management• Concurrent streaming across hosts

• Two ways to specify in cassandra.yaml• initial_tokens: <token1>,<token2>,<token3>,<token4>, …..

Or num_tokens: 256 # recommended as a starter

30


VNODES - HOW DOES IT LOOK

31


STORAGE

Commit Log – Cassandra appends to the commit log first. Uses fsync in the background to force writing of these changes. Durability, can redo writes in case of crash.

Memtable – in memory cache stored by key. After appending to commit log, writes to Memtable. Then write is considered successful. Each CF has different Memtable.

SSTables – Immutable files. When memtable runs out of space or hits a defined key limit it writes out to SSTable asynchronously. When they reach threshold, compaction (Minor) takes place they are merged (uses a structure called Merkle Tree for this).

Bloom Filter – for each SSTable, checks here first if key exists.

32


C* PERSISTENCE MODEL

33


WRITE PATH

1. Determine all applicable replica nodes across all DC’s

2. Coordinator Node sends to all replicas in the local DC

3. Coordinator sends to ONE replica in remote DC

4. Selected remote replica then sends on to other remote replica nodes

5. All respond back to coordinator

• The CL is how long the coordinator blocks before it returns success or failure to client. (remember ANY, ONE, TWO, LOCAL_QUORUM etc.)

• If a replica node is down (or does not respond), HINTED HANDOFF kicks in. The hint is target replica id + mutation data. • Can configure max hint time and after hints no longer stored.• Hinted handoff runs every 10 minutes and tries to send hints to

nodes.

34


READ PATH

• Consistency level determines how many nodes must have data

• Read request goes to a single node, Storage Proxy, it determines nodes that hold replicas of the data, uses Snitch to determine “proximity”.

• CL of ONE - first find it returns

• CL of Quorum – wait for majority. Better guarantee to get most recent.

• After first read, digest calculated of data and used to determine whether nodes not consistent. If not, in background Read Repair is activated.

• Different SSTables may have different columns so need to merge results.

Reads involve checking Bloom Filter to see if SSTable has key. Key cache to get direct access to data and row cache. Very dependent on

configuration.

35


DELETIONS

Distributed systems present unique problem for deletes. If it actually deleted data and a node was down and didn’t receive the delete notice it would try and create record when came back online. So…

C* uses the concept of a Tombstone. The data is replaced with a special value called a Tombstone, works within distributed architecture

Tombstones are cleaned up after defined time (GCGraceSeconds, default 10 days) during compaction.

36


THRIFT TO CQL

• Thrift - original RPC based API and is still fully supported. Used by clients such as Hector, Pycassa, etc. • Interact with Cassandra using the cassandra-cli.

• CQL3 – new and fully supported API. Table orientated, schema driven query language. It resembled (very closely at times) SQL.• Interact with Cassandra using cqlsh.• Cassandra Table – defined as a sparse collection of well known

and ordered columns.• Still uses same underlying storage as Thrift. • More intuitive data modeling for many.

37


APPLICATION DEVELOPMENT

• Interaction with Cassandra can be done using one of supplied clients such as CLI or CQL. Otherwise client applications are built using a language client library.

• Many clients in multiple languages. Including Java, .NET, Python, Scala, Go, PHP, Node.js, Perl, Ruby, etc.• Java:• Hector wraps the underlying Thrift API. Hector is one of the most

commonly used client libraries. • Astyanax is a client library developed by Netflix .• Datastax CQL – newest CQL driver, will be very familiar to JDBC

developers• And many more … (JPA)

• Also exists Datastax OPSCenter and other various GUI’s and REST API (Virgil)

38


HECTOR EXAMPLE

39


DATASTAX CQL DRIVER

From version 1.2. onward. Uses new CQL method for C* interaction.

40


INSTALLATION

You have 3 options

• Cassandra Cluster Manager – installs multi node cluster automatically with just a few parameters. Great for dev and testing (and VERY cool).• https://github.com/pcmanus/ccm

• Datastax Community Edition – Ease of installation and you get free version of OpsCenter. Has a Windows installer (msi) if using Windows.• http://planetcassandra.org/Download/DataStaxCommunityEdition **

• Open Source Version – from Cassandra, most recent versions, RC’s etc.• http://cassandra.apache.org/download/.

• CQL Driver - http://www.datastax.com/download/clientdrivers

** Planet Cassandra - Great source of Cassandra info!!

41

https://github.com/pcmanus/ccm

http://planetcassandra.org/Download/DataStaxCommunityEdition

http://planetcassandra.org/Download/DataStaxCommunityEdition

http://cassandra.apache.org/download/

http://cassandra.apache.org/download/

http://www.datastax.com/download/clientdrivers


DATA MODELING• NoSQL you don’t need to model your data in Cassandra. Or do you?• Flexible schema = no (or little) data modeling done.• Proved to be problematic from application and query perspective.

• Cassandra is extremely good at reading data in order it was stored.

• CQL makes data modeling much easier, nice bridge from SQL world if that knowledge is there.

• Typically in modeling in C*, you will denormalize data to use the storage engine order.

• Essential is to create a good data model is understanding of queries that will be used.

• Remember no joins, no enforced foreign keys. The app (client) heavily influences the model and we don’t think in terms of “normal form”.

42


OUR USE CASES

Our use case involves health care data and comparison data.

General requirement is we have 3 files.• One summary by DRG (diagnosis related group) codes• One broken down by state• One by practitioner.

• We want to be able to store data (entities), perform performant queries by state, DRG code and provider to get comparisons.

• Let’s model the data, have an app load and look at the decisions we will make.• Will use CQL3, no compact storage for backward compatability.• Will use Datastax CQL Driver• OpenCSV for reading files.

43


SAMPLE DATA – DIAGNOSIS RELATED GROUP

DRG Definition Provider Id Provider Name Provider Street Address Provider City039 - EXTRACRANIAL PROCEDURES W/O CC/MCC 10001 SOUTHEAST ALABAMA MEDICAL CENTER 1108 ROSS CLARK CIRCLE DOTHAN

44

State Zip Total Discharges Average Covered Charges Average Total Payments AL 36301 91 32963.07692 5777.241758


RDBMS ENTITY MODEL

45


CASSANDRA DATA MODEL

• Tables (Column Family) for complete data storage

• Index tables with compound keys for query

• Application will handle required joining and foreign keys etc.

• Cassandra handles the quick writes and other important matters (replication and availability).

46


STANDARD ENTITY TABLES

• Used to store all the data for the different inputs

• Can be used for lookups, but remember don’t be afraid to denormalize.

• Following entities• Summary• State• Provider

47


MATERIALIZED VIEWS

• These are optimized for queries.

• Composite columns for utilizing storage order.

• Denormalized for accessing required info

• The index tables make sense initially• Drg codes• Providers• State

• Can use secondary indexes to add in crtieria, e.g. state

48


WHAT ABOUT TIME SERIES?

• Cassandra is highly performant at storing time series (event) data

• First data modeling pattern utilizes single DRG Code per row, the DRG Code is the partition key and they timestamp is the column.• Works well as again we are using C* built-in sorting

• Second pattern, can be used if too many columns so limit row size• Use row partitioning by adding date portion into row key. So all

data for one day is in one row.• Gets interesting when you use TTL, which gives you “automatic”

data expiring!

49


COLLECTIONS

• Can have set, list or map• Set – return sorted when queried• List – when sorted order is not natural order. Also, value can be in

multiple times• Map – key / value as name implies

• Inserts, updates and deletes all allowed• Syntax takes some practice, + and -

• Can expire each element setting a TTL• Each element is stored as a column

50


COUNTERS

• Distributed counters present another problem.

• C* has a counter type

• One PK and a counter column all that is allowed

• Update right away, no inserts

• Can only increment or decrement.

• Test thoroughly under load.

51


SUMMARY

C* Provider highly available, distributed, DC aware DB with tuneable consistency out of the box.

A lot of tools at your disposal.

Work close with ops or devops .

Test, test and test again.

Don’t be afraid to use the C* community.

Thank you!

52

Cassandra Deep Diver & Data Modeling

Technology

r c h i t e c t u r

training c

reward c

long time c

basics acid cap theorem

consistency architecture

cassandra cassandra

tunable consistency