Use Your MySQL Knowledge to Become an Instant Cassandra Guru

Use Your MySQL
Knowledge to Become an
Instant Cassandra Guru
Percona Live Santa Clara 2014
Robert Hodges
CEO
Continuent
Tim Callaghan
VP/Engineering
Tokutek

Who are we?
Robert Hodges
●  CEO at Continuent
●  Database nerd since 1982 starting with M204, RDBMS since
1990, NoSQL since 2012; designed Continuent Tungsten
●  Continuent offers clustering and replication for MySQL and
other fine DBMS types
Tim Callaghan
●  VP/Engineering at Tokutek
●  Long time database consumer (Oracle) and producer
(VoltDB, Tokutek)
●  Tokutek offers Fractal Tree indexes in MySQL (TokuDB) and
MongoDB (TokuMX)

Cassandra!
Cassandra, used by NetFlix, eBay, Twitter, Reddit and
many others, is one of today's most popular NoSQL-
databases in use. According to the website, the largest
known Cassandra setup involves over 300 TB of data
on over 400 machines.
High Performance Reads and Writes
Linear Scalability
High Availability

One Good Thing about Cassandra
●  Cassandra makes it easy to scale capacity
Existing cluster nodes
running out of space
Start new nodes and let them join
Stored data
redistribute
automatically

One Bad Thing about Cassandra
cql> create table foo (
id int primary key,
customerId int,
orderId int,
orderValue double);
OK!
<Me: I think I’m gonna like Cassandra.>
cql> create index idx_foo_1 on foo (customerId, orderId);
ERROR! - Secondary indexes only support 1 column
<Me: I think I just changed my mind. CQL != SQL>
<Me: (much later) secondary indexes aren’t that useful>

Today’s Question
How can you use your MySQL
knowledge to get up to speed
on Cassandra?

What is CQL?
●  Cassandra originally used a Thrift RPC-based API
●  CQL was added in 0.8
○  Simplifies access
○  Smaller learning curve for SQL users
●  So, you’ll feel right at home
○  Create table
○  Data types
○  Insert, update, select, delete
●  But Cassandra isn’t pretending to be relational...

What ISN’T CQL?
●  It’s familiar, and then it’s not!
○  No joins
○  No foreign keys
○  No “not null”
○  No sum(), group(), min(), …
○  Some “ORDER BY” support
○  Single row UPDATE
○  Limited secondary indexing
○  INSERT == UPDATE
■  but it all behaves like REPLACE INTO

The Bottom Line on CQL
●  It’s just a language
●  It’s very similar to SQL
●  Don’t blame CQL
○  What first appears as a “limitation” in Cassandra can
also be a strength
○  CQL enables us to get started quickly
○  Try Cassandra 0.7 for a little while…
Commit this to memory...
“You do not just throw data into Cassandra and
add later indexes to make it fast.”

Schema Design
“Easy to learn and
difficult to master”
- Nolan Bushnell (founder of Atari and Chuck E. Cheese’s Pizza-Time)

Coming from a Relational World?
Tradeoffs are hard
Feature RDBMS Cassandra
Single Point of Failure
Cross Datacenter
Linear Scaling
Data Modeling
*source = Patrick McFadin (@PatrickMcMcFadin)

How is My Data “Organized”?
Relational Model Cassandra Model
Database Keyspace
Table Column Family
Primary Key Row Key
Column Name Column Name/Key
Column Value Column Value

What is a BigTable?
http://static.googleusercontent.com/media/research.google.com/en/us/archive/bigtable-osdi06.pdf
Row
Key
Column Name
Column Value
Timestamp
TTL
Column Name
Column Value
Timestamp
TTL
Column Name
Column Value
Timestamp
TTL
Column Name
Column Value
Timestamp
TTL
Up to 2 billion!
●  Cassandra uses it for the data model
●  Supports extremely wide rows
●  Row lookup is fast and easily distributed
●  Columns are sorted by name

What About Data Types?
●  Unlike some other NoSQL databases, types
are included via CQL
●  Usual suspects
○  ascii, bigint, blob, boolean, decimal, varchar, …
●  More interesting
○  uuid - global uniqueness
○  timeuuid - global uniqueness, sorted by time portion
○  inet - IPv4 or IPv6 address
○  varint - variable precision integer

How Do I Create a Static Table?
CREATE TABLE users (
user varchar,
email varchar,
state varchar,
PRIMARY KEY (user)
);
●  Primary key is hashed for location/placement (more on that later)
○  Meaning you cannot range scan by PK
●  No “not null” or varchar sizing (enforce in applications)
●  Reads (on username) and all writes are easily distributed
●  Schema flexibility without downtime
○  alter table users add column password varchar;
user
email
timestamp
ttl
state
timestamp
ttl
These rows look and feel familiar.

No Auto-Increment?
●  Auto-increment is extremely difficult in a distributed system
●  Use a natural primary key, if possible
○  Remember, small primary keys aren’t important when denormalizing
●  Or, use uuid / timeuuid
○  Generated in your client applications
CREATE TABLE payments (
id timeuuid,
user varchar,
type varchar,
amount decimal,
PRIMARY KEY (id)
);

What About Secondary Indexes?
CREATE TABLE users (
user varchar,
email varchar,
state varchar,
PRIMARY KEY (user));
-- OPTION 1 : create an index
CREATE INDEX idxUBS on users (state);
-- OPTION 2 : create another table (store data twice)
CREATE TABLE usersByState (
state varchar,
user varchar,
PRIMARY KEY (state, user));

Why Not Create Secondary Indexes?
select * from users where state = ‘CA’;
Secondary index
●  Pro : Index is automatically maintained
●  Con : Above query sent to entire cluster (slows everyone down)
Additional table
●  Pro : Above query is directed to single node (important for scalability)
●  Con : Index is manually maintained (insert data twice, which is OK)
General tips
●  Low selectivity is bad - (male/female)
●  Extremely high selectivity is bad - (unique)
●  In general, create additional table and don’t busy entire cluster for reads

How Do I Create a Dynamic Table?
CREATE TABLE payments (
user varchar,
id timeuuid,
type varchar,
amount decimal,
PRIMARY KEY (user, id)
);
- Compound PK creates wide row
- Remainder of columns are “grouped”
- Still 1 “row” per user
user
type
timestamp
ttl
amount
timestamp
ttl
id1
type
timestamp
ttl
amount
timestamp
ttl
id2
●  Affects how data is stored/organized
●  Still returned in “row” format to CSQL
●  Enables “ORDER BY id” and range query on “id” for “user = ?” queries
●  Remember, BigTables support up to 2 billion columns

What About Relationships?
Relational
table emp (empName text PK, deptId int, ...);
index empIdx1 on emp (deptId);
table dept (deptId int PK, deptName text);
Cassandra
table emp (empName text PK, deptId int, deptName text);
table dept (
deptId int,
deptName text,
empName text,
PRIMARY KEY ((deptId, deptName), empName));
Store department name
with employee
Store employee names with
department (up to ~2 billion)
1, Accounting FrankJones FredJones SamSmith

What About Time Series Data?
CREATE TABLE dashboard (
dashboardId text,
event_time timestamp,
event_value double,
PRIMARY KEY (dashboardId, event_time))
WITH CLUSTERING ORDER BY (event_time DESC);
●  great for “last n”, like dashboards
●  “WITH CLUSTERING ORDER BY ()”
○  Data is stored in given order in BigTable row
○  In this case descending by event_time
○  Easy to access most recent data

How Do I Remove Time Series Data?
CREATE TABLE dashboard (
dashboardId text,
event_time timestamp,
event_value double,
PRIMARY KEY (dashboardId, event_time))
WITH CLUSTERING ORDER BY (event_time DESC);
insert into dashboard (dashboardId, event_time, event_value)
values (25, now(), 500.12) using ttl 86400;
●  defined at the “data” level
●  data “magically” disappears
●  no more cron jobs

Are There Other Cool Schemas?
●  collections : sets, lists, maps
○  an alternative to making the row “wide”
○  set<text> : ordered by CQL Type comparator
○  list<text> : ordered by insertion
○  map<int,text> : unordered
○  support for 64,000 objects per collection (but don’t go crazy)
create table users (
user varchar,
emails set<text>,
PRIMARY KEY (user));
insert into users (user, emails) values (‘tmcallaghan’,
{‘tim@tokutek.com’,’timc@tokutek.com’});

MySQL Transactions and Isolation
mysql> BEGIN;
...
mysql> INSERT INTO sample(data) VALUES
(“Hello world!”);
mysql> INSERT INTO sample(data) VALUES
(“Goodbye world!”);
...
mysql> COMMIT;
InnoDB creates MVCC view
of data; locks updated
rows, commits atomically
MyISAM locks table
and commits each row
immediately

Does Cassandra Have Transactions?
“Transactions” includes a lot of things so…
No:
●  Updates on different rows are separate, like MyISAM
●  Failed transactions on replicas might create partial
writes (Cassandra does not guarantee clean-up)
Yes:
●  Updates to single rows are atomic and isolated
●  Updates to rows are durable (logged) as well
(Important: Cassandra rows can be “bigger” than MySQL)

How Does Cassandra Handle Locks?
Cassandra uses timestamps instead
create columnfamily sample(id int primary key, data varchar, n int);
insert into sample(id, data, n) values(1, 'hello', 25);
insert into sample(id, data, n) values(2, 'goodbye', 27);
cqlsh:cbench> update sample set data='goodbye!' where id=2;
cqlsh:cbench> select id, writetime(data),writetime(n) from sample;
id | writetime(data) | writetime(n)
----+------------------+------------------
1 | 1396326015674000 | 1396326015674000 ⇐ Updated together
2 | 1396326144783000 | 1396326027160000 ⇐ Updated separately
Timestamps are the same for columns updated at the same
time

How Does Cassandra Handle Isolation?
●  Row updates are completely isolated until
they are completed
●  Updates can propagate out to replicas at
different times

How Does Cassandra Replication Work?
●  Cassandra is fully multi-master
●  Updates are allowed at any location
●  Updates can happen even when there is a
partition
db1
db3
db2
db4
Client program
issues write
Writes
proxied
to other
instances
Coordinator

How Many Replicas Are There?
The number of replicas and how they are
distributed are properties of keyspaces
CREATE KEYSPACE cbench WITH replication = {
'class': 'SimpleStrategy',
'replication_factor': '3'
};
Keep 3 copies of
data
Strategy class to
distribute replicas

MySQL Partitioning
•  MySQL partitioning breaks a table into <n> tables
o  “PARTITION” is actually a storage engine
•  Tables can be partitioned by hash or range
o  Hash = random distribution
o  Range = user controlled distribution (date range)
•  Helpful in “big data” use-cases
•  Partitions can usually be dropped efficiently
o  Unlike “delete from table1 where timeField <
’12/31/2012’;”

How Does Cassandra Partition Data?
Cassandra splits data by key across hosts
using a “ring” to assign locations
A
B
F
E D
C
G
H
db3
EFGHAB
db4
GHABCD
db2
CDEFGH
db1
ABCDEF
Virtual node D gets
⅛ of hash range
Copy assigned to
host db4 by strategy

Partitioning in Action
Replica placement is based on primary key
insert into sample(id, data) values
(550e8400-e29b-41d4-a716-446655440000,
'hello!');
Primary Key
Run hash function on value, assign to virtual
node, from there to actual hosts

What About Conflicts?
●  Any client can update rows from any location
●  Cassandra resolves conflicts using
timestamps
●  Conflicts are resolved at the cell level
●  The latest timestamp value wins
Id Data Age
352 “bye” 35
UPDATE sample SET data =
‘hello’, age=35 WHERE id=352
‘bye’ WHERE id=352
WINWIN
LOSE

So Is This Like Galera?
●  Galera uses ACID transactions with
optimistic locking
●  It would accept the first transaction and roll
back the second completely
Id Data Age
352 “bye” 35
‘hello’, age=35 WHERE id=352
‘bye’ WHERE id=352
WIN
WIN
LOSE

How Does “Failover” Work?
●  Cassandra does not have failover
●  If a node fails or disappears, others continue
●  Writes or reads may stop if too many
replicas fail
db1 db2
db4
Client program
issues write
Writes
proxied
to other
instances
X
Coordinator

Tunable Consistency
Clients define the level of consistency
cqlsh:cbench> consistency all
Consistency level set to ALL.
cqlsh:cbench> update rtest set k1=4444 where id=3;
Unable to complete request: one or more nodes were unavailable.
cqlsh:cbench> select * from cbench.rtest where id=3;
Unable to complete request: one or more nodes were unavailable.
cqlsh:cbench> consistency quorum
Consistency level set to QUORUM.
cqlsh:cbench> update rtest set k1=4444 where id=3;

What Happens To Failed Writes?
●  Cassandra has several repair mechanisms
for failures
●  Hinted Handoff - The coordinator remembers
the write and replays it when node returns
●  Read Repair - Coordinator for a later read
notices that a value is out of sync
●  Node Repair - Run an external process
(nodetool) to scan for and fix inconsistencies

What Else is There to Learn?
“A lot” would be an understatement, but here
are some topics to consider.
●  You need to “rethink” your data model, so
read up, practice, repeat
●  Data consistency is an application problem
●  Storage/Internals : LSM, Compaction,
Tombstones, Bloom Filters

Summary 1
We liked…
●  CQL + “Tables” made it easy to get started
●  HA
●  Scaling model
Look [out] for…
●  Don’t bother trying to “port” your MySQL application
●  Query patterns are critical, model around them
●  Pay attention to version when reading docs/blogs
○  Cassandra is quickly evolving (0.7, 1.0, 1.2, 2.0, …)

Summary 2
Highly recommended reading
●  Book: “Practical Cassandra: A Developer’s Approach”
●  Datastax has great docs, http://www.datastax.com/docs
●  Presentations
○  http://www.slideshare.net/jaykumarpatel/cassandra-
data-modeling-best-practices
●  Blogs
○  http://planetcassandra.org
○  http://ebaytechblog.com

Questions?
Robert Hodges
CEO, Continuent
robert.hodges@continuent.com
@continuent
Tim Callaghan
VP/Engineering, Tokutek
tim@tokutek.com
@tmcallaghan

Use Your MySQL Knowledge to Become an Instant Cassandra Guru

More Related Content

What's hot (20)

Similar to Use Your MySQL Knowledge to Become an Instant Cassandra Guru (20)

More from Tim Callaghan (12)

Recently uploaded (20)

Use Your MySQL Knowledge to Become an Instant Cassandra Guru