MySQL Cluster Scaling to a Billion Queries

MySQL Cluster
Scaling to Billion Database Queries with MySQL Cluster
Bernd Ocklin
MySQL Cluster Engineering
bernd.ocklin@oracle.com

Servicing the Most Performance-Intensive Workloads

Program Agenda

<Insert Picture Here>
• Databases Are Exciting Again!!!
• Overview of MySQL Cluster
• MySQL Cluster - What’s New
• How is it used?

A converging world ...
Information Banking

Social
Messaging
Networking

Multi-
Gaming
Media

2.1BN USERS
8X DATA GROWTH IN 5 YRS 850M USERS
70+ NEW DOMAINS EVERY 60 SECONDS 20M APPS PER DAY
40% DATA
GROWTH PER YEAR
1 TR VIDEO
PLAYBACKS

$1TR BY 2014
250m TWEETS
PER DAY $700BN IN 2011

5.9BN MOBILE SUBS IN 2011
1 BILLION iOS & ANDROID APPS 370K CALL MINUTES
EVERY 60 SECONDS
DOWNLOADED PER WEEK

Driving new Database Requirements

EXTREME WRITE SCALABILITY REAL TIME USER EXPERIENCE

ROCK SOLID RELIABILITY RAPID SERVICE INNOVATION

No Trade-Offs: Cellular Network

HLR / HSS
Location
Updates

AuC, Call
Routing, Billing

Pre & Post Paid

• Massive volumes of write traffic
• <3ms database response
• Downtime & lost transactions = lost $

Billing, AuC, MySQL Cluster in Action: http://bit.ly/oRI5tF
VLR

No Trade-Offs

Transactional Integrity
Complex REAL TIME USER EXPERIENCE
EXTREME WRITE SCALABILITY Queries
Standards & Skillsets

ROCK SOLID RELIABILITY ELIMNATE BARRIERS TO ENTRY

MySQL Cluster – Users & Applications
Extreme Scalability, Availability and Affordability

• Web
• High volume OLTP
• eCommerce
• User Profile Management
• Session Management & Caching
• Content Management
• On-Line Gaming

• Telecoms
• Subscriber Databases (HLR / HSS)
• Service Delivery Platforms
• VAS: VoIP, IPTV & VoD
• Mobile Content Delivery
• Mobile Payments
• LTE Access http://www.mysql.com/customers/cluster/

Key Benefits
Scaling Auto-sharding + Multi-master
Reads & Writes
Transactional, ACID-compliant relational database

99.999% Shared-nothing design, no Single Point of Failure
Availability
On-Line operations: Scale, Upgrade Schema, etc.

Real-Time High-load, real-time performance
Responsiveness Predictable low latency, bounded access times

Complex, relational queries + Key/Value Access
SQL & NoSQL
APIs MySQL, Memcached, C++, Java, JPA, HTTP / REST

Low TCO, GPL & Commercial editions
Open platform Commodity hardware, management & monitoring tools

Basic architectures
2-tier

Data access
Application Logic
Front-End

Data
Indexes

Basic architectures
3-tier

SQL, JDBC,
Front-End ADO, ...
Application Logic

Data access
(e.g. SQL engine)

Data
Indexes

Basic architectures
4-tier
Front-End

Application
Logic

Data access
(e.g. SQL engine)

Data
Indexes

All services share the same data view

native ClusterJ REST/JSON LDAP memcached SQL, JDBC,
ADO, ...

NDB API

MySQL Cluster Data Nodes

C++ example

    NdbOperation *op = trx>getNdbOperation(myTable);

    op>insertTuple();
    op>equal("key", i);
    op>setValue("value", &value);

    trx>execute( NdbTransaction::Commit );

Java example

Character newCharacter =
session.newInstance(Character.class);

newCharacter.setName(„Yoda“);
newCharacter.setAttributes(„Force“);

Session.persist(newCharacter);

SQL example
(requires MySQL Server)

Mysql> INSERT INTO Charaters (Name, Attributes)
VALUES („Yoda“, „Force“);

High performance and Scalability

Cluster is
• Distributed
• Event Driven
• Asynchronous
• Parallel
• Non-locking

Your friends / Your enemies

•
Disks (life-saver) •
Disks (slow fsync)
•
CPU cache •
Network latency
•
RAM •
Heap allocation
•
Many cores •
NUMA
•
Context switching

Use your friends

Disks (your job saver)
– Log your data to disk (asynchrounsly)
CPU cache
– Align to to it
RAM
– Preallocate!
Many cores
– Distribute to cores (have a model that supports this)

Avoid your enemies

Disks
– Reduce fsyncs
– no swapping
Network latency
– Reduce network round trips
Slow heap allocation
– Pre-allocate all memory, avoid using it
NUMA
– Disable it
Context switching
– Lock to cores
– Get network interrupts out of your way

MySQL Cluster
– A distributed hash table

17 Yoda
143 Albert
12 Bernd
42 Ernest

md5() % <no of nodes>


17 Yoda 12 Bernd
143 Ernest 143 Albert

Best Practice : Primary Keys

• ALWAYS DEFINE A PRIMARY KEY ON THE TABLE!
• A hidden PRIMARY KEY is added if no PK is specified. BUT..
• .. NOT recommended
• The hidden primary key is for example not replicated
(between Clusters)!!
• There are problems in this area, so avoid the problems!
• So always, at least have
id BIGINT AUTO_INCREMENT PRIMARY KEY
• Even if you don't “need” it for you applications

Auto-Sharding (distribution)

Application

[ {id: 12, name: Bernd}, {id: 143, name: Albert},
{id: 42, name: Ernest}, …, {id: 17, name: Yoda}]

{id: 17, … } {id: 143, … } {id: 42, … } {id: 12, … }


Auto-Sharding (distribution)
– Application “knows“ the data location
Application

find({id: 12})

{id: 12, name: Bernd}


Auto-Sharding

• Transparent to the application and data access layer
• No need for application-layer sharding logic – build into the API & kernel
• Partitioning based on hashing all or part of the primary key
• Each node stores primary fragment for 1 partition and back-up fragment for another
• Transparency maintained during failover, upgrades and scale-out
• No need to limit application to single-shard transactions

Adding High Availability
– Introducing Node Groups
Application


{id: 17, … } {id: 143, … } {id: 42, … } {id: 12, … }


- Synchronous Replication
Application


{id: 17, … } {id: 143, … } {id: 42, … } {id: 12, … }


– Synchronous Replication

17 Yoda 12 Bernd
42 Ernest 143 Albert
12 Bernd 17 Yoda
143 Albert 42 Ernest

Handling Scheduled Maintenance
On-Line Operations

• Scale the cluster (add & remove nodes on-line)
• Repartition tables
• Upgrade / patch servers & OS
• Upgrade / patch MySQL Cluster
• Back-Up
• Evolve the schema on-line, in real-time

Adding disk durability

Memory

{id: 17, … } In-memory tables
Data kept in memory but
complemented by logging
to disk.

Disk
Disk based tables
Data kept on disk but
cached in memory.
Logging to disk is
decoupled from
transaction writing.

Shared Nothing

SQL, JDBC,
ADO, ...

No shared components.
Cheap commodity
hardware.
Proper SAN acceptable
but expensive.

– Extreme resilience
Application

Service continuing


Event driven & asynchronous

find(17, {}); {id: 17, name: Albert}

{id: 17} {name: Albert}
TC LQH ACC TUP

Doing things in parallel

find(17, {}); find([12, 17], {});



find(17, {}); {id: 17, name: Albert}

TC LQH {id: 17} {name: Albert}
TC
TC LQH
LQH
TC
TC LQH
LQH ACC TUP
x16 x16


• Primary key reads can be directed to the correct
shard on the API application level
– No waste of resources by doing same operation
on all
• Each data node can handle up to 16 operations in
parallel
• One data node can fully utilize up to 51 physical CPU
cores

Scaleable & Elastic
– Increasing capacity in seconds
Application

Data Nodes

Basic Deployments
1 Data Node

SQL, JDBC,
Front-End ADO, ...
Application Logic
Data Access

Data
Indexes

Basic Deployments
Start scaling with 2 Data Nodes

SQL, JDBC,
Front-End ADO, ...
Application Logic
Data Access
●
Shared Data View
●
Automated Load
Balancing

Data
Indexes

Basic Deployments
Continue Scaling
SQL, JDBC,
ADO, ...
SQL, JDBC,
Front-End ADO, ...
Application Logic OS, App or
Hardware Load
Data access Balancing
(e.g. SQL engine)
Automated
Load Balancing
Data
Indexes

Basic Deployments
Scale on 3 levels
SQL, JDBC,
ADO, ...
Front-End SQL, JDBC,
Application Logic ADO, ...

Data access
(e.g. SQL engine)

Data
Indexes

Typical Deployments
2 Server
Application
SQL, JDBC,
ADO, ...

Web Server

MySQL Server

Data Nodes

Server 1 Server 2

Typical Deployments
4 Server
Application
SQL, JDBC,
ADO, ...
Server 3 Server 4

Web Server

MySQL Server

Data Nodes

Server 1 Server 2

Typical Deployments
Geo Replication
SQL, JDBC,
ADO, ...
Server 3 Server 4 Server 3 Server 4

Server 1 Server 2 Server 1 Server 2

READS
Million / minute

1.200
1.056
1.000

800
• 8 x Commodity Intel Servers
600 • 2 x 6-core processors 2.93GHz
400

200
• x5670 processors (24 threads
0 per total)
2 node 4 node 8 node
• 48GB RAM
UPDATE • Linux
Million / minute

120
109
100 • Infiniband networking
80

60 • flexAsynch benchmark
40
• C++ NoSQL API (NDB API)
20

0
4 node 8 node

Adaptive Query Localization
Scaling Distributed Joins 70x
More
Performance

• Perform Complex Queries
mysqld across Shards
• JOINs pushed down to data nodes
A Data Nodes
• Executed in parallel
Q • Returns single result set to MySQL
L
• Opens Up New Use-Cases
• Real-time analytics
• Recommendations engines
mysqld
• Analyze click-streams

Data Nodes DON’T COMPROMISE
FUNCTIONALITY TO SCALE-OUT !!

MySQL Cluster 7.2 AQL Test Query
Web-Based Content Management System

MySQL
Server

Data Data
Node1 Node2

Copyright 2011 Oracle Corporation 49

Web-Based CMS
70x
More
Performance

87.23 seconds

1.26 seconds

Must Analyze tables for best results
mysql> ANALYZE TABLE <tab-name>;

Memcached Key-Value API
• Persistent, Scalable, HA
Back-End to memcached
• No application changes: re-
uses standard memcached
clients & libraries
• Consolidate Caching &
Database Tiers
• Eliminate cache invalidation
• Simpler re-use of data across
services
• Improved service levels

New • Flexible Deployment
NoSQL • Schema or Schema-less
Access storage

Schema-Free apps
• Rapid application
evolution
• New types of data
constantly added
• No time to get schema
extended
• Missing skills to extend
schema
• Initially roll out to just a
few users
• Constantly adding to live
system


Cluster & Memcached – Schema-Free

key value

<town:maidenhead,SL6>
Application view

SQL view key value


generic table

Cluster & Memcached –
Configured Schema

key value

Application view

SQL view prefix key value


Config tables map.zip

SQL & NoSQL

Clients

JDBC / ODBC
PHP / PERL
Native memcached HTTP/REST Python / Ruby

NDB API

Data Nodes

Mix • SQL: Complex, relational queries
& • Memcached: Key-Value web services
Match • Java: Enterprise Apps
• NDB API: Real-time services

MySQL 5.5 Server Integration
• Configure storage engine per-table
• Choose the right tool for the job
• InnoDB: Foreign Keys, XA Transactions,
Large Rows
• MySQL Cluster: HA, High Write Rates, Real-
Time
• Reduces Complexity, Simplifies
DevOps
• Take advantage of MySQL 5.5
• 3x higher performance
• Improved partitioning, diagnostics, availability,
etc.

Multi-Site Clustering

• Split data nodes across
data centers
• Synchronous replication
Node Group 1 and auto-failover between
Data Node 1 Data Node 2 sites
Synchronous
Synchronous
Replication
Replication
• Improved heartbeating to
handle network partitions
Node Group 2 • Extends HA Options
Data Node 3 Data Node 4 • Active/Active with no
need for conflict
handling

Active/Active Geographic Replication

•Replicating complete
clusters across data
centers
• DR & data locality
• No passive resources
Geographic
Replication •Simplified Active /
Active Replication
• Eliminates requirement
for application & schema
changes
• Transaction-level
rollback

Simplified Provisioning & Maintenance
User Privilege Consolidation

The existence, content and timing of future releases described here is included for information only and may be changed at Oracles discretion.
October 3rd, 2011

MySQL Cluster Manager
Reducing TCO and creating a more agile, highly
available database environment

Automated
Management

Monitoring & High
Recovery Availability
Operation


How Does MySQL Cluster Manager Help?
Example: Initiating upgrade from MySQL Cluster 7.0 to 7.2

Before MySQL Cluster Manager With MySQL Cluster Manager
• 1 x preliminary check of cluster state upgrade cluster --package=7.1 mycluster;
• 8 x ssh commands per server
• 8 x per-process stop commands
• 4 x scp of configuration files (2 x mgmd & 2 x Total: 1 Command -
mysqld)
• 8 x per-process start commands
Unattended Operation
• 8 x checks for started and re-joined processes • Results
• 8 x process completion verifications
• 1 x verify completion of the whole cluster. • Reduces the overhead and complexity
• Excludes manual editing of each configuration of managing database clusters
file. • Reduces the risk of downtime resulting
from administrator error
Total: 46 commands - • Automates best practices in database
2.5 hours of attended operation cluster management


Bootstrap single host Cluster
1. Download MCM from edelivery.oracle.com:
• Package including Cluster
1. Unzip
2. Run agent, define, create & start Cluster!
$> binmcmd –bootstrap
MySQL Cluster Manager 1.1.2 started
Connect to MySQL Cluster Manager by running "D:AndrewDocumentsMySQLmcmbinmcm" -a NOVA:1862
Configuring default cluster 'mycluster'...
Starting default cluster 'mycluster'...
Cluster 'mycluster' started successfully
ndb_mgmd NOVA:1186
ndbd NOVA
ndbd NOVA
mysqld NOVA:3306
mysqld NOVA:3307
ndbapi *
Connect to the database by running "D:AndrewDocumentsMySQLmcmclusterbinmysql" -h NOVA -P 3306
-u root

• Connect to Cluster & start using database
To bootstrap with Cluster 7.2 replace contents of mcm/cluster directory
http://www.clusterdb.com/mysql-cluster/mysql-cluster-manager-1-1-2-creating-a-cluster-is-now-trivial


MySQL Enterprise Monitor 2.3


Evaluate MySQL Cluster CGE
30-Day Trial
• Navigate to
http://edelivery.oracle.com/
and step through (selecting
“MySQL Database” as the
Product Pack)

• Select MySQL Cluster
Manager

Operational Best Practices


When to Consider MySQL Cluster

What are the consequences of downtime or failing to meet
performance requirements?

How much effort and $ is spent in developing and managing HA in
your applications?

Are you considering sharding your database to scale write
performance? How does that impact your application and
developers?

Do your services need to be real-time?

Will your services have unpredictable scalability demands,
especially for writes ?

Do you want the flexibility to manage your data
with more than just SQL ?

Where would I not Use MySQL Cluster?
• “Hot” data sets >3TB
• Replicate cold data to InnoDB
• Long running transactions
• Large rows, without using BLOBs
• Foreign Keys
• Can use triggers to emulate:
• http://dev.mysql.com/tech-resources/articles/mysql-enforcing-foreign-keys.html

• Full table scans
• Savepoints
• Geo-Spatial indexes
• InnoDB storage engine would be the right choice
MySQL Cluster Evaluation Guide
http://mysql.com/why-mysql/white-papers/mysql_cluster_eval_guide.php

MySQL Cluster in Action
Web Reference Architectures
Session Management eCommerce Data Content Management
Refinery Memcache / Application Servers

MySQL Servers MySQL Servers
MySQL Master

Node Group 1 Node Group 2 Node Group 1 Node Group 2

F1 F2 F1 F2 Slave N
F4

Node 3
F3
Node 3

F4
Node 3

F3
Node 3

F1 F2 F1 F2
F3 F4
Node 4

F4
Node 4

F3
Slave 8 Slave 9 Slave 10
Node 4
Node 4

Slave 6 Slave 7
MySQL Cluster Data Nodes MySQL Cluster Data Nodes

Slave 1 Slave 2 Slave 3 Slave 4 Slave 5

• 4 x Data Nodes: 6k Analytics
page hits per second MySQL Master XOR
• Each page hit
generating 8 – 12
database operations

Distributed
Slave 1 Slave 2 Slave 3
Storage

Whitepaper: http://www.mysql.com/why-mysql/white-papers/mysql_wp_high-availability_webrefarchs.php

CUSTOMER PERSPECTIVE
“MySQL Cluster won the performance test hands-
COMPANY OVERVIEW
down, and it fitted our needs perfectly. We
• Leading provider of communications evaluated shared-disk clustered databases, but the
platforms, solutions & services cost would have been at least 10x more.”
• €15.2bn Revenues (2009), 77k employees -- François Leygues, Systems Manager
across 130 countries

CHALLENGES / OPPORTUNITIES
• Converged services driving migration to RESULTS
next generation HLR / HSS systems • Scale out on standard ATCA hardware to
• New IMS platforms for Unified support 60m+ subscribers on a single
Communications platform
• Reduce cost per subscriber and accelerate • Low latency, high throughput with
time to value 99.999%+ availability
• Enabled customers to reduce cost per
subscriber and improve margins
SOLUTIONS • Delivered data management solution at
• MySQL Cluster Carrier Grade Edition 10x less cost than alternatives
• MySQL Support & Consulting Services
http://www.mysql.com/why-mysql/case-studies/mysql-alcatel-casestudy.php
http://www.mysql.com/why-mysql/case-studies/mysql-alcatel-casestudy.php

Shopatron: eCommerce Platform
• Applications
– Ecommerce back-end, user authentication,
order data & fulfilment, payment data &
inventory tracking. Supports several
thousand queries per second

• Key business benefits
– Scale quickly and at low cost to meet
demand
– Self-healing architecture, reducing TCO

• Why MySQL?
– Low cost scalability
– High read and write throughput
– Extreme availability

“Since deploying MySQL Cluster as our eCommerce database, we have had
continuous uptime with linear scalability enabling us to exceed our most stringent SLAs”
— Sean Collier, CIO & COO, Shopatron Inc

http://www.mysql.com/why-mysql/case-studies/mysql_cs_shopatron.php
http://www.mysql.com/why-mysql/case-studies/mysql_cs_shopatron.php 76

COMPANY OVERVIEW CUSTOMER PERSPECTIVE
• Pyro provide comms technology solutions ”MySQL Cluster 7.1 gave us the perfect combination
in Core Network, OSS/BSS & VAS of extreme levels of transaction throughput, low
• Deployed in 120+ networks worldwide latency & carrier-grade availability. We also reduced
• Cell C, one of the largest mobile TCO by being able to scale out on commodity server
operators in South Africa blades and eliminate costly shared storage”
• 560 roaming partners in 186 countries -- Phani Naik, Head of Technology at Pyro Group

• FIFA 2010 world cup opens up network
services to millions of mobile subscribers RESULTS
• International roaming SDP to support up • Supported subscriber and traffic volumes
to 7m roaming subscribers per day • Delivered continuous availability
• Offer local pricing with home network • Implemented in 25% of the time of typical
functionality SDP solutions
• Minimize cost and time to market • Choice in deployment platforms to eliminate
vendor lock-in (migrated from Microsoft)

SOLUTIONS
• MySQL Cluster 7.1 & Services

CUSTOMER PERSPECTIVE
“Telenor has been using MySQL for fixed IP
COMPANY OVERVIEW management since 2003 and are extremely
• Leading telecoms provider across Europe satisfied with its speed, availability and
and Asia. Largest Nordic provider flexibility. Now we also support mobile
• 184m subscribers (Q2, 2010) and LTE IP management with our solution.
Telenor has found MySQL Cluster to be
the best performing database in the world
CHALLENGES / OPPORTUNITIES for our applications.”
• Extend OSS & BSS platforms for new
mobile services and evolution to LTE - Peter Eriksson, Manager, Network Provisioning
• OSS: IP Management & AAA
RESULTS
• BSS: Subscriber Data Management &
Customer Support • Launch new services with no downtime,
due to on-line operations of MySQL
Cluster
• Consolidated database supports
SOLUTIONS Subscriber Data Management initiatives
• MySQL Cluster • MySQL Cluster selected due to 99.999%
• MySQL Support Services availability, real time performance and
linear scalability on commodity hardware

COMPANY OVERVIEW CUSTOMER PERSPECTIVE
• UK-based retail and wholesale ISP & “Since deploying our latest AAA platform, the
Hosting Services MySQL environment has delivered continuous
• 2010 awards for best home broadband uptime, enabling us to exceed our most stringent
and customer service SLAs”
• Acquired by BT in 2007 -- Geoff Mitchell Network Engineer

• Enter market for wholesale services,
demanding more stringent SLAs
• Re-architect AAA systems for data RESULTS
integrity & continuous availability to • Continuous system availability, exceeding
support billing sytems wholesale SLAs
• Consolidate data to for ease of reporting • 2x faster time to market for new services
and operating efficiency
• Agility and scale by separating database
• Fast time to market from applications
• Improved management & infrastructure
efficiency through database consolidation
SOLUTIONS
• MySQL Cluster
• MySQL Server with InnoDB

COMPANY OVERVIEW USER PERSPECTIVE
• Division of Docudesk “MySQL Cluster exceeds our requirements for low
latency, high throughput performance with
• Deliver Document Management SaaS
continuous availability, in a single solution that
minimizes complexity and overall cost.”
CHALLENGES / OPPORTUNITIES -- Casey Brown, Manager of Dev & DBA Services,
Docudesk
• Provide a single repository for customers to
manage, archive, and distribute documents
• Implement scalable, fault tolerant, real time
data management back-end RESULTS
• PHP session state cached for in-service • Successfully deployed document
personalization management solution, eliminating paper
• Store document meta-data, text (as trails from legal processes
BLOBs), ACL, job queues and billing data • Integrate caching and database into one
• Data volumes growing at 2% per day layer, reducing complexity & cost
• Support workload with 50:50 read/write
ratio
SOLUTION • Low latency for real-time user experience
• MySQL Cluster deployed on EC2 and document time-stamping
• Continuous database availability

Getting Started
Learn More
Scaling Web
Databases
Guide

www.mysql.com/cluster/

Evaluate MySQL Cluster 7.2 Bootstrap a Cluster!

Download, No
Download Today Obligation
http://www.mysql.com/ https://edelivery.oracl
downloads/cluster/ e.com/


Summary

Scale Web Services with
Carrier-Grade Availability

Don’t Trade Functionality for Scale

Try it out Today!


Multi-threaded Data Node Extensions

• Scaling out on commodity
hardware is the standard
Application Nodes way to increase
performance
• Add more data nodes and

Node 3
API nodes as required
• MySQL Cluster 7.2
Node 1

increases the ability to also
scale-up each data node
• Increases maximum
number of utilised threads
from 8 to 59
Node 2

Node 4

• Can deliver aX single
thread performance with
bX cores
Node Group 1 Node Group 2


• Threads (post GA!):
• recv: <= 8 Receive threads
Application Nodes
• tc: <= 24 Transaction
Coordinator threads
• ldm: <= 16 Local Query
Handler threads
• send: <= 8 Send threads
• main: 1 Main thread
recv send main • rep: 1 Replication thread
• io: 1 I/O thread
• Engineering guidelines
provided to find the best
tc ldm rep io configuration: ZXZX

ThreadConfig :=<entry> [ ,<entry> ] +

entry :=<type>={ [<param> ]+ }
• Note that extra send,
param := count = N |
cpubind = L |
recv & tc threads
cpuset = L will be part of post-
type := ldm | main | recv | rep | GA maintenance
maint | send | tc | io release.
Example:
ThreadConfig=ldm={count=2,cpubind=1,2},
ldm={count=2,cpuset=6-9},
main={cpubind=12},rep={cpubind=11}

NoSQL with Memcached

• Flexible: set maidenhead 0 0 3
SL6
• Deployment options
STORED
• Multiple Clusters
• Simultaneous SQL Access get maidenhead
• Can still cache in VALUE maidenhead 0 3
Memcached server SL6
• Flat key-value store or map END
to multiple tables/columns

Multi-Site Clustering – changes to
STONITH algorithm
• When heartbeat not received, all data nodes will be asked to
ping all other data nodes
• Each node establishes its list of ‘suspect’ data nodes from whom
they don’t receive a ping response within
ConnectCheckIntervalDelay msecs
• If second period of ConnectCheckIntervalDelay passes
without a ping response then each data node will send a Fail
report to all data nodes naming its suspected node(s)
• On receipt of a Fail message from a suspect node, the receiving
node will consider the originating node as failed rather than the
requested target
• Leaves each side of the temporarily partitioned network with a
viable set of data nodes and arbitration is used to select the
surviving side if there is no longer a clear majority

Multi-Site Clustering – WAN engineering
recommendations based on user experience

• (Obviously) the longer the latency between sites, the
higher the impact to performance
• Target latency should be <= 10 ms; 20 ms
acceptable
• Test with 1000 byte packet, under load
• Bandwidth requirements dependent on traffic but aim
for 1 Gbps+ (100 Mbps for low traffic Cluster)
• Simplest WAN topology possible (fewer points of
failure/failover latency)
• Typical WAN failover times should be short enough
not to trigger STONITH in Cluster

Active-Active Replication - Detecting Conflicts

Geographic Replication – what’s
changed in conflict resolution
• Reflecting GCI (Global Checkpoint Index) removes requirement for
applications to maintain timestamp field in each potentially conflicting
table
• One of the two masters acts as the ‘primary’ and monitors all received
replication events from the ‘secondary’ (including its own ‘reflected GCI’) to
establish when changes not applied in same order on primary and secondary
Clusters
• Primary will then overwrite all conflicting transactions (or optionally just the
conflicting rows) on the secondary – as well as subsequent transactions
influenced by the conflict
• To use, set the function in mysql.ndb_replication to NDB$EPOCH()
or NDB$EPOCH_TRANS()
• Overview & worked example: http://bit.ly/activeactive
• Gory details: http://bit.ly/refcgci

How to Push Privilege Data into Data
Nodes
mysql> SOURCE /usr/local/mysql/share/mysql/ndb_dist_priv.sql;
mysql> CALL mysql.mysql_cluster_move_privileges();

mysql> SHOW CREATE TABLE mysql.userG
*************************** 1. row ***************************
Table: userCreate Table: CREATE TABLE `user` (
`Host` char(60) COLLATE utf8_bin NOT NULL DEFAULT '',
....
....
) ENGINE=ndbcluster DEFAULT CHARSET=utf8 COLLATE=utf8_bin
COMMENT='Users and global privileges‘

•Fully worked example:
http://www.clusterdb.com/mysql-cluster/sharing-user-credential
(http://bit.ly/userpriv)

Scale the Cluster
Capacity & Performance

Application Nodes

Data Node 1 Data Node2

Cluster F1 F2
Mgmt F2 F1

Data Nodes

Scale the Cluster
Capacity & Performance

Application Nodes

Node Group 1 Node Group 2
Node 1

Node 3
Cluster F1 F2 Cluster
Mgmt F3 F4 Mgmt
Node 2

Node 4
F3 F4
F1 F2
Data Nodes

Node Failure Detection & Self-Healing Recovery

On-Line Scaling & Maintenance
1. New node group added
2. Data is re-partitioned
3. Redundant data is deleted
4. Distribution is switched to share
load with new node group

• Can also update schema on-
line
• Upgrade hardware &
software with no downtime
• Perform back-ups on-line

Only MySQL Can…..

blend the agility & innovation of the web….

….with the trust & capability of the network.

No Trade-Offs: eCommerce
• Integrated Service Provider
platform
• eCommerce
• Payment processing
• Fulfillment
• Supports 1k+ manufacturers &
18k retail partners
• Requirements
• Scaling, On-Demand
• HA: failures & on-line upgrades
• High batch & real time loads
• Low TCO: capex and opex
http://mysql.com/customers/view/?id=1080

No Trade-Offs: Flight Control
• US Navy aircraft carriers
• Consolidated flight operations
management system
• Maintenance records
• Fuel loads
• Weather conditions
• Flight deck plans
• Requirements
• No Single Points of Failure
• Complete redundancy
• Small footprint, harsh environment
• 4 x MySQL Cluster nodes,
Linux and Windows

MySQL User Conference Session: http://bit.ly/ogeid3

Creating & running your first Cluster
- the “manual” way (without MCM)

• Up & running in 10-15 minutes using Quick Start guides from
http://dev.mysql.com/downloads/cluster/
• Versions for Linux, Windows & Solaris


ACID Compliant Relational Database
• SQL & NoSQL interfaces

Write-Scalable & Real-Time
• Distributed, multi-master, auto-sharding, optimized in-memory structures & indices

99.999% Availability
• Shared-nothing, integrated clustering & sub-second recovery, local & geographic
replication, on-line operations

Low Barriers to Entry
• Open-source, elastic, multiple APIs, management tools, commodity hardware

MySQL Cluster Scaling to a Billion Queries

Recomendados

Recomendados

Mais conteúdo relacionado

Mais procurados

Mais procurados (20)

Destaque

Destaque (12)

Semelhante a MySQL Cluster Scaling to a Billion Queries

Semelhante a MySQL Cluster Scaling to a Billion Queries (20)

Último

Último (20)

MySQL Cluster Scaling to a Billion Queries