couchbase live europe 2015: couchbase in amadeus critical systems

28
Couchbase in Amadeus critical systems © 2014 Amadeus IT Group SA Dietmar Fauser Amadeus VP R&D Architecture, Quality and Governance Didier Spezia Senior Expert Database London, March 2015

Upload: couchbase

Post on 16-Jul-2015

314 views

Category:

Software


13 download

TRANSCRIPT

Change the Year in the Copyright field1. 2. 3.

4.

Couchbase in Amadeus critical systems

© 2

014 A

madeus I

T G

roup S

A

Dietmar Fauser

Amadeus VP R&D

Architecture, Quality and Governance

Didier Spezia

Senior Expert Database

London, March 2015

Change the Year in the Copyright field1. 2. 3.

4.

Page 2

Agenda

© 2

014 A

madeus I

T G

roup S

A

1. Introduction

2. Shopping platform, low latency stores

3. R-Box, distribution and ultra-high scalability

4. Hotel shopping and booking, ultra-high availability

5. Conclusion

Change the Year in the Copyright field1. 2. 3.

4.

Introduction

© 2

014 A

madeus I

T G

roup S

A

© 2

014 A

madeus I

T G

roup S

A

Amadeus, the big picture

Corporations

Online agencies

Airlines

Hotels

Cruise lines / Ferries

Tour Operators

Car rental companies

Rail

Travel providers

Global Distribution

System

IT solutions

Travel agents

AirlinesDirect distribution

Airports,

Airline sale offices,

Call centres…

Travel buyers

Leisure

Traditional agencies

Business

Amadeus is present across the entire journey

Travel booking solutions for travel distributors:

• Travel agencies, traditional and online

• Airlines (direct distribution)

Offer management and optimization toolsfor:

• Airlines: inventory, departure control

• Travel agencies: management systems

• Hotels : inventory, booking

• Railways : inventory

• Companies : business travel management

Change the Year in the Copyright field1. 2. 3.

4.

Page 6

© 2

014 A

madeus I

T G

roup S

A

Amadeus, airlines partners

121 airlines contracted for

Reservation and Inventory

800 millions passengers boarded each year by 2015(a)

+93 airlines

(List of airlines shown is a sample and not exhaustive. Passengers generated by airlines that are Altéa customers with Altéa Inventory and/or Departure

Control modules implemented or have contracted the migration to any of these two modules. Estimated annual PBs calculated as follows: (i) annual

PBs migrated to Altéa by 31 December 2010 plus (ii) annual PBs estimated to be migrated to Altéa by 31 December 2013 based on contracts signed

through 31 December 2010; all annual PB volume data based on actual 2010 traffic volumes where migrated before 1 January 2011, remainder

estimated using historic traffic volumes and therefore exclude any volume trend assumptions

(a) 2015 estimated annual PB calculated by applying the IATA’s regional air traffic growth projections to the latest available annual PB figures to the 121

airlines contracted to Altéa as of 31 December 2012

Change the Year in the Copyright field1. 2. 3.

4.

Page 7

© 2

014 A

madeus I

T G

roup S

A

Amadeus, Operational Figures

_ Handles up to 30,000 customer

transactions/sec

_ <0.5 sec response time

_ 16 Petabytes of storage

_ 3.7 million net bookings/day

_ 1.6 billion transactions/day

_ 500,000+ database disk IO/second

_ 4,800+ IT changes/month

Change the Year in the Copyright field1. 2. 3.

4.

Page 8

© 2

014 A

madeus I

T G

roup S

A

Couchbase & Amadeus

_ 2013

• Signed Premium Partnership

• Prototyped 2 use cases with Couchbase (air shopping/inventory)

• Low latency, high-throughput key/value stores

_ 2014

• Couchbase went live for Air Shopping

• Implementing a new use case (hotel shopping & booking)

• Ultra-high availability

_ 2015

_ Air Shopping ramp-up

_ Couchbase will go live for Air Inventory

_ Couchbase production readiness for hotel shopping & booking

_ Implementing a new use case (R-Box)

_ Ultra-high scalability

Change the Year in the Copyright field1. 2. 3.

4.

Shopping Platform

High-throughput, low latency,

key/value stores

2

© 2

014 A

madeus I

T G

roup S

A

Change the Year in the Copyright field1. 2. 3.

4.

Page 10

© 2

014 A

madeus I

T G

roup S

A

The Shopping Flow

Priced solutions

End user

Amadeus

1 Travel request

2 Route Explorer 3 Connection Builder 5 Faring

6

1A Agencies Airline.com Self-serviceCall Centers B2T / B2BAgent.com

Flight selection

4 Flt Validation

Availability calculation

Change the Year in the Copyright field1. 2. 3.

4.

Page 11

© 2

014 A

madeus I

T G

roup S

A

Availability Cache Architecture

Availability Data Services

Miss: read Database

Memcached Servers, 12 servers * 1TB

Miss: Populate Memcached

Availability Services (600 servers)

MySQL Farm, 7 * 4 servers

Shard 1 Shard 2 Shard 3 Shard 4 Shard 5 Shard 6 Shard 7

Change the Year in the Copyright field1. 2. 3.

4.

Page 12

_ Online rebalancing• Extending capacity currently takes careful operations, many

days of preparation and many days to warm up the new cache

_ Cold starts• Memcached (very rare) outages are (very) disruptive

_ Operational costs• Leveraging local storage with PCI-E flash cards• The current infrastructure is complex and requires lot of effort

for day to day operations

_ The storage logic implemented in the application layer is complex

© 2

014 A

madeus I

T G

roup S

A

Problems we are trying to solve

Change the Year in the Copyright field1. 2. 3.

4.

Page 13

© 2

014 A

madeus I

T G

roup S

A

Availability Cache New Architecture

Availability Data Services

Couchbase Servers, 30 servers * 1TB

Availability Services (600 servers)

Change the Year in the Copyright field1. 2. 3.

4.

Page 14

© 2

014 A

madeus I

T G

roup S

A

Availability Data Workloads

GET 2.6 M/s

SET 1 M/s

Change the Year in the Copyright field1. 2. 3.

4.

Page 15

_Couchbase is fast

• Microseconds adds up quickly when exploring thousands of travel solutions

• Our cache queries returns in less than 0.5ms

_Couchbase is predictably fast

• Long tail latencies would kill our application response time

• Couchbase leverages our memory capacity

_Couchbase online rebalancing works without impacting our query latency

© 2

014 A

madeus I

T G

roup S

A

Couchbase benefits

Change the Year in the Copyright field1. 2. 3.

4.

Page 16

_Ability to read on replica (2.1.0)

• In failover scenarios, we prefer to be non consistent rather than a suffering from a few seconds brownout

_Multithreaded read/write operations (2.1.0)

• Our workload is write intensive

• We want to get the most of our Fusion IO PCI-E SSDs

_Rack awareness (2.5.0)

• Our data centre has 6 firecells, we can lose one of them without losing any data

© 2

014 A

madeus I

T G

roup S

A

We asked for it, we got it

Change the Year in the Copyright field1. 2. 3.

4.

Page 17

(benchmarked on a subset of 15% of our data)

© 2

014 A

madeus I

T G

roup S

A

Couchbase 3.0.1

_ Improvement (compared to 2.5.1)

• Automatic, optimized resource utilization

_Consequences

• Same volume processed with less writing threads

• Increase of the write throughput

+ 30% write throughput

- 20% CPU consumption on I/O threads

Change the Year in the Copyright field1. 2. 3.

4.

R-Box

Distribution

Ultra-high scalability

3

© 2

014 A

madeus I

T G

roup S

A

Change the Year in the Copyright field1. 2. 3.

4.

Page 19

R-Box: an extension of the shopping platform

Projecting availability calculations in remote data centres

Change the Year in the Copyright field1. 2. 3.

4.

Page 20

What’s in the box?

265ced1609a17cf1

a5979880a2ad364653895ae8

R-Box

Shell Layer

Core Layer

Data Synchronization

Live Data(Flight, Counters, Yields…)

Data Flow Log/Stat Flow Polling Flow

APP

Data Synchronization

Caches Data(ABR, RFD, Codeshare)

File Distribution(bittorrent)

Log &Stat Flow

Log Server

Administration Tool

AdminAPI

(XML)

Admin Server

Polling Request(XML, EDI)

ESBCouchbase NoSQL

APP APP APP APP APP

Actually, a mini data centre

Administration and management

Storage for live data: Couchbase

Storage for static data

Middleware components (ESB)

Application servers

Monitoring and logs

Change the Year in the Copyright field1. 2. 3.

4.

Hotel shopping and booking

Ultra-high availability

4

© 2

014 A

madeus I

T G

roup S

A

Change the Year in the Copyright field1. 2. 3.

4.

Page 22

_ Largest hotel group worldwide

• A bit more than 4,700 properties

• About 680,000 rooms

• Occupancy rate: 60-70%

_Expected traffic: 80 M bookings/year

_Amadeus to provide

• Shopping services – 10,000 TPS

• Booking

_Strict SLAs

Intercontinental Hotels Group (IHG)

Signed in March 2014

Change the Year in the Copyright field1. 2. 3.

4.

Page 23

(for mission critical and high-throughput applications)

© 2

014 A

madeus I

T G

roup S

A

Ultra-high availability

_General trend of the market

_Our target:

_Complex infrastructure

• Network, storage, databases, application servers, etc …

_Not possible to achieve using conventional database engines

99.999 %99.95 %

Change the Year in the Copyright field1. 2. 3.

4.

Page 24

© 2

014 A

madeus I

T G

roup S

A

Technical requirements

. . . . . . .

Service Isolation Zone

Computing Silo

Compute Silo

Computing Silo

.

.

.

ESB

.

.

.

Storage

Global

DC

. . . . . . .

Load Balancer

.

.

.

Service Isolation Zone

Computing Silo

Computing Silo

Computing Silo

.

.

.

ESB

StorageGlobal

Service Isolation Zone

Computing Silo

Computing Silo

Computing Silo

ESB

Storage

_ Document store for booking

_ Key/value store for shopping

_ Master/master replication between DC

_ Full consistency for 1 DC

_ Eventual consistency between DC

_ Contain shopping traffic in isolation zone

_ Redundant clusterware

Change the Year in the Copyright field1. 2. 3.

4.

Page 25

Cross data centre replication, allowing for creative topologies

© 2

014 A

madeus I

T G

roup S

A

Couchbase XDCR ring topology

Primary CB

Secondary CB

Shopping Shopping Shopping

Primary CB

Secondary CB

Shopping Shopping Shopping

Data Centre 1 Data Centre 2

Change the Year in the Copyright field1. 2. 3.

4.

Conclusion5

© 2

014 A

madeus I

T G

roup S

A

Change the Year in the Copyright field1. 2. 3.

4.

Page 27

Couchbase delivers on what Amadeus values:

_ Scalability

• Horizontal

• Vertical

_ Predictability of raw performance

_ Operability

_ Complex distributed topologies (XDCR)

© 2

014 A

madeus I

T G

roup S

A

Key takeaways

Change the Year in the Copyright field1. 2. 3.

4.

Thank you

© 2

014 A

madeus I

T G

roup S

A

You can follow us on:

AmadeusITGroup

amadeus.com/blogamadeus.com