How Hailo fuels its growth using NoSQL storage and analytics - Dave Gardner (Hailo)

Preview:

DESCRIPTION

Presented at JAX London 2013 Hailo, the taxi app, has served more than 5 million passengers in 15 cities and has taken fares of $100 million this year. I'm going to talk about how that rapid growth has been powered by a platform based on Cassandra and operational analytics and insights powered by Acunu Analytics. I'll cover some challenges and lessons learned from scaling fast!

Citation preview

Hailo and NoSQL

David Gardner, Architect at Hailo

JAXLONDON2013

JAXLONDON2013

JAXLONDON2013

1.  Why choose NoSQL

2.  A whistle-stop tour of Cassandra

3.  Adoption of Cassandra at Hailo

What this talk is about

JAXLONDON2013

What is Hailo?

Hailo is The Taxi Magnet. Use Hailo to get a cab wherever you are, whenever you want.

JAXLONDON2013

JAXLONDON2013

JAXLONDON2013

JAXLONDON2013

•  The world’s highest-rated taxi app – over 11,000 five-star reviews

•  Over 500,000 registered passengers

•  A Hailo hail is accepted around the world every 4 seconds

•  Hailo operates in 15 cities on 3 continents from Tokyo to Toronto in nearly 2 years of operation

Facts and figures

JAXLONDON2013

•  Hailo is a marketplace that facilitates over $100M in run-rate transactions and is making the world a better place for passengers and drivers

•  Hailo has raised over $50M in financing from the world's best investors including Union Square Ventures, Accel, the founder of Skype (via Atomico), Wellington Partners (Spotify), Sir Richard Branson, and our CEO's mother, Janice

Hailo is growing

JAXLONDON2013

Why choose NoSQL?

JAXLONDON2013

“NoSQL DBs trade off traditional features to better support new and emerging use cases”

Andy Gross, Riak

http://www.slideshare.net/argv0/riak-use-cases-dissecting-the-solutions-to-hard-problems

JAXLONDON2013

•  More widely used, tested and documented software

•  Ad-hoc querying

•  Talent pool with direct experience

What are we trading off?

JAXLONDON2013

•  High availability

•  Scalability

•  Operational simplicity

What do we get back in return?

JAXLONDON2013

Cassandra 101

JAXLONDON2013

Consistent hashing Vector clocks * Gossip protocol Hinted handoff Read repair http://www.allthingsdistributed.com/files/amazon-dynamo-sosp2007.pdf

Columnar SSTable storage

Append-only Memtable

Compaction

http://labs.google.com/papers/bigtable-osdi06.pdf

Amazon Dynamo + Google Big Table

JAXLONDON2013

coordinator node

Client

tokens are integers from 0 to 2127

three replicas (RF=3)

JAXLONDON2013

Consistency level (CL)

Level Description

ONE 1st Response

QUORUM N/2 + 1 replicas

LOCAL_QUORUM N/2 + 1 replicas in local data centre

EACH_QUORUM N/2 + 1 replicas in each data centre

ALL All replicas

How many replicas must respond to declare success?

JAXLONDON2013

Big Table

•  Sparse column based data model •  SSTable disk storage •  Append-only commit log •  Memtable (buffer and sort) •  Immutable SSTable files •  Compaction

http://research.google.com/archive/bigtable-osdi06.pdf http://www.slideshare.net/geminimobile/bigtable-4820829

JAXLONDON2013

Plus timestamp, used for Last Write Wins (LWW) conflict resolution

Name

Value

Column

JAXLONDON2013

we can have millions of columns

* theoretically up to 2 billion!

Name

Value

Column

Name

Value

Column

Name

Value

Column

JAXLONDON2013

Name

Value

Column

Name

Value

Column

Name

Value

Column

Row Key

Row

JAXLONDON2013

Column Family

Column Row Key Column Column

we can have billions of rows

Column Row Key Column Column

Column Row Key Column Column

JAXLONDON2013

Write Memtable

SSTable SSTable Commit Log

Memory

Disk

buffers writes and sorts data

flush on time or size trigger

immutable

JAXLONDON2013

Cassandra at Hailo

JAXLONDON2013

Hailo launched in London in November 2011

•  Launched on AWS

•  Two PHP/MySQL web apps plus a Java backend

•  Mostly built by a team of 3 or 4 backend engineers

•  MySQL multi-master for single AZ resilience

JAXLONDON2013

Why Cassandra?

•  A desire for greater resilience – “become a utility” Cassandra is designed for high availability

•  Plans for international expansion around a single consumer app Cassandra is good at global replication

•  Expected growth Cassandra scales linearly for both reads and writes

•  Prior experience I had experience with Cassandra and could recommend it

JAXLONDON2013

The path to adoption

•  Largely unilateral decision by developers – a result of a startup culture

•  Replacement of key consumer app functionality, splitting up the PHP/MySQL web app into a mixture of global PHP/Java services backed by a Cassandra data store

•  Launched into production in September 2012 – originally just powering North American expansion, before gradually switching over Dublin and London

JAXLONDON2013

One year on...

•  Further breakdown of functionality into Go/Java SOA

•  Migrating all online databases to Cassandra

JAXLONDON2013

Development perspective

JAXLONDON2013

“Cassandra just works”

Dom W, Senior Engineer

JAXLONDON2013

Use cases

1.  Entity storage

2.  Time series data

JAXLONDON2013

CF = customers

126007613634425612: createdTimestamp: 1370465412 email: dave@cruft.co givenName: Dave familyName: Gardner locale: en_GB phone: +447911111111

JAXLONDON2013

Considerations for entity storage

•  Do not read the entire entity, update one property and then write back a mutation containing every column

•  Only mutate columns that have been set

•  This avoids read-before-write race conditions

JAXLONDON2013

JAXLONDON2013

CF = stats_db

2013-06-01: 55374fa0-ce2b-11e2-8b8b-0800200c9a66: {“action”:”… a48bd800-ce2b-11e2-8b8b-0800200c9a66: {“action”:”… b0e15850-ce2b-11e2-8b8b-0800200c9a66: {“action”:”… bfac6c80-ce2b-11e2-8b8b-0800200c9a66: {“action”:”…

JAXLONDON2013

CF = stats_db

LON123456: 13b247f0-ce2c-11e2-8b8b-0800200c9a66: {“action”:”… 20f70a40-ce2c-11e2-8b8b-0800200c9a66: {“action”:”… 2b44d3b0-ce2c-11e2-8b8b-0800200c9a66: {“action”:”… 338a22f0-ce2c-11e2-8b8b-0800200c9a66: {“action”:”…

JAXLONDON2013

JAXLONDON2013

Considerations for time series storage

•  Choose row key carefully, since this partitions the records

•  Think about how many records you want in a single row

•  Denormalise on write into many indexes

JAXLONDON2013

Analytics

•  With Cassandra we lost the ability to carry out analytics eg: COUNT, SUM, AVG, GROUP BY

•  We use Acunu Analytics to give us this abilty in real time, for pre-planned query templates

•  It is backed by Cassandra and therefore highly available, resilient and globally distributed

•  Integration is straightforward (HTTP POST)

NSQ Acunu C* events

JAXLONDON2013

JAXLONDON2013

AQL

SELECT SUM(accepted), SUM(ignored), SUM(declined), SUM(withdrawn) FROM Allocations WHERE timestamp BETWEEN '1 week ago' AND 'now’ AND driver='LON123456789’ GROUP BY timestamp(day)

JAXLONDON2013

Get a picture of driver supply

SELECT COUNT DISTINCT(driverId) FROM driverLocs WHERE timestamp BETWEEN '1 day ago' AND 'now' GROUP BY timestamp(hour) SELECT COUNT FROM driverLocs WHERE timestamp BETWEEN '1 day ago' AND 'now' GROUP BY latitude(0.01), longitude(0.01)

JAXLONDON2013

JAXLONDON2013

Operational perspective

JAXLONDON2013

“Allows a team of 2 to achieve things they wouldn’t have considered before Cassandra existed”

Chris H, Operations Engineer

JAXLONDON2013

JAXLONDON2013

3 clusters

6 machines per region

3 regions (stats cluster is a long story)

Operational Cluster

Stats Cluster

ap-southeast-1 us-east-1 eu-west-1

us-east-1 eu-west-1

AZ1

eu-west-1

AZ1

AZ2 AZ2

AZ3 AZ3

AZ1

us-east-1

AZ1

AZ2 AZ2

AZ3 AZ3

AZ1

ap-southeast-1

AZ1

AZ2 AZ2

AZ3 AZ3

JAXLONDON2013

JAXLONDON2013

AWS VPCs with Open VPN links

3 AZs per region

m1.large machines

Provisoned IOPS EBS

Operational Cluster

Stats Cluster

~ 1TB/node

~ 200GB/node

JAXLONDON2013

Backups

•  SSTable snapshot

•  Used to upload to S3, but this was taking >6 hours and consuming all our network bandwidth

•  Now take EBS snapshot of the data volumes

JAXLONDON2013

Encryption

•  Requirement for NYC launch

•  We use dmcrypt to encrypt the entire EBS volume

•  Chose dmcrypt because it is uncomplicated

•  Our tests show a 1% performance hit in disk performance, which concurs with what Amazon suggest

JAXLONDON2013

Datastax Ops Centre is a quick win

JAXLONDON2013

Multi DC

•  Something that Cassandra makes trivial

•  Would have been very difficult to accomplish active-active inter-DC replication with a team of 2 without Cassandra

•  Rolling repair needed to make it safe (we use LOCAL_QUORUM)

•  We schedule “narrow repairs” on different nodes in our cluster each night

JAXLONDON2013

Compression

•  Our stats cluster was running at ~1.5TB per node

•  We didn’t want to add more nodes

•  With compression, we are now back to ~600GB

•  Easy to accomplish

•  `nodetool upgradesstables` on a rolling schedule

JAXLONDON2013

Management perspective

JAXLONDON2013

“The days of the quick and dirty are over”

Simon V, EVP Operations

JAXLONDON2013

Technically, everything is fine…

•  Our COO feels that C* is “technically good and beautiful”, a “perfectly good option”

•  Our EVPO says that C* reminds him of a time series database in use at Goldman Sachs that had “very good performance”

…but there are concerns

JAXLONDON2013

People who can attempt to query MySQL

People who can attempt to

query Cassandra

JAXLONDON2013

JAXLONDON2013

Lessons learned

JAXLONDON2013

There might be a gulf in experience

JAXLONDON2013

10 Average years experience per team member

MySQL Cassandra

JAXLONDON2013

Lesson learned

•  Have an advocate - get someone who will sell the vision internally

•  Learn the theory - teach each team member the fundamentals

•  Make an effort to get everyone on board

JAXLONDON2013

Things can drift into failure

JAXLONDON2013

JAXLONDON2013

JAXLONDON2013

JAXLONDON2013

JAXLONDON2013

JAXLONDON2013

Lesson learned

•  Be pro-active with Cassandra, even if it seems to be running smoothly

•  Peer-review data models, take time to think about them

•  Big rows are bad - use cfstats to look for them

•  Mixed workloads can cause problems - use cfhistograms and look out for signs of data modeling problems

•  Think about the compaction strategy for each CF

JAXLONDON2013

EBS is terrible

JAXLONDON2013

Lessons learned

•  EBS is nearly always the cause of Amazon outages

•  EBS is a single point of failure (it will fail everywhere in your cluster)

•  EBS is slow

•  EBS is expensive

•  EBS is unnecessary!

JAXLONDON2013

Management need to know the trade offs

JAXLONDON2013

Lessons learned

•  Keep the business informed – explain the tradeoffs in simple terms

•  Sing from the same hymn sheet

•  Make sure there solutions in place for every use case from the beginning

JAXLONDON2013

People who can attempt to query MySQL People who can

attempt to query Cassandra

JAXLONDON2013

Conclusions

JAXLONDON2013

We like Cassandra

•  Solid design

•  HA characteristics

•  Easy multi-DC setup

•  Simplicity of operation

JAXLONDON2013

Lessons for successful adoption

•  Have an advocate, sell the dream

•  Learn the fundamentals, get the best out of Cassandra

•  Invest in tools to make life easier

•  Keep management in the loop, explain the trade offs

JAXLONDON2013

The future

•  We will continue to invest in Cassandra as we expand globally

•  We will hire people with experience running Cassandra

•  We will focus on expanding our reporting facilities

•  We aspire to extend our network (1M consumer installs, wallet) beyond cabs

•  We will continue to hire the best engineers in London, NYC and Asia

JAXLONDON2013

Questions?

Recommended