How LinkedIn uses memcached, a spoonful of SOA, and a sprinkle of SQL to scale

JDBC – We don’t need no
stinking JDBC. How
LinkedIn uses memcached,
a spoonful of SOA, and a
sprinkle of SQL to scale.
David Raccah & Dhananjay Ragade
LinkedIn Corporation

Goal of this Presentation
What you will learn
How LinkedIn built a cheap
and scalable system to
store our member’s profiles,
and how you can do the
same

2

Agenda

> Review system ilities
> What happened to databases?
> SOA What
> Discuss existing Best Practices
> Pixie Dust and Kool-Aid are not so bad
> What LinkedIn’s got up their sleeve
> How it all came together…
> Q&A

3

Terminology of the ilities
the terms of large successful systems
> Performance
 Not an “ility” but without it, no ility will save you
> Availability
 Availability is the proportion of time a system is in a
functioning condition
> Reliability
 The probability that a functional unit will perform its
required function for a specified interval under
stated conditions.
 The ability of something to "fail well" (fail without
catastrophic consequences)

4

Terminology of the ilities
the terms of large successful systems
> Scalability
 Slow with multiple users vs. single user
> Manageability
 The ability to manage all parts of a large moving
system
> Serviceability
 The ability to service an arm of the system without
bleeding to death (e.g. change out a database from
a working system). Bleeding is OK in a high
performance system – death is NOT acceptable.

5

Agenda

> SOA What
> Q&A

6

Databases
The systems that drive the enterprise … or….
> RDBMS – Relational Data
Base Management System
Attribute
> KVSS – Key Value Storage
System
> Enterprise Search Engines

7

Database Server History….

8

Database mind set has changed…
From data access to data management to….
> Initially it was all about remote
data access with an index
> Then it moved to ACID data
management and tooling
> Then it became an Application
Server with data affinity
> Now we have come full circle
and people have figured out
that scaling is more important
than relationships, transactions,
and data and behavioral affinity.

9

Database Mantra that Rule the Roost
ACID
> Atomicity – All or nothing
> Consistency – Data in the
system should never get in a
contradictory state.
> Isolation: Two requests
cannot interfere with one
another.
> Durability: No do over – once
the data is persisted, it
cannot change.

10

Anti-Database Rules
BASE
> Basically Available
 Support partial failures within your
architecture (e.g. sharding)
> Soft state
 State may be out of synch for
some time
> Eventually consistent
 Eventually all data is made
consistent (as long as the
hardware is reliable)

11

Database Scalability
Or lack thereof…
> Databases work. Look at:
 Hotmail
 Facebook
 eBay
> Databases scale with hardware
> They do not scale horizontally
well
 Partition management is
nonexistent and RYO is a mess
 Many use them as ISAM and
not even relational

12

Database Tools and language
Duh…
> Defacto standards for tools and
languages abound for relational
databases
> Easy to manage the data within
a partition and easy to write
code to operate on said data
> Terrifying but nice to use
extensions include running
Java within the Data Engine, so
that you could run your
application within the big iron

13

Database’s other features
Which are the pain points….
> Constraints – Nice idea until
you start partitioning.
2PC is the anti-scalability
pattern (Pat Helland)
> Computation – this feature turns out to cause more
pain as cost rises with scale and are incompatible
with most languages and tools.
> Replication & backup
 Nice tools that are indeed important and useful
> ACL support & Data Engine optimizations
 Used for sure, but exist to circumvent deficiencies

14

Key Value Storage Systems
BigTable, Hive, Dynamo– the Wild Wild West
> Reliable – Proven on web
> Available – redundant (locally)
> Scalable – no constraints
> Limited ACIDity
> No Standard and not portable
> Almost no:
 Constraints or relationships
 Computation or transactions

15

Enterprise Search Engines
Index yes – storage device no
> A great inverted index
> Finds data quickly
> However, what it returns is
commonly an ID to the
entity(s) in question
> Real-Time solutions are
available but not fully
deployed today
> Limited ACIDity/transactions
> Scalable, available, reliable
16

Agenda

> SOA What
> Q&A

17

SOA
Service Oriented Architecture
> SOA may be overkill for most
enterprises
> Still a Tiered and layered
architecture – which is what
SOA hoped to formulate and
standardize is a solid approach
> Services (not SOA) allow for
efficient reuse of business
processes and aggregation
services within a complex
development organization
18

Agenda

> SOA What
> Q&A

19

Best Practices
Storage and architecture
> Store critical data redundantly
and reliably with a cluster
 Google via BigTable, Facebook
via MySQL, eBay via replicated &
sharded DB
> Layer services on top of the
storage device to manage data
integrity and complexity
 LinkedIn, Amazon, eBay

20

Best Practices
> Create a bus to route
replicated data to consumers –
e.g. search, data mining, etc.
 Almost all sites
> Parallelization via things like
scatter/gather
 Almost all search topologies
(Google, Yahoo, Live),
 Facebook, etc.

21

Best Practices
> Keep the system stateless
 eBay, Google, etc.
> Partition data and services
 Facebook, eBay
> Cache data
> Replicate your data
> Route requests to where the
behavior and/or data exists
> Degrade gracefully with load

22

Best Practices
> Tiering systems
 Latency vs. Affinity
 Traversal versus affinity – you need to
analyze the cost and make a decision
 Scaling vs. parallelizing
 Do you need to keep tiering all
systems to keep the scalability
uniform?
 Complexity vs. diminished
dependencies
 Does the reduced dependencies make
up for the increased system
complexity?

23

Agenda

> SOA What
> Q&A

24

Pixie Dust and Kool-Aid
Building on the past

25

Pixie Dust and Kool-Aid
Building on the past
> So what do we want:
 Reliable
 Available
 Scalable
 ACIDity on simple transactions
 Standard and portable interface
 Data Optimizations
 Cache and replicate
 Low cost BASE architecture

26

Agenda

> SOA What
> Q&A

27

LinkedIn’s Data Services
Mixture of standards and pixie dust
> Front a database with a service
> Cache data
> Route to and partition the data
service
> Scale and replicate services in a
horizontal manner
> Keep all writes ACID and
subsequent reads ACID as well

28

Mixture of standards and pixie dust
> Databases are reliable
> Scale out at the service
> Replicate and cache
> Partitioning comes from the front
tier and business servers that
front the data services

29

Immediate replication vs. eventual replication
> Caching needs a consistency algorithm
> Techniques for immediate replication
 Paxos
 Chubby, Microsoft AutoPilot, Zoo Keeper
 N Phase Commit (2PC and 3PC)
> Techniques for eventual consistency
 BASE (Basically Available, Soft-state,
Eventual Consistency
 Inktomi, Dynamo, AWS

30

LinkedIn’s approach
> Keep core data ACID
> Keep replicated and cached data BASE
> Replicate data via the data bus
> Cache data on a cheap memory
(memcached)
> Use a hint to route the client to his /
her’s ACID data

31

Databus – the linchpin of our replication

32

LinkedIn’s approach

33

Core DS
> Keep core data ACID in the DB
> All writes come here.
> Databus source for all replication
> The last line of defense for a
cache miss
> Manages sharding

34

RepDS
> Manages cache consistency and
replication
> Manages the freshness of the
caller
> Reads come from cache

35

RepReader
> RepReader is the typical tip of the
iceberg problem
> All read operations are sourced
from the cache unless the caller’s
freshness token is out of the window

36

Freshness Token (AKA Pixie Dust)
> The freshness token = Pixie Dust for
CUD operations
> It also allows us to give the caller
control over whether they are content
with BASE data, even if they did no
CUD operation.

37

For the love of Pixie dust and Kool-Aid
> We use commodity hardware and
software to run our service
> We use Pixie Dust to keep costs down
and keep our customer happy
> We keep OPS and the exec-staff
happy with our special brand of Kool-
Aid

38

Agenda

> SOA What
> Q&A

39

Profile Re-architecture
Changing planes in mid-flight

> Original LinkedIn System
> Use of XML for i18n
> Phased Transition

40

Problems from the original system
Anthropology 101
> Be fair… it worked well for
a startup
> Many tables in one big
DB
> Too many similar object
hierarchies
> No well defined domains

41

Why XML?
Flexibility
> Profile has many fields
> 1NF for I18n ==> too many
tables
> StAX for fast parsing
> Easier to version the profile
> Human readable
> JSON? ProtoBuf?

42

Issues with XML
<good/> <bad/> <ugly/>
> XML schema design tradeoffs
and analytics impact
> XML is verbose
> StAX is unfriendly
> XML in the DB caused us
some performance
headaches

43

Phased Transition
Evolving a living, breathing organism
> Successive iterations avoid breakages
> No major site downtime
> Easier to sanity check
> Does not hold other teams hostage
> Phases LinkedIn went through

44

Double Writes Topology
Safety first

45

After Legacy Tables Dropped
Auld Lang Syne

46

Wrap up
The moral of the story is…
> Keep your system BASE
> Use commodity hardware
> Use pixie dust (AKA data freshness token)
> Evolve slowly - no big bang!

47

David Raccah & Dhananjay Ragade
draccah@linkedin.com
dragade@linkedin.com
Linkedin Corporation

49

Performance
Often mixed up with scalability
> Performance
 A numerical value given to
a single system when
asked to do a task under
nominal load
 If the system responds
poorly without load, it will
assuredly continue its
molasses response time
under load

51

Availability
Often mixed up with reliability
> Availability
 A numerical value
given to a system
that defines the
proportion of time a
system is in a
functioning condition.
 Most common
scoring system is
called nines – which is defined as the uptime versus
the uptime and downtime – five nines = 0.99999

52

Reliability
The ability for a system to perform its functionality
> Reliability
 A system can be 100% available
and still be 100% unreliable (e.g.
non consistent caching)
 A person can consistently give
you the wrong answer
 Architecture is defined as the
balance of the ilities and cost

53

Scalability
the term that many think is the holy grail
> Scalability
 The ability for a system to manage
more traffic or to be “scaled” as
more traffic appears
 System slows with multiple users
vs. single user
 Route, Partition, Orchestrate,
replicate, and go asynch
 Split the system horizontally
 Rarely scale vertically

54

The rest of the ilities
the ones that people tend to ignore till its too late
> Manageability
 It is a double-edged
sword which can be
easily ignored
> Serviceability
 Here complexity starts to
rear its ugly head
> Maintainability
 Of course maintainability
tends to run upstream of
complexity

55

How LinkedIn uses memcached, a spoonful of SOA, and a sprinkle of SQL to scale

More Related Content

Viewers also liked (20)

Similar to How LinkedIn uses memcached, a spoonful of SOA, and a sprinkle of SQL to scale (20)

More from LinkedIn (20)

Recently uploaded (20)

How LinkedIn uses memcached, a spoonful of SOA, and a sprinkle of SQL to scale