UnConference for Georgia Southern Computer Science March 31, 2015

Topics
Agile in the real world
Career
Kafka
DevOps
Cloud and aaS
NoSQL: MongoDB
NoSQL: Redis
BloomFilters
Speed Round (5 topics in 8 minutes)
Big Data (Hadoop, Spark)
Actor Systems (Akka)
Streaming (Storm, InfoSphere Streams)

Scrum, Kanban

Processes for managing work

Team is not just software engineers
− QA, test automation
− Product Analyst
− Documentation
− Production/Operations Engineers

Co-location

Kanban

Limits amount of work team is working on

Visually displays what is being worked on,
waiting

Some online tools, but teams usually work on
the walls

Message Queues

What if you had billions of messages each day?

Any couldn't lose any of them?

'current' technology is a queue, held in memory
or RDBMS

Lots of problems with holding in memory

Lots of RDBMS activity

Broker needs to know who consumed a
message so he can delete it

Changing the Paradigm

Kafka uses disk as the primary store

Kernel level writes directly to disk controller =
FAST

Kafka doesn't care if anyone consumed an
event

Consumers ask for a specific offset in a topic

Consumers can listen for new events

Use Cases

Huge volumes of input data > memory

Consumers at different times/rates

Replay for support, developers

Old Way *

Someone gives you a 2” thick specification

You code it for 6 months

Test it for 3 months

Give it to operations to run

Get bug reports

Fix them, deploy them monthly

*unfortunately a lot of people still do this

Old Way - Operations

Manually build a server

Hope you got everything on it

Deploy the software, hope you got everything

Manually test that the software works

If the server fails, repeat

Upgrades take hours

DevOps and Agile Way

Customer presents stories to team

Team asks questions, scopes and defines tests

Team develops code, tests, talks with client

Production engineers learn what team is
building, what tech etc.

Continuous Integration builds make sure it
works

Minimum Viable Product is produced

Operations runs a command, code is deployed

DevOps - Operations

Template for a server is defined (OS, software etc.)

Template for software release is defined (versions,
dependencies, automated tests)

Deployment: Stock OS server is booted

Deployment command is run

UrbanCode/Chef downloads Server templates,
application templates and automated tests. Runs all of
them, tells operations it is ready.

Server failure? Boot new machine and repeat

Scale? Boot new machine and repeat

DevOps Software

Lots of features added to a release to support
in production

Alerts, monitoring, metrics

Automated testing

Interfaces for 'probing' what is going on inside

New Tech is vetted with operations during
development, not 'throw it over the wall'

Cloud

Most hyped technology in the last 10 years

(Big Data is a close 2nd
)

Basic idea: someone deploys thousands
(millions in Amazon's case) of servers and
makes it easy for you to use them

No capital costs. Pay for what you use

Infinite* capacity if you built it right

Vendors are continually adding new features

Why?

$0.20 an hour servers

IT Doesn't know about it ...

Site gone viral? Spin up 100 more app servers,
bigger database server (see DevOps
discussion)

No capital up front. $100/month until the hockey
stick, then revenue covers usage

As a Service (*aaS)

Business model for making $$ on the cloud

Offer a service to business at a cost less than if
they did it themselves

IaaS – Infrastructure as a Service (Amazon,
SoftLayer)

PaaS – Platform as a Service (BlueMix, Azure,
AppEngine)

SaaS – Software as a Service (Silverpop,
Hotmail, Gmail, LinkedIn)

AaS terminology

Virtual Machine

Software Defined Networking

Multi-tenancy

Scalability, Redundancy and Disaster Recovery

Netflix

Tips

See how far you can get with PaaS, but be
wary of vendor-specific features

Look for standard databases (MySQL,
MongoDB, Redis, SQL Server) PaaS before
you stand up your own

Docker

Shut it down when you aren't using it!

Document Databases

Schema-less

Store JSON, search by JSON, retrieve JSON

Document is the transaction layer

Sharding & Partitioning

Lots of scale problems are addressed by splitting the
data

Partitioning is the data is grouped by a key in the data.
The location is usually fixed (or takes a lot to move).

Sharding is the data is grouped by a key, but the
location can change.

Example: shard on event timestamp.
− Partition: all 12/2015 go into the same disk
− Shard: all 12/2015 could start in the same disk, but
be divided by smaller date ranges as volume
increases

“Internet Scale”

Adding new nodes is easy

Shards 'move' based on load

Perfect for event storage with no or in place
changes

But ...

Not for documents that grow

JSON is tricky to query

Schema-less sounds great, until you have to
manage database someone else built

No substitute for good design

Data Structure Server

CS 101 data structures: list, set, string

CS 101 operations: push, pop, add, delete

Very, very fast. Very simple

100% in memory

Single threaded

Lots of people use it as an application cache

Big O notation documented for every operation

So?

Lets you build shared state across lots of
consumers

Pushes operations to the server, so they are
atomic

For example: set the 1000th
bit on a bit array
and increment a population counter. Without
pulling the data back to the client

Use Cases

BloomFilters

Work queue for thousands of worker threads
(Akka) without continual database polling

Cache of web sessions

“For the cost of an Oracle Enterprise License I
can give you 64 cores and 3 TB of memory”

Bloom Filters
From WikiPedia (Don't tell my kid's teacher!)
"A Bloom filter is a space-efficient probabilistic data structure, conceived by
Burton Howard Bloom in 1970, that is used to test whether an element is a
member of a set. False positive matches are possible, but false negatives are
not, thus a Bloom filter has a 100% recall rate"

Hashing

Apply 'x' hash functions to the key to be stored/queried

Each function returns a bit to set in the bitset

Mathematical equations to determine how big to make the bitset, how
many functions to use and your acceptable error level

http://hur.st/bloomfilter?n=4&p=1.0E-20

False Positives

Perfect hash functions aren't worth the cost to develop

Sometimes existing bits for a key are set by many other keys

Make sure you understand the business impact of a false positive

Remember, never a false negative

Question

How would you keep track of the number of
unique visitors to a website today?

What if I wanted to know if a specific user had
visited today?

SQL Query?

List of visitors in memory?

Examples

Redis-backed
− Unique visitor counts this week
− White lists/black lists
− What ad to show a visitor? (hash cookie id)
− Client side joins (time based population counts)

Databases
− Is this key in the page?

Fail Fast

Speed Round
Things we could talk about for hours. In minutes

Version Control

Git is free. Use it

Local repository, keep copies of all your work

Very helpful when you're working on a project
and need to revert

Plus a good thing to talk about on an interview

Google Docs

Huh?

Get a Gmail account (or custom domain, more
about this later)

Use Gmail's 10 GB of free storage to put all
your class assignments.

No risk of losing school work. Can access from
any computer

IntelliJ

For the Java Users

Eclipse is 'difficult'. Sorry, but it is

(So is vi, don't get me started on emacs)

Community edition is free

Syntax highlighting, multiple language support

Clustering

Technology for building a set of processes that
work together

Auto detect (and launch) of process failures

Keeps your service available 24x7

ZooKeeper, Mesos are common technologies

If done right, failure of a server isn't a 'wake up
the team!' event.

Replication

Usually requires Clustering

One process is the 'master' of the data or
processing.

One or more 'slaves' listen to updates/actions
of the master silently

If the master dies, Clustering will promote a
slave to master and the processing continues

Sometimes you can read from slaves for read
performance increases since they should be
identical (or nearly) to the master

Bonus: Online Resources

Let me Google that for you

HighScalability.com

StackOverflow.com

Infoq.com

Feedly.com – RSS isn't dead yet

What?

2nd
most hyped concept in the last 5 years
(Cloud is #1)

Basically, we generate too much data

Companies are afraid to throw it away

But don't know what to do with it, or if it is even
valuable

How to store it?

Often unstructured

Lots and lots of it (sensor logs)

Can't store it on one disk, can't risk losing it if
that disk is lost

Can't back it up

HDFS – Hadoop File System

Distributed, multiple copies, multiple servers

How to access it?

Not in a RDBMS

Not in a typical NoSQL (Redis, MongoDB etc.)

Hadoop is one common tool for making sense
out of it

Still need to write code (Cascading is my
favorite)

Other Approaches

Spark
− Actor based approach to processing data
− Uses HDFS, but holds results in memory as long as
possible
− Still need to write code
− Very, very fast

Hbase
− Non-relational database with columns and rows
− Most JDBC drivers can talk to it

Careers

Data Scientist
− Apply knowledge of statistics and algorithms to the
big data
− Find actionable insights in the data to drive the
business
− Build Predictive models based on data to determine
what might happen next
− Both coding and math skills

Your phone is more powerful than
the Space Shuttle's computers

Typical phone has 2 cores (some 4)

Typical laptop has 4 cores

Typical server has 8 cores

Being able to do things in parallel is necessary
to scale

Multi-threading is really hard

Books written about it

Courses taught about it

Java got it wrong TWICE so far

Concurrency bugs are very difficult to debug
and fix

Servers are so cheap now, concurrency across
machines is very common

Actors

So rather than writing the code yourself, why
not rely on a concurrency model?

Actors are called with a piece of work to do,
respond with an answer

All the threading, clustering, messaging in
handled by it.

Akka for Java. Scala Actors (or Akka for Scala)

Use case

Redis list with all the tasks to be performed

Akka actor tells 'manager' it is ready for work

Manager pops item from Redis into a key
specific for that worker and tells the worker
what to do.

When done, the worker tells the manager and
the key is removed (or more tasks are created)

Scales to thousands of machines, 10 threads
per machine

Real time and over time processing

Some events need to be responded to
immediately

Some events need to be compared to other
events in the last 'x' time periods to understand
what to do about them

Some events can be discarded quickly, but
need to evaluate them to determine that

Used to be $$$$ to use

Storm changed that

Open Source Event Processing System used
by Twitter

You write code to evaluate events as they
occur, not to query a set of events at rest

Not 'everyone in the last hour who bought more
than $100'

Rather 'this person just bought >$100' so do
something. Don't wait

Use Cases

Combined with Kafka for event ingestion

Every purchase request, check if the
geographic coordinates match the last 5
purchases, or their home zip code + 5 miles

Every download of a white paper, see if they
are a sales prospect and alert the salesperson
the customer is on the site and interested

Careers

Security and anti-fraud

Online marketing

Securities Trading (Wall Street)

Career
Managing your career before you have one

Linked In

Get one

This is not Facebook or Instagram

Who would you want to work with in 5 years?

Who wouldn't you?

Stay in touch with Intern/co-op co-workers

Lots of useful information, follow 'smart' people

Internships

Get one, even on campus

LinkedIn to search for them

Nepotism is okay!

Open Source

Find an interesting project

Does it have a 'startup guide'? Or a poorly
written one?

Get an example working and submit a
documentation update

Another thing to talk about during an interview

Study Groups

Get several!

No, not organized cheating

What you understand someone else doesn't

Eventually you won't be the smartest in a group

Interview discussion topic

Resumes

Gold Award? Eagle Scout?

List courses in major

Describe your impact on the business (even if a
cook or waiting tables)

One page. Really.

Vanity Emails – time to get rid of them
− Seriously consider a Google-hosted domain ($50
year)

Interviews

Prepare, prepare, prepare

Details about course work

How will your internship help the company?

Test First Development during coding examples

Ask questions

Be ready: What else do you do for fun?

Life

Have one outside of computers

Have non-CS friends

Get a hobby

Exercise (Google 'The Hackers diet')

Why? Burn out

Other things
Learn about Mobile
Learn about JavaScript client libraries (JQuery,
Angular.js etc.) even if you want to do backend
work
Learn about security, both building secure and the
tools for testing
Data Science has a lot of buzz (Learn R,
Predictive Analytics)
Internet of Things (IoT)

Finally
Never stop asking why?
Never stop taking things apart
Never stop listening and learning

UnConference for Georgia Southern Computer Science March 31, 2015

Recommended

More Related Content

What's hot (20)

Similar to UnConference for Georgia Southern Computer Science March 31, 2015 (20)

Recently uploaded (20)

UnConference for Georgia Southern Computer Science March 31, 2015