Thanks for coming.

I know the day's been running a little bit long,

so I will hopefully not keep you too long.

Welcome to the session between you and happy hour.

Hopefully, we'll get you out of here just in time

to get some drinks and get some snacks.

My name is Sandeep Parikh.

I run the cloud solutions architecture

team for Americas East, for Google Cloud Platform,

and today, I want to talk to you about building

a collaborative data platform.

Effectively, you've got teams of individuals spread

across your company, that have to work together

in some meaningful way, and you need

them to share data, talk to each other,

basically build something together.

So, how do you enable that?

So, we're going to walk through some of the ways

that we tell a lot of customers, and developers, and partners

how to do that stuff in the real world.

First, we're going to meet our hypothetical team

of individuals, and we're going to learn

about the tasks that have to do, and the tools that they

typically use.

And then, we'll jump into mapping

the tools they use to the tools available in Google Cloud,

so you can start to see how they might

get a consistent set of tools and work together.

And then, we're going to talk about how

we enable collaboration.

This is how we do things, like set up teams

to do self-service and easy data discovery.

I'm probably going to trivialize a handful of details,

but stay with me, it'll make sense as we get through it.

And then, I want to cover what the new workflow's

like once you get all this stuff up and running,

and then give you a couple little things

and tips on how to get started.

Does that make sense to everybody?

Head nods?

OK, good.

All right, so let's meet the team.

All right, so first, we've got about four individuals.

We've got a data scientist, a data engineer, a business

analyst, and an app developer.

Ultimately, they all have three things in common.

They use different data sets, those data sets

are all in different formats, and they all

use different tools, right?

So, there's not a ton of overlap here,

and that's kind of the challenge.

So, let's assume for a second, or let's imagine for a second,

we've got this team, and their job

is to build a recommendation engine for a retail site.

So, what are the kinds of things that each person on the team

is going to need to be successful

in order to do their job?

So, the first thing we'll cover is the data,

and then we'll talk about the tools.

So, from a data perspective, Alice, the data scientist,

needs kind of cleansed, normalized data.

Things that have no longer log files, right,

not just raw rows from a database,

but things that actually make sense.

And those could be, for example, stuff

like purchase history, product history, product metadata,

click streams, product reviews.

All the stuff that it would take to actually craft and model

a recommendation problem.

Then, we've got Bob, the data engineer.

He's going to have to take all of the raw data

and turn it into something useful.

So, he's probably going to need things like,

logs information, product transactions or purchase

history, product metadata, as well.

This is all the stuff that's coming in

straight off the application, that he's

got to turn into something useful for other parts

of the organization.

Then, there's Ned.

Ned is the business analyst, and he

needs a lot of aggregated data, things

that he can use to generate statistics

and understanding about how the business is performing.

Not just, again, rows or log files, but again,

that next level up from, sort of, the basics.

And then, finally, we have Jan, the app developer.

She's going to need users and product predictions,

or recommendations.

She's going to be able to-- she needs

to be able to build new microservices that

expose things like this recommendation engine

to the end user.

So, that can be things like a recommendations API,

access to wish lists, some knowledge around similar users

or similar products, how things complement each other.

But, you can see there's a little bit of overlap here,

but not a ton, right.

They all need something different

and they kind of need it in a slightly different format.

Now, if you think about sort of a--

I tried to draw a sequence diagram here,

and it got a little complicated, so I simplified it

a little bit.

But, ultimately, Alice and Bob have to agree and talk

to each other.

Bob and Ned have to agree and talk to each other.

Ned and Alice have to talk each other,

and Jan and Alice have to talk to each other.

You know, there's a lot of communication happening.

And, listen, I love talking to my coworkers just as much

as the next person, but if you're

spending all of your time trying to explain to them what you

need, when you need it, where it came from,

what it should look like, it becomes challenging, right.

It slows down the ability for them

to make progress and work quickly.

Ultimately, this is the problem that we

see a lot of folks having out in the, world, right?

They each have to find data sets.

They each have to figure out who owns it.

They have to figure out how things are related

to each other, and they have to manage all that by themselves,

which, again, takes time away from their ultimate task.

Just by a show of hands, does this sound familiar to anybody?

Oh, that's a scary number, actually.

Maybe I shouldn't have asked that question.

No, but that this is great, right?

This is how we have to-- we have to find, understand,

what the problem is, then try to solve it together.

All right, so let's talk about tools.

That's the first step.

We need to understand what everyone's

working with in order to kind of make the next jump.

So, Alice is the data scientist, huge fan of Python,

like myself.

It's like her Swiss army knife, she uses it

for just about everything.

Typically, does a lot of her work

with IPython or Jupyter notebooks,

pulls data down, does it on her workstation,

you know, slices dices data, uses

a lot of pretty common data science frameworks,

uses things like, NumPy, SciPy, scikit-learn,

Pandas, because she has a history,

and/or likes data frames.

That's a typical approach, from a data science perspective.

Bob is a huge fan of Java, right?

He does lots of big ETL pipeline work,

does a lot of orchestrating data from one format to another,

or multiple sources of data together into something

that looks cohesive on the other end.

So, he's typically using tools like MapReduce or Spark.

In some cases, he might be using Apache Beam, which

you'll actually hear more about over the next couple days,

as well.

Ned is like a spreadsheet guy.

He loves SQL and he loves spreadsheets.

So, what he likes to do is build custom reports

and dashboards, very SQL-driven, got to have some kind of access

to a data warehouse, so he can do giant--

you know, kind of cube rotations and understanding.

Basically, he's got to be able to prove to the business

how products are performing, why things like a recommendation

engine is even necessary.

So, it's his responsibility to kind of facilitate that.

And then, we've got Jan, the app developer.

Jan's a typical kind of polyglot app developer.

Likes writing stuff on microservices, likes delivering

features very quickly, you know, may have a language preference

one way or the other, but has to maintain a lot of the existing

infrastructure.

So, that could be things like node apps.

That could be simple things like Python apps with Flask or Ruby

apps with Sinatra, all the way up

to like kitchen sink frameworks, like Django or Rails.

Just depends, right, whatever the use case is.

So, we've gotten that far.

So, we understand their task, we know

who all the key players are, and we know the things

that they like to use.

So, the next step for us is, how do we

figure out which tools they should

be using in Google Cloud?

All right, we can skip that slide.

So, what we'll do is let's lay out

a handful of the tools that are available.

And we're not going to cover everything that's

in GCP right now, because we don't have enough time,

and I want you guys to have snacks at some point soon.

So, what we'll do is kind of focus on the core set of things

that are critical for each of those individual roles

to have access to.

So, the first part, you know, for things like applications,

we've got Compute.

So, with things like virtual machine

or virtual instances with Compute Engine,

manage Kubernetes clusters with Container Engine,

or just kind of manage container deployments,

or Docker container deployments with App Engine.

It's a good place to start.

On the storage front, cloud storage for objects,

Cloud SQL for relational data, Cloud Datastore for, you know,

NoSQL, or non-relational datasets,

and Cloud Bigtable for wide column.

And then, on the data and analytics,

are then the more processing kind of side,

Dataproc for managed Hadoop and Spark clusters,

Datalab for IPython notebooks, BigQuery for kind

of an analytical data warehouse, and data flow for running

managed Beam, Apache Beam pipelines.

And then, the last column, or the last kind of group,

is machine learning.

So we've got cloud machine learning,

which is kind of running TensorFlow models,

and running them at scale, things

like the natural language API, the speech API, or the vision

API.

As you start to think about the team

that we just introduced, and you start

looking at this sort of bucket of tools, of this set of puzzle

pieces, you can start to see where things

are going to kind of fit together

There's a consistent set of tools

here that we can map back on to the team.

So let's walk through each one of those individually,

really quickly.

So, for data science workloads, Alice has a couple of options.

We already talked about the fact that she's a fan of Python,

so that maps really well.

So, with Python, she has two options.

She can either use Cloud Datalab,

or she can use Cloud Dataproc with Jupyter.

With Datalab, what she gets is a complete Python dev

environment.

You know, it's bundled in with like NumPy, SciPy, Matplotlib,

so she can kind of kick off her work, and build those charts,

and build those, kind of understanding,

or on the data set as she would.

Additionally, on top of that though,

it's also got built-in support for TensorFlow and BigQuery.

This means she's got a complete environment to go and start