CARTER MORGAN: Welcome to another episode
CARTER MORGAN: Welcome to another episode

of "VM End-to-End," a show where we have a VM skeptic
of "VM End-to-End," a show where we have a VM skeptic

and a VM enthusiast get together and hash out
and a VM enthusiast get together and hash out

all things VM related.
all things VM related.

Brian, thanks again for coming in and spitting knowledge
Brian, thanks again for coming in and spitting knowledge

at me.
at me.

BRIAN DORSEY: I'm so excited about this one.
BRIAN DORSEY: I'm so excited about this one.

Disks, disks, disks.
Disks, disks, disks.

CARTER MORGAN: Yeah, you would--
CARTER MORGAN: Yeah, you would--

I don't understand this.
I don't understand this.

I've been trying to figure this out for three episodes now.
I've been trying to figure this out for three episodes now.

What is it about disks that is so exciting for you?
What is it about disks that is so exciting for you?

BRIAN DORSEY: I think part of it,
BRIAN DORSEY: I think part of it,

for me, is that I think almost all useful systems have
for me, is that I think almost all useful systems have

to store their information somewhere, right?
to store their information somewhere, right?

Your system has a memory, and you care about the data.
Your system has a memory, and you care about the data.

Most of the time, that's on a disk somewhere,
Most of the time, that's on a disk somewhere,

or ends up on a disk in the end.
or ends up on a disk in the end.

And so I think, for me, it's where systems get real.
And so I think, for me, it's where systems get real.

And--
And--

CARTER MORGAN: OK.
CARTER MORGAN: OK.

BRIAN DORSEY: --that's the big deal.
BRIAN DORSEY: --that's the big deal.

CARTER MORGAN: I could see that.
CARTER MORGAN: I could see that.

It's a fundamental unit of computing.
It's a fundamental unit of computing.

And so that's what's exciting.
And so that's what's exciting.

BRIAN DORSEY: Yeah.
BRIAN DORSEY: Yeah.

And on that unit, we think of them
And on that unit, we think of them

as physical hard disks or maybe an SSD.
as physical hard disks or maybe an SSD.

But let me turn this back on you.
But let me turn this back on you.

In the cloud, in Compute Engine, what do you think a disk is?
In the cloud, in Compute Engine, what do you think a disk is?

CARTER MORGAN: OK.
CARTER MORGAN: OK.

All right.
All right.

Going back to a slice of a data-center model, to me,
Going back to a slice of a data-center model, to me,

a disk would be a bunch of pieces
a disk would be a bunch of pieces

from real physical machines spread out
from real physical machines spread out

over a data warehouse or a data center,
over a data warehouse or a data center,

pulled together, and presented to me, the user,
pulled together, and presented to me, the user,

as one logical disk that I can work with and operate on.
as one logical disk that I can work with and operate on.

BRIAN DORSEY: Exactly.
BRIAN DORSEY: Exactly.

Spot-on.
Spot-on.

CARTER MORGAN: Yes.
CARTER MORGAN: Yes.

BRIAN DORSEY: That is exactly it.
BRIAN DORSEY: That is exactly it.

And in fact, some of that-- so it's a network-attached disk,
And in fact, some of that-- so it's a network-attached disk,

right?
right?

And it speaks of a protocol.
And it speaks of a protocol.

And even physical disks, these days, speak of packet protocol.
And even physical disks, these days, speak of packet protocol.

So SATA is serial ATA.
So SATA is serial ATA.

And they talk about the blocks.
And they talk about the blocks.

So here's a block of data, or give me that block of data.
So here's a block of data, or give me that block of data.

And ours speak of similar protocol, but over the network,
And ours speak of similar protocol, but over the network,

to a group of disks.
to a group of disks.

CARTER MORGAN: So I guess, what is
CARTER MORGAN: So I guess, what is

it that's exciting about that, and especially
it that's exciting about that, and especially

the network-attached part, because what I care about
the network-attached part, because what I care about

is performance.
is performance.

I care about backups and scalability.
I care about backups and scalability.

I care about reliability more so than a fundamental unit
I care about reliability more so than a fundamental unit

of computing.
of computing.

Where do these disks sit in relation to those?
Where do these disks sit in relation to those?

BRIAN DORSEY: Yeah.
BRIAN DORSEY: Yeah.

OK.
OK.

This is a meaty question.
This is a meaty question.

So I think, actually, let's talk about each of these in a row.
So I think, actually, let's talk about each of these in a row.

So performance, to start.
So performance, to start.

I think performance is tricky.
I think performance is tricky.

What does it mean to be fast?
What does it mean to be fast?

And I think there's actually two different parts of that,
And I think there's actually two different parts of that,

for disks.
for disks.

One is throughput.
One is throughput.

In a bulk, how much data can you copy over
In a bulk, how much data can you copy over

a certain amount of time?
a certain amount of time?

And the other is latency.
And the other is latency.

So how quickly can you get a response back
So how quickly can you get a response back

to a specific question?
to a specific question?

CARTER MORGAN: Right.
CARTER MORGAN: Right.

It feels like a physical disk is going
It feels like a physical disk is going

to be faster than that because I don't have to communicate over
to be faster than that because I don't have to communicate over

a network.
a network.

That seems like common sense to me.
That seems like common sense to me.

BRIAN DORSEY: And I agree it seems like common sense.
BRIAN DORSEY: And I agree it seems like common sense.

And it's true for one and not for the other.
And it's true for one and not for the other.

And it's weird why that is.
And it's weird why that is.

So as physical disks are getting faster over time--
So as physical disks are getting faster over time--

sorry, as they're getting larger and faster--
sorry, as they're getting larger and faster--

but they're getting larger faster
but they're getting larger faster

than they're getting faster.
than they're getting faster.

So there's more and more capacity on the disk.
So there's more and more capacity on the disk.

And the window to read and write that
And the window to read and write that

is not getting as big as fast.
is not getting as big as fast.

So if you're talking over the network
So if you're talking over the network

to hundreds or thousands of disks,
to hundreds or thousands of disks,

you can actually get more overall throughput.
you can actually get more overall throughput.

So the throughput side is an area
So the throughput side is an area

where we are actually faster than a single physical disk.
where we are actually faster than a single physical disk.

But sometimes-- there is a little extra step there.
But sometimes-- there is a little extra step there.

So if you need the absolute lowest latency read,
So if you need the absolute lowest latency read,

you might need something faster.
you might need something faster.

So we have a product, local SSD, to do that.
So we have a product, local SSD, to do that.

CARTER MORGAN: And that's basically
CARTER MORGAN: And that's basically

just literally attaching a very fast hard drive to a machine.
just literally attaching a very fast hard drive to a machine.

BRIAN DORSEY: Yeah, it's a--
BRIAN DORSEY: Yeah, it's a--

CARTER MORGAN: OK.
CARTER MORGAN: OK.

BRIAN DORSEY: --physical SSD on the host
BRIAN DORSEY: --physical SSD on the host

that we are running your VM on.
that we are running your VM on.

CARTER MORGAN: OK.
CARTER MORGAN: OK.

So that seems, to me, like--
So that seems, to me, like--

I come from video game programming.
I come from video game programming.

So is it something I would use for maybe
So is it something I would use for maybe

an inventory-management system?
an inventory-management system?

Something like, I recall items in inventory often.
Something like, I recall items in inventory often.

Is that a good fit for this?
Is that a good fit for this?

BRIAN DORSEY: It would seem like it, but I think, in this case,
BRIAN DORSEY: It would seem like it, but I think, in this case,

probably not.
probably not.

So the local SSDs are actually intended for temporary storage
So the local SSDs are actually intended for temporary storage

because they are mapped to this physical host.
because they are mapped to this physical host.

And we want our VMs to be long-lived and able to move
And we want our VMs to be long-lived and able to move

around.
around.

So there are cases where we can't
So there are cases where we can't

have the data live as long as you might like on a local SSD.
have the data live as long as you might like on a local SSD.

So they're best fit for things like caches or data
So they're best fit for things like caches or data

that can be regenerated by processing or transcoding
that can be regenerated by processing or transcoding

workloads or things like that.
workloads or things like that.

CARTER MORGAN: Intermediate calculations
CARTER MORGAN: Intermediate calculations

in math or vectors or whatnot.
in math or vectors or whatnot.

OK.
OK.

BRIAN DORSEY: Absolutely.
BRIAN DORSEY: Absolutely.

CARTER MORGAN: All right.
CARTER MORGAN: All right.

That makes a lot of sense, then.
That makes a lot of sense, then.

So then what does Compute Engine offer
So then what does Compute Engine offer

when it comes to different types of physical disks
when it comes to different types of physical disks

or different types of PDs, persistent disks?
or different types of PDs, persistent disks?

BRIAN DORSEY: Yeah.
BRIAN DORSEY: Yeah.

I think there are multiple different types,
I think there are multiple different types,

and I'm probably not going to go into all the details.
and I'm probably not going to go into all the details.

But the basic breakdown is an area
But the basic breakdown is an area

that I think you didn't hit here that I
that I think you didn't hit here that I

think does matter to everybody.
think does matter to everybody.

And that's cost.
And that's cost.

CARTER MORGAN: Oh, that's a big one.
CARTER MORGAN: Oh, that's a big one.

Yeah.
Yeah.

BRIAN DORSEY: So it basically boils down
BRIAN DORSEY: So it basically boils down

to a price-to-performance ratio.
to a price-to-performance ratio.

So when you have a workload that is really
So when you have a workload that is really

performance-sensitive, there are versions of persistent disks
performance-sensitive, there are versions of persistent disks

that can hit that.
that can hit that.

And they tend to be a little more expensive.
And they tend to be a little more expensive.

And if you've got a workload that is mostly
And if you've got a workload that is mostly

about storing information, and you don't access it that often,
about storing information, and you don't access it that often,

there are less expensive modes of PD to do that.
there are less expensive modes of PD to do that.

CARTER MORGAN: And that makes sense to me.
CARTER MORGAN: And that makes sense to me.

If I was a small-business owner, I
If I was a small-business owner, I

might not need the speed and scale
might not need the speed and scale

that, say, a company like the size of Google
that, say, a company like the size of Google

would operate at.
would operate at.

BRIAN DORSEY: And--
BRIAN DORSEY: And--

CARTER MORGAN: So then--
CARTER MORGAN: So then--

BRIAN DORSEY: Sorry, I just cut you off there.
BRIAN DORSEY: Sorry, I just cut you off there.

CARTER MORGAN: --to not get into-- oh--
CARTER MORGAN: --to not get into-- oh--

BRIAN DORSEY: But--
BRIAN DORSEY: But--

CARTER MORGAN: Mm-hmm.
CARTER MORGAN: Mm-hmm.

BRIAN DORSEY: --to that point, I think lots of workloads need
BRIAN DORSEY: --to that point, I think lots of workloads need

both as--
both as--

or, systems need both.
or, systems need both.

You've got a long-term storage and a high-speed portion
You've got a long-term storage and a high-speed portion

as well.
as well.

CARTER MORGAN: OK.
CARTER MORGAN: OK.

That's actually a really good point.
That's actually a really good point.

So then, OK, talking about systems--
So then, OK, talking about systems--

and like you said, in systems in the while
and like you said, in systems in the while

might need two types of disk.
might need two types of disk.

What about backing up these kind of systems
What about backing up these kind of systems

that are in production?
that are in production?

Do these virtual-network disks actually
Do these virtual-network disks actually

help with reliability or backing up data?
help with reliability or backing up data?

BRIAN DORSEY: A lot.
BRIAN DORSEY: A lot.

Absolutely.
Absolutely.

So like you said, it's this logical collection of blocks,
So like you said, it's this logical collection of blocks,

right?
right?

So in many ways, I think it's like having
So in many ways, I think it's like having

most of a SAN, a storage area network.
most of a SAN, a storage area network.

So you get most of the features you would want from that.
So you get most of the features you would want from that.

And by that, I mean you can take a snapshot of a disk, which
And by that, I mean you can take a snapshot of a disk, which

makes a point-in-time copy without slowing the disk down.
makes a point-in-time copy without slowing the disk down.

So you can make backups of live systems any time you want.
So you can make backups of live systems any time you want.

And the other thing you can do is,
And the other thing you can do is,

you can take an existing disk, and you can just
you can take an existing disk, and you can just

make it larger, even while it's connected to a machine.
make it larger, even while it's connected to a machine.

I feel like everybody has run out of disk space
I feel like everybody has run out of disk space

at some point.
at some point.

CARTER MORGAN: I have a laptop, and it's run out of space.
CARTER MORGAN: I have a laptop, and it's run out of space.

And there's nothing I can really do about it
And there's nothing I can really do about it

other than plugging in an extra hard drive.
other than plugging in an extra hard drive.

So that's a feature that makes a lot of sense to me.
So that's a feature that makes a lot of sense to me.

You said something earlier about snapshots and whatnot, though.
You said something earlier about snapshots and whatnot, though.

And I'm curious, what's exciting about that?
And I'm curious, what's exciting about that?

How does that become usable in a production system?
How does that become usable in a production system?

BRIAN DORSEY: So there's the straight-up backup use case,
BRIAN DORSEY: So there's the straight-up backup use case,

which I think is important.
which I think is important.

But then one of the big ways we scale computers these days is--
But then one of the big ways we scale computers these days is--

there's vertical, and there's horizontal, right?
there's vertical, and there's horizontal, right?

So vertical might be like making the disk bigger,
So vertical might be like making the disk bigger,

like we talked about.
like we talked about.

And then horizontal is, add more computers to the mix doing it.
And then horizontal is, add more computers to the mix doing it.

And so if you have a computer that's
And so if you have a computer that's

set up the way you want it to work,
set up the way you want it to work,

you can take a snapshot of that.
you can take a snapshot of that.

And then you can make multiple images
And then you can make multiple images

of that that are running on multiple machines
of that that are running on multiple machines

and have them work as a group to do the work.
and have them work as a group to do the work.

CARTER MORGAN: OK.
CARTER MORGAN: OK.

That is actually really cool, because you're
That is actually really cool, because you're

going to get reliability and consistency there.
going to get reliability and consistency there.

You're not going to get human errors.
You're not going to get human errors.

BRIAN DORSEY: Yeah.
BRIAN DORSEY: Yeah.

CARTER MORGAN: I am a huge fan of that.
CARTER MORGAN: I am a huge fan of that.

Is there anything else there that you want to touch on?
Is there anything else there that you want to touch on?

Because, like you said, you're going to get backups from this.
Because, like you said, you're going to get backups from this.

You're going to be able to scale vertically and horizontally
You're going to be able to scale vertically and horizontally

from this.
from this.

I feel like those are some really big benefits to disks
I feel like those are some really big benefits to disks

and why they're useful, and even more
and why they're useful, and even more

useful than having a physical non-cloud-native disk.
useful than having a physical non-cloud-native disk.

So what do we need to follow up with in the future?
So what do we need to follow up with in the future?

We said cost.
We said cost.

BRIAN DORSEY: Yeah.
BRIAN DORSEY: Yeah.

CARTER MORGAN: I missed cost, and that's really important.
CARTER MORGAN: I missed cost, and that's really important.

So let's definitely have an episode on cost.
So let's definitely have an episode on cost.

What else am I missing?
What else am I missing?

BRIAN DORSEY: I think a whole episode there makes sense.
BRIAN DORSEY: I think a whole episode there makes sense.

And then, I think, with the scaling stuff,
And then, I think, with the scaling stuff,

you get more moving parts, right?
you get more moving parts, right?

So if you go horizontal, you'll have more computers.
So if you go horizontal, you'll have more computers.

You'll have more disks.
You'll have more disks.

You start-- need to think about networking patterns.
You start-- need to think about networking patterns.

And so it's useful to have some automation
And so it's useful to have some automation

to set that up and configure it in a consistent way.
to set that up and configure it in a consistent way.

CARTER MORGAN: Yeah, and I'm actually
CARTER MORGAN: Yeah, and I'm actually

excited about that because I'm a big fan of automating
excited about that because I'm a big fan of automating

as much as I can.
as much as I can.

I make mistakes.
I make mistakes.

I don't want to risk it.
I don't want to risk it.

Brian, thanks again for just taking the time.
Brian, thanks again for just taking the time.

I've got to admit, slowly, you're
I've got to admit, slowly, you're

winning me over, because--
winning me over, because--

BRIAN DORSEY: Love it.
BRIAN DORSEY: Love it.

CARTER MORGAN: --there seems to be a lot of benefits.
CARTER MORGAN: --there seems to be a lot of benefits.

Thank you.
Thank you.

That's all for us this episode.
That's all for us this episode.

If you're watching this at home, let me know anything
If you're watching this at home, let me know anything

that you think we missed talking about or anything
that you think we missed talking about or anything

that you thought was very cool.
that you thought was very cool.

I'm getting won over right now.
I'm getting won over right now.

Are you with me or no?
Are you with me or no?

Let us know in the comments below.
Let us know in the comments below.

[MUSIC PLAYING]
[MUSIC PLAYING]