All language subtitles for 16 - SAA Spot Instances Spot Fleet English

Okay, so let's do a deep dive

on EC2 spot instances.

We know that with a spot instance,

you can get a discount up to 90% compared to On-Demand.

So that's huge cost saving.

How does that work?

We define a max spot price that we're willing to pay

for that spot instance.

And then as long as the instance is having a spot price

that is less than the max price we're willing to pay,

then we keep that instance.

So that max spot price, I will show it to you

in the next slide.

But the hourly spot will vary based on offer and capacity,

and so it will go up and it will go down.

And if somehow the current spot price

goes over the max price you have defined,

then you have two options.

And for these two options,

you have a two minutes grace period.

So that gives you a little bit of time to do these things.

And the options can be either you're stopping your instance,

that means that you shut down everything you were doing,

then you stop your instance.

And so if one day, the spot price goes below your max price,

then you can restart your instance

and continue where you left it off.

Or if you don't need the state on your EC2 instance,

you can just choose to terminate your instance

and let it go.

And then that would assume

that anytime you would restart your work,

you can start with a fresh new EC2 instance.

So two strategies based on the type of workload you have,

but you have a two minutes grace periods to do so.

The other strategy, if you don't want your spot instance

to be reclaimed by AWS is to use a spot block.

A spot block is when you block a spot instance

for a specified timeframe.

This can be between one to six hours,

and you get that block without any interruption,

at least on paper, because in the documentation,

but it's really rare, in rare situations,

the instance may be reclaimed.

But overall, when you consider a spot block

think that the instance will not be reclaimed,

that's the whole purpose of the spot block.

So when do we use spot instances?

Well, for batch jobs, data analysis,

workloads that are resilient to failures,

and as I said, they're not great

for critical jobs or databases.

Now let's have a look at the spot instance pricing.

So this is a graph that I obtained directly

from the AWS console.

As you can see here, this gives us the price

for a m4.large instance over three months in the spot zone.

As you can see, you have six lines for US-East-1,

the region and you have six availability zones,

so six prices.

So as you can see, the spot price does vary

based on the AZ you're considering.

As we can see, over three months, it has varied quite a bit.

For example, if you look at that yellow line,

it has varied between 0.4 to 0.45,

all the way down to almost 0.35, let's say and so on.

Let's assume that we are defining a user defined max price

to be this black horizontal bar dotted line.

We'd see is that if the yellow line is above

the dotted line, then the current spot price is going

to be more than the max price we have defined.

In this instance, right here,

we're going to lose our spot instance.

So we have to terminate it or stop it.

But if the price goes down,

so if the yellow line goes down our dotted line,

that means that we are willing to pay that price.

We are gaining that spot instance, in this meantime.

This is how we can optimize the cost

of our workload very easily.

As you can see, the On-Demand price is $0.10 per hour,

so that's really, really cheap.

And then the spot instance price is even cheaper.

It's around 0.4, I would say.

It's a 60% saving in that instance,

for the spot instance pricing, so lots of savings.

And obviously, if you were to define

your user defined max price

to be something like $0.05 per hour,

then no in-spot instances will be reclaimed

because you're always willing

to pay more than the actual spot instance price.

So no one can predict how the price goes over time.

But as you can see, it can be rather stable.

It can fluctuate a little bit, but overall,

it still provides you huge savings compared to On-Demand.

Next, how do you terminate a spot instance?

So we have to first understand how a spot request works.

For this, let's consider a spot request.

With a spot request,

you are defining how many instances you want,

your maximum price you're going to pay,

the launch specification, so the AMI and so on,

and when your request is valid from and until,

but it could be infinite, and then the request type.

It's very important to understand

there's two types of requests.

You can do one-time request for spot instances

or a persistent request for post instances.

Let's have a look at the difference.

If it's a one-time request,

as soon as your spot request is fulfilled,

your instances are going to be launched.

And then your spot request will go away

because it was a one-time request type.

So in this case, your spot request is going to go away

and you're fine.

But if it's a persistent request type,

that means that we want this number of instances to be valid

as long as the spot request is valid from to valid until.

That means that if somehow your instances do get stopped,

or interrupted based on the spot price,

then your spot requests will go back into action.

When things can be validated

we'll restart spot instances for you.

So if somehow you stop a spot instance, in persistent mode,

and your spot request is still active,

your spot request automatically will be smart enough

to restart a launch instance for you.

And this is something you can also see in this diagram,

but I won't spend too much time on it.

So what you understand out of it

is that if you want to cancel a spot request,

it needs to be in the open state,

the active state, or the disabled state.

That means that basically, it's not failed,

it's not cancelled or it's not closed.

When you want to cancel a spot request,

so if you take the spot request on the left hand side

and you cancel it, it's not going to terminate any instances

that you have launched from before.

So it is still your responsibility

to terminate these instances

and not the responsibility of AWS.

So as such, if you want to terminate spot instances

for good and not have them relaunched,

you need to first cancel the spot request,

and then you terminate the associated spot instances.

Because if you were to terminate the spot instances first,

remember, it goes back into the spot request

and a spot request says, okay, you wanted six instances,

but I can see you have zero right now,

I'm going to launch six instances for you.

So again, the right order to cancel

and to terminate spot instances is

to first cancel the spot request to make sure

that no new instances will be launched by AWS.

Then you terminate the associated spot instances.

I'm spending a bit of time on this

because this is something that can come up in the exam.

Okay, next spot fleets.

So this is the ultimate way to save money.

Spot fleets is going to be a way to define for you

to get a set of spot instances, and optionally,

On-Demand instances of what's called a fleet.

And the spot fleet will try its best

to meet the target capacity

with the price constraints you defined.

So it will launch from possible launch pools.

It can have different instance types,

different OS and availability zones.

You're going to define multiple launch pools,

a multiple instance size, multiple everything,

and then the fleet will choose the best

and most appropriate launch pool for you.

When your spot fleet either reaches your budget

or reaches the capacity you wanted,

then it will stop launching instances.

And so you define a strategy

to allocate spot instances in your spot fleet.

That's something to remember.

The first one is lowest price,

and that's going to be the most popular one

going into the exam,

which is that the spot fleets will launch instances

from the pool that has the lowest price

and that gives you a lot of cost optimization.

This is a great option if you have a very short workload.

It can also have a diversified way

to launch your spot instances, in which case,

the spot instances will be distributed

across all the pools that you have defined from before,

which is great for availability and long workloads

because if one pool goes away,

then your other pools are still active.

And then finally, you have capacityOptimized

to launch with the optimal capacity based on the number

of instances and choose the right pool for that.

The idea is that spot fleets can be complicated,

but what you have to remember is that using the spot fleet,

you're able to define multiple launch pools,

multiple instance types

until you're just interesting in real power.

And then if you use the lowestPrice discounts,

or the lowest price rate strategy for spot fleets,

then the spot fleets will automatically request

the spot instances with the lowest price for us.

So spot fleet gives us an extra saving

based on spot instances, because it's smart enough

to choose the right spot instance pools

to allow us to get the maximum amount of savings.

This is the benefit of spot fleets.

Again, you need to understand the difference here

of doing a very simple spot instance request

or you know exactly the type of instance you want

and the AZ you want, to using the spot fleet and saying,

okay, you can choose all these instance types

and all these AZ and what I need from you

is to give me the lowest price, for example.

That's hopefully is giving you a lot of insights

into how spot fleets and spot instances are different.

I hope you like this lecture

and I will see you in the next lecture.