WEBVTT

1
00:00:00.210 --> 00:00:02.020
<v Instructor>Okay, so let's do a deep dive</v>

2
00:00:02.020 --> 00:00:04.680
on EC2 spot instances.

3
00:00:04.680 --> 00:00:06.570
We know that with a spot instance,

4
00:00:06.570 --> 00:00:10.870
you can get a discount up to 90% compared to On-Demand.

5
00:00:10.870 --> 00:00:12.420
So that's huge cost saving.

6
00:00:12.420 --> 00:00:13.950
How does that work?

7
00:00:13.950 --> 00:00:17.190
We define a max spot price that we're willing to pay

8
00:00:17.190 --> 00:00:18.440
for that spot instance.

9
00:00:18.440 --> 00:00:23.080
And then as long as the instance is having a spot price

10
00:00:23.080 --> 00:00:25.910
that is less than the max price we're willing to pay,

11
00:00:25.910 --> 00:00:27.810
then we keep that instance.

12
00:00:27.810 --> 00:00:29.520
So that max spot price, I will show it to you

13
00:00:29.520 --> 00:00:30.520
in the next slide.

14
00:00:30.520 --> 00:00:33.880
But the hourly spot will vary based on offer and capacity,

15
00:00:33.880 --> 00:00:35.720
and so it will go up and it will go down.

16
00:00:35.720 --> 00:00:38.500
And if somehow the current spot price

17
00:00:38.500 --> 00:00:40.940
goes over the max price you have defined,

18
00:00:40.940 --> 00:00:42.510
then you have two options.

19
00:00:42.510 --> 00:00:43.910
And for these two options,

20
00:00:43.910 --> 00:00:46.120
you have a two minutes grace period.

21
00:00:46.120 --> 00:00:48.170
So that gives you a little bit of time to do these things.

22
00:00:48.170 --> 00:00:51.550
And the options can be either you're stopping your instance,

23
00:00:51.550 --> 00:00:54.020
that means that you shut down everything you were doing,

24
00:00:54.020 --> 00:00:55.450
then you stop your instance.

25
00:00:55.450 --> 00:00:58.293
And so if one day, the spot price goes below your max price,

26
00:00:58.293 --> 00:01:00.220
then you can restart your instance

27
00:01:00.220 --> 00:01:02.660
and continue where you left it off.

28
00:01:02.660 --> 00:01:05.720
Or if you don't need the state on your EC2 instance,

29
00:01:05.720 --> 00:01:08.400
you can just choose to terminate your instance

30
00:01:08.400 --> 00:01:09.250
and let it go.

31
00:01:09.250 --> 00:01:10.460
And then that would assume

32
00:01:10.460 --> 00:01:12.210
that anytime you would restart your work,

33
00:01:12.210 --> 00:01:15.000
you can start with a fresh new EC2 instance.

34
00:01:15.000 --> 00:01:17.650
So two strategies based on the type of workload you have,

35
00:01:17.650 --> 00:01:20.530
but you have a two minutes grace periods to do so.

36
00:01:20.530 --> 00:01:23.540
The other strategy, if you don't want your spot instance

37
00:01:23.540 --> 00:01:27.140
to be reclaimed by AWS is to use a spot block.

38
00:01:27.140 --> 00:01:30.410
A spot block is when you block a spot instance

39
00:01:30.410 --> 00:01:32.330
for a specified timeframe.

40
00:01:32.330 --> 00:01:35.000
This can be between one to six hours,

41
00:01:35.000 --> 00:01:38.410
and you get that block without any interruption,

42
00:01:38.410 --> 00:01:40.970
at least on paper, because in the documentation,

43
00:01:40.970 --> 00:01:43.540
but it's really rare, in rare situations,

44
00:01:43.540 --> 00:01:45.430
the instance may be reclaimed.

45
00:01:45.430 --> 00:01:47.850
But overall, when you consider a spot block

46
00:01:47.850 --> 00:01:49.880
think that the instance will not be reclaimed,

47
00:01:49.880 --> 00:01:52.290
that's the whole purpose of the spot block.

48
00:01:52.290 --> 00:01:54.510
So when do we use spot instances?

49
00:01:54.510 --> 00:01:56.530
Well, for batch jobs, data analysis,

50
00:01:56.530 --> 00:01:58.200
workloads that are resilient to failures,

51
00:01:58.200 --> 00:01:59.840
and as I said, they're not great

52
00:01:59.840 --> 00:02:02.410
for critical jobs or databases.

53
00:02:02.410 --> 00:02:05.610
Now let's have a look at the spot instance pricing.

54
00:02:05.610 --> 00:02:07.270
So this is a graph that I obtained directly

55
00:02:07.270 --> 00:02:09.250
from the AWS console.

56
00:02:09.250 --> 00:02:11.760
As you can see here, this gives us the price

57
00:02:11.760 --> 00:02:16.676
for a m4.large instance over three months in the spot zone.

58
00:02:16.676 --> 00:02:20.293
As you can see, you have six lines for US-East-1,

59
00:02:21.732 --> 00:02:23.360
the region and you have six availability zones,

60
00:02:23.360 --> 00:02:24.270
so six prices.

61
00:02:24.270 --> 00:02:27.200
So as you can see, the spot price does vary

62
00:02:27.200 --> 00:02:29.550
based on the AZ you're considering.

63
00:02:29.550 --> 00:02:33.660
As we can see, over three months, it has varied quite a bit.

64
00:02:33.660 --> 00:02:35.470
For example, if you look at that yellow line,

65
00:02:35.470 --> 00:02:38.900
it has varied between 0.4 to 0.45,

66
00:02:38.900 --> 00:02:43.720
all the way down to almost 0.35, let's say and so on.

67
00:02:43.720 --> 00:02:47.630
Let's assume that we are defining a user defined max price

68
00:02:47.630 --> 00:02:51.333
to be this black horizontal bar dotted line.

69
00:02:52.340 --> 00:02:56.300
We'd see is that if the yellow line is above

70
00:02:56.300 --> 00:02:59.380
the dotted line, then the current spot price is going

71
00:02:59.380 --> 00:03:02.090
to be more than the max price we have defined.

72
00:03:02.090 --> 00:03:03.770
In this instance, right here,

73
00:03:03.770 --> 00:03:05.640
we're going to lose our spot instance.

74
00:03:05.640 --> 00:03:07.790
So we have to terminate it or stop it.

75
00:03:07.790 --> 00:03:09.500
But if the price goes down,

76
00:03:09.500 --> 00:03:13.080
so if the yellow line goes down our dotted line,

77
00:03:13.080 --> 00:03:15.930
that means that we are willing to pay that price.

78
00:03:15.930 --> 00:03:18.810
We are gaining that spot instance, in this meantime.

79
00:03:18.810 --> 00:03:20.740
This is how we can optimize the cost

80
00:03:20.740 --> 00:03:23.170
of our workload very easily.

81
00:03:23.170 --> 00:03:28.160
As you can see, the On-Demand price is $0.10 per hour,

82
00:03:28.160 --> 00:03:29.350
so that's really, really cheap.

83
00:03:29.350 --> 00:03:32.610
And then the spot instance price is even cheaper.

84
00:03:32.610 --> 00:03:35.430
It's around 0.4, I would say.

85
00:03:35.430 --> 00:03:38.500
It's a 60% saving in that instance,

86
00:03:38.500 --> 00:03:40.990
for the spot instance pricing, so lots of savings.

87
00:03:40.990 --> 00:03:43.000
And obviously, if you were to define

88
00:03:43.000 --> 00:03:44.550
your user defined max price

89
00:03:44.550 --> 00:03:48.560
to be something like $0.05 per hour,

90
00:03:48.560 --> 00:03:51.110
then no in-spot instances will be reclaimed

91
00:03:51.110 --> 00:03:52.030
because you're always willing

92
00:03:52.030 --> 00:03:54.930
to pay more than the actual spot instance price.

93
00:03:54.930 --> 00:03:57.770
So no one can predict how the price goes over time.

94
00:03:57.770 --> 00:03:59.760
But as you can see, it can be rather stable.

95
00:03:59.760 --> 00:04:01.930
It can fluctuate a little bit, but overall,

96
00:04:01.930 --> 00:04:05.373
it still provides you huge savings compared to On-Demand.

97
00:04:06.670 --> 00:04:09.780
Next, how do you terminate a spot instance?

98
00:04:09.780 --> 00:04:13.380
So we have to first understand how a spot request works.

99
00:04:13.380 --> 00:04:15.630
For this, let's consider a spot request.

100
00:04:15.630 --> 00:04:16.920
With a spot request,

101
00:04:16.920 --> 00:04:19.940
you are defining how many instances you want,

102
00:04:19.940 --> 00:04:22.030
your maximum price you're going to pay,

103
00:04:22.030 --> 00:04:24.630
the launch specification, so the AMI and so on,

104
00:04:24.630 --> 00:04:27.280
and when your request is valid from and until,

105
00:04:27.280 --> 00:04:30.180
but it could be infinite, and then the request type.

106
00:04:30.180 --> 00:04:31.370
It's very important to understand

107
00:04:31.370 --> 00:04:32.940
there's two types of requests.

108
00:04:32.940 --> 00:04:35.820
You can do one-time request for spot instances

109
00:04:35.820 --> 00:04:38.930
or a persistent request for post instances.

110
00:04:38.930 --> 00:04:40.950
Let's have a look at the difference.

111
00:04:40.950 --> 00:04:42.710
If it's a one-time request,

112
00:04:42.710 --> 00:04:45.560
as soon as your spot request is fulfilled,

113
00:04:45.560 --> 00:04:47.630
your instances are going to be launched.

114
00:04:47.630 --> 00:04:50.490
And then your spot request will go away

115
00:04:50.490 --> 00:04:52.570
because it was a one-time request type.

116
00:04:52.570 --> 00:04:55.370
So in this case, your spot request is going to go away

117
00:04:55.370 --> 00:04:56.520
and you're fine.

118
00:04:56.520 --> 00:04:58.770
But if it's a persistent request type,

119
00:04:58.770 --> 00:05:02.620
that means that we want this number of instances to be valid

120
00:05:02.620 --> 00:05:06.810
as long as the spot request is valid from to valid until.

121
00:05:06.810 --> 00:05:10.830
That means that if somehow your instances do get stopped,

122
00:05:10.830 --> 00:05:12.980
or interrupted based on the spot price,

123
00:05:12.980 --> 00:05:16.340
then your spot requests will go back into action.

124
00:05:16.340 --> 00:05:18.350
When things can be validated

125
00:05:18.350 --> 00:05:21.220
we'll restart spot instances for you.

126
00:05:21.220 --> 00:05:25.200
So if somehow you stop a spot instance, in persistent mode,

127
00:05:25.200 --> 00:05:27.400
and your spot request is still active,

128
00:05:27.400 --> 00:05:30.090
your spot request automatically will be smart enough

129
00:05:30.090 --> 00:05:32.690
to restart a launch instance for you.

130
00:05:32.690 --> 00:05:35.300
And this is something you can also see in this diagram,

131
00:05:35.300 --> 00:05:36.920
but I won't spend too much time on it.

132
00:05:36.920 --> 00:05:38.230
So what you understand out of it

133
00:05:38.230 --> 00:05:42.600
is that if you want to cancel a spot request,

134
00:05:42.600 --> 00:05:44.670
it needs to be in the open state,

135
00:05:44.670 --> 00:05:46.600
the active state, or the disabled state.

136
00:05:46.600 --> 00:05:48.130
That means that basically, it's not failed,

137
00:05:48.130 --> 00:05:50.030
it's not cancelled or it's not closed.

138
00:05:51.164 --> 00:05:54.130
When you want to cancel a spot request,

139
00:05:54.130 --> 00:05:56.050
so if you take the spot request on the left hand side

140
00:05:56.050 --> 00:06:00.123
and you cancel it, it's not going to terminate any instances

141
00:06:00.123 --> 00:06:01.900
that you have launched from before.

142
00:06:01.900 --> 00:06:04.230
So it is still your responsibility

143
00:06:04.230 --> 00:06:05.980
to terminate these instances

144
00:06:05.980 --> 00:06:09.060
and not the responsibility of AWS.

145
00:06:09.060 --> 00:06:12.220
So as such, if you want to terminate spot instances

146
00:06:12.220 --> 00:06:14.390
for good and not have them relaunched,

147
00:06:14.390 --> 00:06:17.500
you need to first cancel the spot request,

148
00:06:17.500 --> 00:06:21.410
and then you terminate the associated spot instances.

149
00:06:21.410 --> 00:06:25.200
Because if you were to terminate the spot instances first,

150
00:06:25.200 --> 00:06:27.300
remember, it goes back into the spot request

151
00:06:27.300 --> 00:06:29.920
and a spot request says, okay, you wanted six instances,

152
00:06:29.920 --> 00:06:31.580
but I can see you have zero right now,

153
00:06:31.580 --> 00:06:33.960
I'm going to launch six instances for you.

154
00:06:33.960 --> 00:06:36.060
So again, the right order to cancel

155
00:06:36.060 --> 00:06:38.220
and to terminate spot instances is

156
00:06:38.220 --> 00:06:40.510
to first cancel the spot request to make sure

157
00:06:40.510 --> 00:06:44.010
that no new instances will be launched by AWS.

158
00:06:44.010 --> 00:06:47.200
Then you terminate the associated spot instances.

159
00:06:47.200 --> 00:06:48.200
I'm spending a bit of time on this

160
00:06:48.200 --> 00:06:51.290
because this is something that can come up in the exam.

161
00:06:51.290 --> 00:06:53.100
Okay, next spot fleets.

162
00:06:53.100 --> 00:06:55.780
So this is the ultimate way to save money.

163
00:06:55.780 --> 00:06:58.680
Spot fleets is going to be a way to define for you

164
00:06:58.680 --> 00:07:02.990
to get a set of spot instances, and optionally,

165
00:07:02.990 --> 00:07:05.650
On-Demand instances of what's called a fleet.

166
00:07:05.650 --> 00:07:08.380
And the spot fleet will try its best

167
00:07:08.380 --> 00:07:09.940
to meet the target capacity

168
00:07:09.940 --> 00:07:12.090
with the price constraints you defined.

169
00:07:12.090 --> 00:07:15.060
So it will launch from possible launch pools.

170
00:07:15.060 --> 00:07:17.330
It can have different instance types,

171
00:07:17.330 --> 00:07:20.260
different OS and availability zones.

172
00:07:20.260 --> 00:07:22.093
You're going to define multiple launch pools,

173
00:07:22.093 --> 00:07:24.210
a multiple instance size, multiple everything,

174
00:07:24.210 --> 00:07:27.353
and then the fleet will choose the best

175
00:07:27.353 --> 00:07:31.240
and most appropriate launch pool for you.

176
00:07:31.240 --> 00:07:34.490
When your spot fleet either reaches your budget

177
00:07:34.490 --> 00:07:36.470
or reaches the capacity you wanted,

178
00:07:36.470 --> 00:07:39.170
then it will stop launching instances.

179
00:07:39.170 --> 00:07:41.330
And so you define a strategy

180
00:07:41.330 --> 00:07:44.110
to allocate spot instances in your spot fleet.

181
00:07:44.110 --> 00:07:45.350
That's something to remember.

182
00:07:45.350 --> 00:07:47.500
The first one is lowest price,

183
00:07:47.500 --> 00:07:49.350
and that's going to be the most popular one

184
00:07:49.350 --> 00:07:50.570
going into the exam,

185
00:07:50.570 --> 00:07:54.420
which is that the spot fleets will launch instances

186
00:07:54.420 --> 00:07:57.320
from the pool that has the lowest price

187
00:07:57.320 --> 00:08:00.080
and that gives you a lot of cost optimization.

188
00:08:00.080 --> 00:08:02.640
This is a great option if you have a very short workload.

189
00:08:02.640 --> 00:08:05.470
It can also have a diversified way

190
00:08:05.470 --> 00:08:08.266
to launch your spot instances, in which case,

191
00:08:08.266 --> 00:08:11.680
the spot instances will be distributed

192
00:08:11.680 --> 00:08:14.730
across all the pools that you have defined from before,

193
00:08:14.730 --> 00:08:17.170
which is great for availability and long workloads

194
00:08:17.170 --> 00:08:18.910
because if one pool goes away,

195
00:08:18.910 --> 00:08:21.240
then your other pools are still active.

196
00:08:21.240 --> 00:08:23.730
And then finally, you have capacityOptimized

197
00:08:23.730 --> 00:08:26.980
to launch with the optimal capacity based on the number

198
00:08:26.980 --> 00:08:30.400
of instances and choose the right pool for that.

199
00:08:30.400 --> 00:08:32.610
The idea is that spot fleets can be complicated,

200
00:08:32.610 --> 00:08:35.610
but what you have to remember is that using the spot fleet,

201
00:08:35.610 --> 00:08:37.590
you're able to define multiple launch pools,

202
00:08:37.590 --> 00:08:38.570
multiple instance types

203
00:08:38.570 --> 00:08:40.430
until you're just interesting in real power.

204
00:08:40.430 --> 00:08:43.620
And then if you use the lowestPrice discounts,

205
00:08:43.620 --> 00:08:46.170
or the lowest price rate strategy for spot fleets,

206
00:08:46.170 --> 00:08:48.760
then the spot fleets will automatically request

207
00:08:48.760 --> 00:08:52.150
the spot instances with the lowest price for us.

208
00:08:52.150 --> 00:08:55.180
So spot fleet gives us an extra saving

209
00:08:55.180 --> 00:08:57.680
based on spot instances, because it's smart enough

210
00:08:57.680 --> 00:08:59.990
to choose the right spot instance pools

211
00:08:59.990 --> 00:09:03.090
to allow us to get the maximum amount of savings.

212
00:09:03.090 --> 00:09:05.370
This is the benefit of spot fleets.

213
00:09:05.370 --> 00:09:07.410
Again, you need to understand the difference here

214
00:09:07.410 --> 00:09:10.790
of doing a very simple spot instance request

215
00:09:10.790 --> 00:09:12.840
or you know exactly the type of instance you want

216
00:09:12.840 --> 00:09:15.810
and the AZ you want, to using the spot fleet and saying,

217
00:09:15.810 --> 00:09:17.330
okay, you can choose all these instance types

218
00:09:17.330 --> 00:09:19.990
and all these AZ and what I need from you

219
00:09:19.990 --> 00:09:23.270
is to give me the lowest price, for example.

220
00:09:23.270 --> 00:09:25.840
That's hopefully is giving you a lot of insights

221
00:09:25.840 --> 00:09:28.560
into how spot fleets and spot instances are different.

222
00:09:28.560 --> 00:09:29.560
I hope you like this lecture

223
00:09:29.560 --> 00:09:31.510
and I will see you in the next lecture.

