ML Implementations and Operations

Question 1

What is the name of the library used for making containers compatible with Sagemaker?

Answer 1

sagemaker-containers

Question 2

What is the file structure of a training container on Sagemaker?

Answer 2

/opt/ml
├── input
│ ├── config
│ │ ├── hyperparameters.json
│ │ └── resourceConfig.json
│ └── data
│ └── <channel_name>
│ └── <input></input>
├── model
│
├── code
│ └──

│
└── output
└── failure

</channel_name>

Question 3

What is the file structure of a docker image?

Answer 3

WORKDIR
* nginx.conf
* predictor.py
* serve/
* train/
* wsgi.py

Question 4

What are production variants of a model?

Answer 4

They variants of a model that you can test on live traffic through the usage of weighted traffic redistribution. Allows you to perform A/B testing

Question 5

True or False: Shadow Testing is an example of a production variant of a model

Answer 5

True

Question 6

What is Sagemaker Neo?

Answer 6

It is a service that optimizes machine learning model execution for specific devices, allowing them to run on edge devices

Question 7

True or False: Sagemaker Neo is composed of a singular compiler that optimizes your code

Answer 7

False, it is composed of a compiler and a runtime

Question 8

What other AWS Service would you use with Sagemaker Neo to deploy a model to the Edge?

Answer 8

IoT Greengrass

Question 9

True or False: Sagemaker Neo optimized models can be deployed to an HTTP endpoint

Answer 9

False, only HTTPS

Question 10

True or False: There are no restirction to which instance a Sagemaker Neo optimized model can be deployed to

Answer 10

False, it must be deployed to the same type of instance as the one where it was optimized

Question 11

True or False: IoT Greengrass uses Lambda inference applications

Answer 11

True

Question 12

True or False: You can use KMS integrations to encrypt all the data on Sagemaker Notebooks and jobs

Answer 12

False, all notebook data can be encrypted, but on jobs only data /opt/ml/ and /tmp can be encrypted

Question 13

True or False: When running a training job, inter-node training communication may be optionally encrypted with no impacts on performance (AKA inter-container traffic encryption)

Answer 13

False, it can result in increased training cost and time for deep learning

Question 14

True or False: You can run Sagemaker jobs inside VPCs, but Notebooks are internet enabled by default. If this is disabled, your notebook needs an interface endpoint (PrivateLink) or NAT
Gateway, and allow outbound connections, for training and hosting to
work. The same is true for inference and training images.

Answer 14

True

Question 15

What on Sagemaker can Cloudwatch log and monitor?

Answer 15

-Ground Truth (active workers, how much they are doing)
-Instance health metrics
-Invocations and latency of endpoints

Question 16

What are some frameworks you can use to perform distributed training on Sagemaker?

Answer 16

Horovod and Parameter Servers

Question 17

True or False: Sagemaker does not support Spot Training

Answer 17

False

Question 18

What is Sagemaker Elastic Inference?

Answer 18

It is a feature that accelerates inference at a fraction of the cost of using a GPU

Question 19

True or False: Sagemaker Elastic Inference works by leeting you select instance types optimized for inference when you choose your instance

Answer 19

False, it allows you to use an accelerator alongside a CPU instance

Question 20

True or False: Elastic Inference cannot be applied to Sagemaker notebooks

Answer 20

False, it can

Question 21

What built-in Sagemaker algorithms work with Elastic Inference?

Answer 21

-Image Classification and Object Detection

Question 22

True or False: Elastic Inference works with both pre-built and custom MXNet, Tensorflow and PyTorch containers

Answer 22

True

Question 23

Which use cases are good for Serverless Inference?

Answer 23

Use cases where traffic is unpredictable

Question 24

What is Sagemaker Inference Recommender?

Answer 24

It is a Sagemaker feature that recommends the best instance type and configuration for your model deployments

Question 25

True or False: Sagemaker Inference Recommender automates load testing model tuning

Answer 25

True

Question 26

How does Sagemaker Inference Recommender work?

Answer 26

It benchmarks your model registered on Model Registry for various instace types and endpoint configurations and generates metrics for all of them

Question 27

True or False: Sagemaker Inference Recommender has 2 types of recommendation, Instance Recommendation and Endpoint Recommendation, and each of the need their own benchmarks

Answer 27

True

Question 28

How do Sagemaker Deployments interact with AZs?

Answer 28

Sagemaker automatically tries to distribute instances across AZs, but this only works if you have more than 1 instance and subnets configured in different AZs

Question 29

What is Amazon Sagemaker Operators for Kubernetes?

Answer 29

It is a framework that integrates Sagemaker with Kubernetes to facilitate MLOps

Question 30

True or False: Amazon Sagemaker Operators for Kubernetes enables Hybrid ML workflows (on-prem + cloud)

Answer 30

True

Question 31

What are Sagemaker Projects?

Answer 31

They are Sagemaker Studio's native MLOps solution with CI/CD.

Question 32

True or False: Sagemaker Projects uses code repositories for building and deploying solutions using Step-Functions

Answer 32

False, it uses Sagemaker Pipelines

Question 33

What is a Sagemaker inference pipeline?

Answer 33

It is linear sequence of 2-15 containers where the output of one container is piped to the next in a way as to make the whole inference work as a pipeline. Useful for dealing with pre-processing, post-processing, etc

Question 34

True or False: Sagemaker Inference Pipelines are compatible with Scikit-learn containers and SparkML, which can be run with Glue or EMR

Answer 34

True