Machine learning and analytics services
The analytics services in the order data passes through them, and the pre-trained artificial intelligence services sorted by what you feed them, so a scenario can be answered without knowing anything about the models underneath.
Lesson 9 of 11 in objective 3. Cloud technology and services, part of AWS Certified Cloud Practitioner (CLF-C02).
Athena asks once; Redshift answers every morning
Amazon Athena runs SQL directly against files already sitting in Amazon S3, with no cluster to create and no loading step, and bills by the amount of data each query scans. That pricing is the reason it is the right answer to a compliance question about a year of logs that will probably never be asked again: the cost of asking once is the cost of one scan, and the cost of not asking is nothing at all.
Amazon Redshift is the opposite economics and a genuinely different tool: a data warehouse with columnar storage, excellent when the same analytical questions are asked over data you have deliberately loaded, and a poor answer when the alternative is a single query. Amazon EMR runs the managed big-data frameworks such as Spark and Hadoop, for work SQL cannot express. AWS Glue is the catalog that tells Athena what the files contain, plus the ETL that reshapes them. Amazon QuickSight is the business intelligence layer that draws results from a source such as Athena or Redshift — it never reads raw files itself, and an option offering it as the query engine is testing the layers.
A stream you read, or a delivery you do not have to
Two streaming shapes are worth telling apart. Amazon Kinesis Data Streams keeps an ordered, replayable stream that you write your own consumers against, which is what you want for real-time processing or for several independent readers of the same events. Amazon Data Firehose — the delivery service previously named Kinesis Data Firehose — receives records, buffers them by size or by time, optionally converts or compresses them, and writes them to a destination such as S3, Redshift or OpenSearch, with nothing for you to run or scale.
The requirement that separates them is almost always phrased as work you do not want to do. "Capture these events into S3 within a minute or two, with no cluster and no consumer application to build" is Firehose in one sentence. A queue is the tempting neighbour and is wrong for the same reason: something still has to poll it, batch the messages and write the files, and that something is exactly the application the team said it would not build.
Reach for a pre-trained service first
Sort the pre-trained services by their input and most scenario questions answer themselves. Images and video go to Amazon Rekognition. Documents and scans go to Amazon Textract, which is the one that understands a form's structure and can turn a scan into key-value pairs rather than a wall of words. Text goes to Amazon Comprehend for analysis or Amazon Translate for languages. Speech goes to Amazon Transcribe on the way in and Amazon Polly on the way out. Conversations go to Amazon Lex, and enterprise search to Amazon Kendra.
Amazon SageMaker sits underneath all of them as the build-your-own option: the platform for building, training and deploying a model when nothing pre-trained fits. That makes it the wrong answer to a stem that says the team has no machine learning expertise and does not want to train anything, which is how such stems are usually written. The habit worth forming is the order — pre-trained service first, custom model only when the catalogue has no answer.
Worth carrying in
- Amazon Athena
- Serverless SQL over files in S3. Billed by data scanned. The question you ask once.
- AWS Glue
- The data catalog that says what those files contain, plus managed ETL.
- Amazon Redshift
- The warehouse. Load data in, then ask the same questions every morning.
- Amazon EMR
- Managed Spark, Hadoop and friends, for work SQL cannot express.
- Amazon QuickSight
- Dashboards and business intelligence on top of a source such as Athena or Redshift.
- Amazon Kinesis Data Streams
- An ordered, replayable stream you write your own consumers against.
- Amazon Data Firehose
- Managed delivery of a stream into S3, Redshift or OpenSearch. Nothing to run.
- Amazon Rekognition
- Images and video: objects, scenes, faces, moderation.
- Amazon Textract
- Documents and scans: text, form fields and tables.
- Amazon Comprehend
- Natural language on text you already have: entities, key phrases, sentiment.
- Amazon Transcribe
- Speech to text. Amazon Polly is text to speech.
- Amazon Translate
- Machine translation between languages.
- Amazon Lex
- Conversational interfaces — the chatbot and voice-bot service.
- Amazon Kendra
- Enterprise search across an organisation's own documents.
- Amazon SageMaker
- Build, train and deploy your own models. The answer when nothing pre-trained fits.
What the exam does with this
- Asked once, over files already in S3, with nothing to stand up: Athena. Asked every morning over data you loaded: Redshift. The FREQUENCY in the stem decides it, not the size of the data.
- QuickSight draws the answer and never reads raw files. Offering it as the query engine is a test of whether the layers are separate in your head.
- Firehose lands a stream in storage with nothing for you to run; a Kinesis data stream is what you write consumers against. "No consumer application to build" is pointing at Firehose.
- Textract for a document, Rekognition for a photograph, Comprehend for text that already exists. A scanned form is a document, however much it looks like an image.
- SageMaker is wrong whenever the stem says the team has no machine learning staff and a pre-trained service already covers that input.
- Objective
- 3. Cloud technology and services
- Share of the exam
- 34% (the whole objective)
- Questions in this lesson
- 3
- Signed for by a person
- 0
Partly checked. None of the 3 questions here has been read against the cited source by a person. 3 questions have been checked against their cited clause by an automated pass — which is not the same thing, and is not a signature.
Only questions a person has signed for are used in mock exams here. That is the whole difference between the two kinds of checking above.
Questions in this lesson
- Compliance has asked a one-off question about a year of web server logs that are already sitting in Amazon S3 as compressed files. The answer is needed this week, the question is unlikely to be asked again, and nobody wants to stand up a cluster or load the data anywhere first. Which service answers it? machine-checked
- An insurer receives claim forms as scanned PDFs. It needs the printed values out of the form's fields — policy number, date, amount — as structured text it can put into a database. The team has no machine learning expertise and does not want to train anything. Which service should it start with? machine-checked
- Clickstream events arrive continuously from a busy website and must be captured and written into Amazon S3 within a minute or two of arriving, so that analysts can query them there. The team wants no cluster and no consumer application to build. Which service does this? machine-checked
Drill this lesson
A lesson is one sitting: the trainer draws a short run from these questions alone and spaces the ones you get wrong.
Practise Machine learning and analytics services
The rest of objective 3
- Ways in: console, CLI, SDKs and infrastructure as code
- Regions, Availability Zones and edge locations
- EC2 and choosing an instance type
- Containers, serverless, auto scaling and load balancing
- Relational, NoSQL and in-memory databases, and moving them
- VPCs, subnets, traffic controls and private connectivity
- Object storage and the S3 storage classes
- Block, file and hybrid storage, lifecycle rules and backup
- Machine learning and analytics services — you are here
- Messaging, integration and business applications
- Developer tools, end-user computing, front end and IoT