AWS AI/ML and Analytics Services: SageMaker, Athena, Kinesis and More

Before memorizing individual services, learn the split that organizes the whole AI/ML category. AWS offers machine learning at two levels. At the bottom is Amazon SageMaker, a fully managed platform where data scientists and ML engineers build, train, and deploy their own custom machine learning models. SageMaker covers the entire ML lifecycle — preparing data, writing and running training jobs, tuning models, and hosting them behind endpoints for predictions. If a question mentions data scientists, custom models, training, or the phrase "machine learning workflow," the answer is SageMaker. AWS now calls this ML platform Amazon SageMaker AI (the name "Amazon SageMaker" also covers a newer unified data and AI platform); the exam guide still says Amazon SageMaker, and for CLF-C02 both names mean the build-your-own-model service.

Above SageMaker sits a family of pre-trained AI services. With these, you do not need to train anything. AWS has already built and trained the model; you simply call an API with your input — text, audio, an image, a document — and get a result back. No ML expertise is required, and that is precisely their selling point. A few of them also offer optional customization for specialised needs — Amazon Rekognition Custom Labels learns your own objects from labelled images, and Amazon Comprehend custom classification and custom entity recognition learn your own categories — but the defining fact is that they work out of the box with no model of your own. Every other AI service in scope for CLF-C02 belongs to this family: Lex, Kendra, Comprehend, Polly, Rekognition, Textract, Transcribe, and Translate.

This split is itself exam material. A question might describe "a company with no machine learning expertise that wants to add image analysis to its application" — the phrase no ML expertise steers you toward a pre-trained service (Rekognition), not SageMaker. Conversely, "a data science team needs a managed environment to train and deploy a custom fraud-detection model" is SageMaker, because the team is building something of its own.

Think of it this way: SageMaker is the workshop; the pre-trained services are finished appliances. CLF-C02 rarely goes deeper into SageMaker than this — you will not be asked about notebooks, instance types, or algorithms. You only need to recognize it as the build-your-own-ML platform.

Five pre-trained services handle human language, and the exam distinguishes them by the direction of the conversion. Amazon Transcribe converts speech to text: feed it audio — call recordings, meeting audio, video soundtracks — and it returns a transcript. Amazon Polly goes the opposite way, converting text to speech: give it written text and it returns lifelike spoken audio, used for voice assistants, audio versions of articles, and accessibility features. The names help — Polly the parrot speaks; Transcribe produces a transcript. Exams love to swap these two, so lock in the directions now.

Amazon Translate does exactly what its name says: it translates text between languages. Any scenario about localizing content, translating user reviews, or supporting a multilingual audience points here. There is no trick to it, which makes it an easy point.

Amazon Comprehend is the natural language processing (NLP) service. It reads text and extracts meaning: sentiment (is this review positive or negative?), entities (names of people, places, organizations), key phrases, and the dominant language. When a question says a company wants to "analyze customer feedback for sentiment" or "find common themes in support tickets," the answer is Comprehend. It analyzes text it is given — it does not transcribe, translate, or speak.

Amazon Lex builds conversational interfaces — chatbots and voice bots. It uses the same technology that powers Alexa: automatic speech recognition plus natural language understanding, so a bot can figure out what a user wants ("book a hotel room") and gather the details through dialogue. Trigger phrases: "chatbot," "virtual agent," "conversational interface," "interactive voice response." Lex handles the conversation; if the scenario is merely converting audio to text with no back-and-forth, that is Transcribe instead.

Amazon Rekognition analyzes images and videos. It detects objects and scenes ("this photo contains a dog and a bicycle"), recognizes and compares faces, reads text that appears inside a photo, identifies celebrities, and flags inappropriate content for moderation. The trigger words are visual: photos, video frames, facial recognition, content moderation, object detection. A social platform that wants to automatically block explicit uploads, or a security system matching faces against a watchlist, uses Rekognition.

Amazon Textract extracts text and structured data from scanned documents — PDFs, forms, tables, invoices, IDs. Simple OCR only gives you a blob of characters; Textract goes further and understands structure, returning form fields as key-value pairs ("Name: Jane Smith") and preserving table rows and columns. Any scenario about digitizing paperwork, automating invoice or loan-application processing, or "extracting data from scanned forms" is Textract. The dividing line with Rekognition matters: documents go to Textract, photos and videos go to Rekognition — even though both technically involve images.

Amazon Kendra is intelligent enterprise search. It comes with ready-made connectors that index an organization's scattered content — wikis, file shares, SharePoint, S3 buckets, help articles — and lets employees ask natural-language questions like "How many days of parental leave do we get?" and receive a specific answer extracted from the right document, not just a list of links. Machine learning powers its understanding of the question. Trigger phrases: "search across internal documents," "employees can't find information," "natural language search of company knowledge." Kendra is a finished search service you point at your content; contrast it with Amazon OpenSearch Service (below), a search engine developers build into their own applications. Kendra searches content that already exists; it does not create, transcribe, or translate anything.

Notice that every service in this section and the previous one is pre-trained. You configure and call them, and they are usable out of the box with AWS's pre-trained models (optional custom training, where it exists, is an add-on). That single fact eliminates SageMaker from most matching questions.

The exam guide names four analytics services explicitly, and together they form a pipeline you should be able to recite: move and prepare data (Glue), stream it in real time (Kinesis), query it (Athena), visualize it (QuickSight).

Amazon Athena is a serverless, interactive query service that runs standard SQL directly against data stored in Amazon S3. There are no clusters to launch and no data to load — you point Athena at files in S3 (CSV, JSON, Parquet, logs) and start querying, paying per query based on data scanned. Trigger phrases: "query data in S3 using SQL," "serverless," "no infrastructure to manage," "pay only for queries you run." Athena is ideal for ad hoc analysis of logs and data lakes.

AWS Glue is serverless ETL — extract, transform, load. It discovers data, catalogs its schema in the Glue Data Catalog, and runs jobs that clean, reshape, and move data between stores (for example, converting raw CSVs in S3 into optimized Parquet, or loading transformed data into Redshift). Trigger phrases: "prepare data for analytics," "ETL," "data catalog." Athena and Redshift Spectrum use the Glue Data Catalog to know what the data in S3 looks like, and EMR can be configured to use it too. That is the line between Glue and EMR: EMR can also run Spark transformations, but it is a big data framework service (you run clusters, or use EMR Serverless) that can use the Glue Data Catalog as its metastore (recommended when the catalog must persist or be shared across clusters) but offers no serverless, managed central catalog of its own — so a need for serverless ETL plus a central catalog points to Glue.

Amazon Kinesis is the family of services for real-time streaming data — collecting, processing, and analyzing data as it arrives: clickstreams, IoT sensor readings, application logs, live video. The single most reliable trigger word on the entire analytics list is "real time" (or "streaming"). If data must be acted on within seconds of being generated, the answer is Kinesis. Two members of the family are worth telling apart. Amazon Kinesis Data Streams retains records — 24 hours by default, extendable up to 365 days — so several consumer applications can each read the same records independently and you can replay older data, for example to backfill a consumer application added later. Amazon Data Firehose (formerly Kinesis Data Firehose) simply delivers streaming data to destinations such as S3, Redshift or OpenSearch Service; it is not something applications read from or replay.

Amazon QuickSight is the business intelligence (BI) service: it builds interactive dashboards and visualizations from your data sources (Athena, Redshift, S3, databases) and shares them with business users. Trigger phrases: "dashboards," "visualizations," "BI reports for management." QuickSight displays insight; it does not store, transform, or stream data. AWS now offers it as Amazon Quick Sight, the BI capability within Amazon Quick; the exam guide still uses the name Amazon QuickSight, and both names mean the same service.

Amazon Redshift is the AWS data warehouse: a fast, petabyte-scale system for running complex SQL analytics on large volumes of structured data, usually loaded into it. You run it as provisioned clusters or as Amazon Redshift Serverless (a workgroup with no clusters to manage, billed for the compute it uses), and Redshift Spectrum lets it also query data that stays in S3. A warehouse aggregates data from many sources — sales, marketing, operations — so analysts can run heavy reporting and historical analysis. Trigger phrases: "data warehouse," "petabyte-scale analytics," "complex queries across years of structured business data," and recurring dashboard and reporting queries with high concurrency. Keep Redshift firmly separated from transactional databases (Task 3.4 territory): a warehouse is for analysis, not for powering an application's day-to-day reads and writes.

Amazon EMR (Elastic MapReduce) runs managed open-source big data frameworks such as Apache Spark, Hadoop, Hive, and Presto. Instead of buying and configuring a cluster of servers for large-scale distributed data processing, you let EMR provision, run, and scale the cluster — or use EMR Serverless to run Spark and Hive jobs without managing clusters at all. Either way you pay for the compute your jobs use, and by default a cluster keeps its own Hive metastore, though EMR can be pointed at the Glue Data Catalog instead so table definitions persist and are shared. The trigger is the framework name: any question mentioning Spark or Hadoop answers EMR.

Amazon OpenSearch Service is a managed search and log-analytics engine (the successor to Amazon Elasticsearch Service). Two use cases dominate: powering full-text search inside an application (product search on an e-commerce site) and interactively searching and analyzing large volumes of logs for operational monitoring. Trigger phrases: "full-text search," "log analytics," "analyze application logs in near real time." Don't confuse it with Amazon CloudWatch Logs: CloudWatch Logs collects and queries logs for operational monitoring, but it is not a full-text search engine you build into a customer-facing application.

Two services need only a line each. Amazon MSK (Managed Streaming for Apache Kafka) is fully managed Apache Kafka — choose it when a question names Kafka explicitly; otherwise generic streaming questions answer Kinesis. AWS Data Exchange is a catalog where you find, subscribe to, and use third-party data sets — weather feeds, financial data, demographics — directly in AWS. Once you subscribe, the data and each update the provider publishes are delivered into your own AWS account — for example exported to S3 — where you use it with Athena, Glue, EMR, Redshift or SageMaker. Trigger phrase: "subscribe to third-party data."

Foundational exams reward instant recognition, so rehearse these two tables until each trigger phrase produces the service name without thought. Every row is a one-line identity. First, the AI/ML services.

Two habits make the AI/ML rows stick. First, identify the input and output: audio in, text out is Transcribe; text in, audio out is Polly; image in, labels out is Rekognition; document in, structured data out is Textract. Second, check whether the scenario involves building anything — only SageMaker builds; everything else is called. Now the analytics services.

A quick decision rule covers most analytics questions: where is the data and how fast do you need it? Data sitting in S3 that you want to query occasionally → Athena. Data arriving continuously that must be handled instantly → Kinesis. Data from many systems loaded centrally for heavy, repeated reporting → Redshift. Messy data that needs cleaning first → Glue. Results that executives need to see → QuickSight.

Most wrong answers in this task come from five look-alike pairs. Resolve each one with a single distinguishing question.

Athena vs Redshift. Both run SQL, and both can reach data in S3 (Redshift through Spectrum), so ask: what is the workload pattern?

"Occasionally query log files already in S3" → Athena. "Central warehouse for sustained enterprise reporting" → Redshift.

Transcribe vs Polly. Same domain, opposite directions. Transcribe: speech in, text out. Polly: text in, speech out. Read the question's input and output and the answer falls out.

Rekognition vs Textract. Both process visual input, so ask: photo or paperwork? Rekognition analyzes photos and videos — objects, faces, scenes, moderation. Textract processes scanned documents and forms, extracting text, key-value pairs, and tables. A scanned invoice is Textract even though it is technically an image.

Kendra vs OpenSearch Service. Both are "search," so ask: who is searching what, and how? Kendra is ML-powered enterprise search with ready-made connectors (SharePoint, wikis, file shares, S3) where employees ask natural-language questions across internal company documents and get answers extracted from them. OpenSearch is a search and log-analytics engine that developers build into their own applications — product search, log exploration.

AWS Glue vs Amazon EMR. Both can transform data with Spark, so ask: do you want serverless ETL with a catalog, or a big data framework? Glue is serverless ETL and owns the Glue Data Catalog. EMR runs Spark, Hadoop and similar frameworks on clusters (or EMR Serverless) and can use the Glue Data Catalog as its metastore, but has no serverless, managed central catalog of its own. Existing Spark or Hadoop jobs → EMR; serverless ETL plus a catalog → Glue.

Kinesis Data Streams vs Data Firehose vs SQS vs SNS. SQS and SNS belong to Task 3.8, but they appear as distractors here, so ask: what happens to a record after it is read?

CLF-C02 sometimes chains two or three services in a single scenario, and pipelines are also the best way to cement the matches. Walk through this one slowly.

A company records thousands of customer support calls every day. Management wants to know whether customer sentiment is improving over time and to display the trend on a dashboard for executives.

Break the requirement into tasks and match each one. First, the raw material is audio, but sentiment analysis works on text — so the recordings must be converted: speech to text is Amazon Transcribe. Second, the transcripts must be analyzed for sentiment — NLP on text is Amazon Comprehend. Third, the results must be visualized as a trend for executives — BI dashboards are Amazon QuickSight. The full pipeline: Transcribe → Comprehend → QuickSight, likely with the transcripts and results stored in S3 between stages.

Now stretch the same scenario to test your other matches. If the company wanted supervisors to see sentiment on calls while they are still in progress, the words "real time" bring Kinesis into the pipeline to stream the data. If analysts wanted to run occasional SQL queries over the transcript files sitting in S3 — "how many calls mentioned refunds last month?" — that is Athena. If the raw transcripts needed cleaning and reformatting before analysis, AWS Glue would run the ETL. And if the company instead wanted a chatbot to deflect simple support calls entirely, that is Amazon Lex.

Each stage is matched on its own, so the same method handles other chains. French training videos that need German subtitles are speech to text and then text to another language: Transcribe → Translate (Translate works on text, so it cannot take the audio directly). Scanned expense receipts whose amounts must appear in a finance team's charts are document extraction and then BI: Textract → QuickSight, usually with the extracted data stored in S3 and queried through Athena in between.

Practice decomposing every multi-service question this way: list the tasks in order, match each task to its one-line service identity, and ignore services whose identity was never invoked. The wrong options in these questions are services that do a plausible-sounding but different job — Polly instead of Transcribe, Redshift instead of Athena. Your trigger tables are the defense.