Application Integration, Analytics and Other In-Scope Services

Beginner⏱60 minutes📖49 min readLesson 11 of 16

Application Integration, Analytics and Other In-Scope Services

How AWS moves data between applications and how it turns raw data into answers: SQS queues and FIFO ordering, SNS topics and fanout, EventBridge rules and schedules, Step Functions workflows, the API Gateway front door with throttling, then the analytics stack - Athena, Glue, Kinesis, Firehose, OpenSearch, QuickSight, EMR and Redshift - closing with one-line coverage of every remaining in-scope service from Rekognition to CodePipeline, plus the lookalike services that are explicit distractors.

Domain 3 (Cloud Technology and Services, 34% of the score) hides two task statements that carry a huge, scattered surface area: Task 3.7 "Identify AWS artificial intelligence and machine learning (AI/ML) services and analytics services", with its skill statements naming Amazon SageMaker AI, Amazon Lex, Amazon Athena, Amazon Kinesis, AWS Glue and Amazon Quick Sight, and Task 3.8 "Identify services from other in-scope AWS service categories", which explicitly lists application integration (EventBridge, SNS, SQS), business applications (Connect, SES), customer enablement (AWS Support), developer tools (CodeBuild, CodePipeline, X-Ray), end-user computing (AppStream 2.0, WorkSpaces, WorkSpaces Secure Browser), frontend (Amplify) and IoT (IoT Core). Add Amazon API Gateway from the Networking row and this single lesson spans material the exam draws from two domains.

plaintext

====================================================================
 CLF-C02 LESSON 11 — INTEGRATION + ANALYTICS QUICK CARD
          (facts as of Oct 2026 — verify current before use)
====================================================================
 APPLICATION INTEGRATION
  SQS             pull queue, store-and-forward buffer
                  Standard : at-least-once, best-effort order
                  FIFO     : exactly-once processing, strict order
                  visibility timeout 30 s default (max 12 h)
                  retention 60 s - 14 days, max message 1 MiB
  SNS             push pub/sub topic -> N subscriptions
                  email / SMS / HTTP+S / Lambda / SQS
                  fanout = SNS topic -> N queues -> N workers
  EVENTBRIDGE     JSON events matched by rules -> targets
                  5 targets per rule - 300 rules per bus
                  default / custom / partner buses
                  AWS management events ingested free
  STEP FUNCTIONS  serverless workflow orchestration
                  Standard : up to 1 yr, $0.000025/transition
                  Express  : up to 5 min, $1.00 per million
 --------------------------------------------------------------------
 API GATEWAY       the "front door": REST / HTTP / WebSocket
                   10,000 RPS + 5,000 burst -> HTTP 429
                   REST $3.50/M - HTTP $1.00/M (as of Oct 2026)
 --------------------------------------------------------------------
 ANALYTICS
  ATHENA          serverless ANSI SQL over S3, $5.00/TB scanned
  GLUE            serverless ETL + Data Catalog, $0.44/DPU-hour
  KINESIS STREAMS ordered shards: 1 MB/s + 1,000 rec/s write
  DATA FIREHOSE   no shards: land data in S3/Redshift/OpenSearch
  OPENSEARCH      managed domains: logs, search, Dashboards
  QUICK SIGHT     BI dashboards, Reader $3/user/month
  EMR             managed Hadoop/Spark clusters, per-second
  REDSHIFT        petabyte-scale warehouse, from $0.543/hour
 --------------------------------------------------------------------
 ML (one line each) Rekognition images/video - Comprehend NLP -
   Textract documents - Transcribe speech to text - Polly text to
   speech - Translate - Lex chatbots - SageMaker AI build/train -
   deploy - Amazon Q generative-AI assistants
 OTHER IN-SCOPE    Connect contact center - SES email - CLI /
   CodeBuild / CodePipeline / X-Ray - IoT Core - Amplify -
   AppStream 2.0 / WorkSpaces / WorkSpaces Secure Browser -
   AWS Marketplace
====================================================================

ℹ️Note

Scope discipline. Every service presented as examinable below sits on the official CLF-C02 in-scope list or in a Domain 3/4 task statement (verified against the exam guide, 2026-10). Four services that look perfectly at home here are explicitly out of scope and appear only so you can eliminate them: Amazon AppFlow, Amazon Managed Streaming for Apache Kafka (Amazon MSK), Amazon WorkDocs and AWS AppFabric. Two things this research could not verify and therefore never teaches as fact: whether Amazon Bedrock (absent from the in-scope list) or other generative-AI services are examined, and the status of AWS CodeCommit (on neither the in-scope nor the out-of-scope list).

In this lesson you will:

read Amazon SQS as the decoupling answer and separate standard from FIFO delivery guarantees;
work the queue mechanics: visibility timeout, long polling, retention, dead-letter queues;
design an SNS fanout and say why it is SNS plus SQS, not SNS alone;
route event-driven architectures through EventBridge buses, rules, targets and schedules;
pick Step Functions Standard vs Express and price a workflow in state transitions;
use API Gateway as the front door: REST vs HTTP vs WebSocket, throttling, Lambda integration;
price an Athena query and a Glue job, and explain why partitioning pays;
separate Kinesis Data Streams from Data Firehose, then place OpenSearch, QuickSight, EMR and Redshift;
one-line every remaining in-scope service: ML suite, Connect, SES, DevTools, IoT Core, EUC, Amplify, Marketplace;
study five AWS-published case studies (Capital One, Paytm, Amazon Prime Video, NASA JPL and the NFL);
check what the 2026 service-list changes did to the names, categories and distractors this exam tests;
practise with 12 exam-style questions plus four interactive checks.

1. Amazon SQS: queues, decoupling, standard vs FIFO

1.1 What a queue actually decouples

Amazon SQS is a pull-based, store-and-forward buffer between a producer and a consumer. Without it the web tier calls the worker tier directly: the worker's outage becomes the web tier's outage, and neither side can scale independently. With a queue in between, the producer writes a message and returns; workers poll at their own rate and at their own count. AWS's decision guide describes SQS in one line worth memorising: "Messages are persisted … at-least-once delivery."

Property	Standard (default)	FIFO
Delivery guarantee	At-least-once: "a message may be delivered more than once"	Exactly-once processing
Ordering	Best-effort ordering only	Strict first-in, first-out order
Throughput (as of Oct 2026)	Near-unlimited API calls per second	300 messages/s; 3,000/s with batching; high-throughput mode up to 70,000/s
Duplicate handling	None: duplicates are expected	5-minute deduplication interval
Exam stem	"scale consumers independently", "may arrive twice, out of order"	"exactly once, in order"

Because standard queues deliberately hand you duplicates, AWS's own guidance is that "you must design your applications to be idempotent": a duplicate must not double-charge, double-ship or double-email. Idempotency is the consumer's job; the queue only promises persistence.

1.2 The mechanics that appear in question stems

Mechanic	Documented value (as of Oct 2026)	Why the exam cares
Visibility timeout	Default 30 s, range 0 s, 12 h	A received message is hidden, not deleted; if the worker dies or runs long, the message reappears, that is where duplicates come from
Long polling	Wait time 0 to 20 s	Waiting instead of hammering the queue cuts empty responses and request cost
Retention	Default 4 days; range 60 s: 1,209,600 s (14 days)	Unprocessed messages survive an outage for up to 14 days
Message size / delay	1 KiB: 1,024 KiB (1 MiB); delay up to 15 min	Put large payloads in S3 and pass the key
Dead-letter queue	A queue that holds repeatedly failing messages	Isolate poison messages instead of retrying forever

Pricing (as of Oct 2026, verify before use): 1 million requests per month free, then approximately $0.40 per million requests.

1.3 From point-to-point to a queue: what actually changes

Without a queue (point-to-point)	With Amazon SQS in the middle
Producer must know the consumer's address and be up when it calls	Producer writes to a Regional endpoint and returns immediately
A consumer outage is a producer outage	Consumers fail, restart and scale without touching the producer
Adding a consumer means redeploying or reconfiguring the producer	Add worker fleets; competing consumers share the same queue
Traffic spike = the producer must throttle itself	Queue absorbs the spike; workers drain it at their own rate
Retry logic is hand-written per caller	Visibility timeout, redelivery, long polling and DLQ are service features
One-to-one, tightly coupled in time	One-to-many possible (queue per consumer family), decoupled in time

💡 Tip: Use Ctrl + Mouse Wheel to zoom, or click the fullscreen icon

Worked example E1: choosing the FIFO tier. A payment stream must be processed exactly once and in order at 4,000 messages per second. Standard FIFO caps at 300/s, or 3,000/s with batching: still below the required 4,000/s, so the documented answer is high-throughput FIFO mode, up to 70,000 messages per second with batching. Had ordering not been required, standard queues would have absorbed the rate with no throughput ceiling to reason about.

📚 Did you know? The retention and visibility numbers explain most "duplicate message" stems on this exam: a message that is not deleted within the 30-second default visibility timeout becomes visible again, so duplicates are normal on standard queues; which is why AWS tells you to build idempotent consumers rather than promising you will never see a message twice (as of Oct 2026; verify current before use).

2.1 Push, not poll

Amazon Simple Notification Service is the push counterpart to SQS's pull: a publisher sends to a topic, and the topic fans the message out to every subscription. AWS describes it as a "push mechanism, eliminating the need to periodically check or 'poll'". Subscriptions can be Amazon SQS queues, AWS Lambda functions, HTTP/S endpoints, email, SMS and mobile push; which is why one publish can simultaneously page a human and feed a machine.

	Amazon SQS	Amazon SNS
Model	Pull, consumers poll a queue	Push, topic delivers to subscribers
Message shape	Competing consumers, one gets each message	Every subscription receives a copy
Documented guarantee	At-least-once (standard)	At-least-once for HTTP/S; exactly-once for Lambda and Amazon SQS subscribers
Failure handling	Visibility timeout + dead-letter queue	Retry policy, then an SQS dead-letter queue on the subscription
Scale (as of Oct 2026)	1 M requests/month free	100,000 Standard topics per account; up to 12.5 million subscriptions per topic
Billing unit	Per request	Each 64 KB chunk of a payload counts as one request

2.2 The fanout pattern

The canonical answer to "one alert must reach many independent consumers" is SNS → N SQS queues → N worker fleets. SNS replicates the message; each queue then absorbs its own retries, backlog and scaling, so a slow consumer never throttles the others. SNS alone is fine for light subscribers (email, SMS), but the pattern the guide trains is SNS with an SQS queue per consumer family.

💡 Tip: Use Ctrl + Mouse Wheel to zoom, or click the fullscreen icon

2.3 Subscription types at one-exam depth

Subscription	What it delivers	Typical stem
Amazon SQS queue	A copy into a queue that the consumer family drains	fanout with independent retries
AWS Lambda	Direct asynchronous invocation of a function	run code per event, no polling
HTTP/S	A POST to your endpoint	our own webhook receiver
Email	A message to a mailbox	page a human
SMS	A text message	alert an on-call phone
Mobile push	A notification to a device	reach an app user

Worked example E2, fanout arithmetic and SNS billing. One alarm publishes to six subscriptions (email, SMS, Lambda, three queues): SNS performs six deliveries from a single publish, and each queue retries independently. Now add billing, SNS counts each 64 KB chunk as one request, so a 150 KB payload bills as ceil(150 / 64) = 3 requests. Keep payloads small, or store the body in S3 and publish the reference.

⚠️ Push and pull are the fastest marks in Domain 3. "Consumers poll" ⇒ SQS; "a service pushes / one publish reaches many endpoints" ⇒ SNS. And remember: fanout = SNS + SQS, because only a queue gives each consumer an independent retry budget and a place to dead-letter its own failures.

3. Amazon EventBridge: event buses, rules and schedules

3.1 Buses, rules, targets

EventBridge carries JSON events. An event lands on a bus, a rule matches it by event pattern (or by schedule), and the rule invokes up to five targets. The three bus types are the examinable split:

Bus	What arrives there	Exam stem
Default	Events from AWS services; AWS management events are ingested by the event bus for free	"react to an EC2 state change"
Custom	JSON events you publish yourself	"our order service emits order.created"
Partner	Events from SaaS partners (Datadog, Shopify, Stripe and others)	"trigger on a third-party SaaS event"

Hard limits (as of Oct 2026): a single rule can send an event to up to five targets, and there are up to 300 rules per event bus. Billing counts each 64 KB chunk of a payload as one event; custom events are $1.00 per million, partner cross-account delivery $0.05 per million, and AWS Pipes $0.40 per million.

3.2 Schedules

Rules can also fire on cron or rate expressions, but AWS now labels scheduled rules a legacy feature and steers you to EventBridge Scheduler, which offers 14,000,000 free invocations per month before charging $1.00 per million. "Every night at 02:00 UTC" is a schedule question, and the current answer is EventBridge Scheduler.

💡 Tip: Use Ctrl + Mouse Wheel to zoom, or click the fullscreen icon

AWS publishes a three-way decision guide, and its wording is quotable exam material:

Service	What AWS's decision guide says	Persistence	Documented guarantee
Amazon SQS	"Messages are persisted"	Yes: stored until consumed and deleted	At-least-once delivery
Amazon SNS	"delivered to subscribers in real-time"	No: pushed and gone	Exactly-once for Lambda and Amazon SQS subscribers; at-least-once for HTTP/S
Amazon EventBridge	"Events do not persist; processed in real-time."	No: matched and invoked immediately	Content filtering, schedules, SaaS and cross-account routing

Event-driven architecture in one sentence: producers publish facts that happened, never RPC calls; consumers subscribe to the facts they care about. That inversion is why EventBridge questions usually mention content filtering, SaaS sources, cross-account routing or schedules: things SQS and SNS do not do.

📚 Did you know? EventBridge's cheapest line item is free: AWS management events are ingested by the event bus for free, and EventBridge Scheduler includes 14,000,000 free invocations per month (as of Oct 2026). The billing surprise instead is granularity: each 64 KB chunk of a payload is billed as one event, so a 200 KB event counts as four events (verify current before use).

4. AWS Step Functions: orchestrating multi-step workflows

4.1 Standard vs Express

Step Functions is serverless workflow orchestration written in Amazon States Language: it sequences Lambda, ECS, SQS, SNS, DynamoDB and Glue calls and adds retries, catch blocks, Parallel and Map states, and waits. It is not a queue and not a notifier; it is the thing that knows what step comes next.

	Standard	Express
Maximum duration	Up to one year	Up to five minutes
Execution model	Exactly-once workflow execution	At-least-once execution
Start rate	2,000 executions/s	100,000 executions/s
Billing (as of Oct 2026)	Per state transition: $0.000025 each in US East (N. Virginia), 4,000 free per month	$1.00 per million requests plus duration and memory
Exam stem	Long-running, human approval, audit trail	High-volume, short, fault-tolerant event processing

Worked example E3, pricing the same workload two ways. A workflow runs 1,000,005 state transitions in a month. Subtract the free allowance: 1,000,005 − 4,000 = 996,005; multiply by the unit price: 996,005 × $0.000025 ≈ $24.90 on the Standard workflow. If every execution instead finishes in seconds and tolerates at-least-once execution, the same 1,000,000 Express requests cost $1.00. Same business logic, a 25× price gap, the exam wants you to notice the unit: state transitions vs requests.

4.2 Choosing among the four integration services

💡 Tip: Use Ctrl + Mouse Wheel to zoom, or click the fullscreen icon

Question in the stem	Answer
Buffer work, let consumers scale independently, tolerate retries	Amazon SQS
One publish must reach email + SMS + several queues	Amazon SNS (fanout)
Content-filtered events, cron/rate schedules, SaaS or cross-account	Amazon EventBridge
A business process with ordered steps, retries and a days-long wait	AWS Step Functions

Amazon MQ is not the answer on this exam. Legacy AMQP/MQTT broker migration is not on the CLF-C02 in-scope list, so a "we need an enterprise message broker" stem still resolves to one of the four services above.

4.3 Worked example E8: walking a claims approval end to end

Step	Service	Why that service
1. Validate the submission	AWS Lambda	Short, stateless code per claim
2. Check the text for fraud signals	Amazon Comprehend	Pre-trained NLP, no ML expertise needed
3. Wait for a human decision (hours to days)	Step Functions (Standard)	Executions may run up to one year with a task token for the approver
4. Notify the applicant	Amazon SNS	One publish reaches email and the internal queue
5. Store the result	Amazon DynamoDB	Managed key-value persistence for the record
6. Retry a step that fails three times	Step Functions retry/catch	Built into Amazon States Language rather than hand-written

The whole point of the example: Step Functions owns the sequence and the waits, Lambda owns the single steps, SNS owns the notification. A question that asks "what waits three days for a human?" is asking for Step Functions Standard, not for an SQS delay (15 minutes maximum) and not for a 5-minute Express execution.

5. Amazon API Gateway: the front door

5.1 Three API types

AWS positions API Gateway as the "front door" for back-end services, handling "traffic management … authorization and access control, throttling, monitoring, and API version management" in front of Lambda, EC2 or any HTTP endpoint.

Type	Choose it when the stem says	Extras (as of Oct 2026)
REST API	API keys, per-client throttling, AWS WAF integration, private API endpoints, caching	Full-featured, $3.50 per million calls
HTTP API	Cheapest minimal proxy to Lambda or an HTTP service, JWT authorizers	$1.00 per million calls
WebSocket API	Persistent two-way connections: chat, live feeds, collaborative apps	Connect, disconnect, route to the back end

Free tier (12 months, as of Oct 2026): 1,000,000 API calls per month for REST APIs and 1,000,000 per month for HTTP APIs.

5.2 Throttling

By default throttling is applied per account per Region: 10,000 requests per second (RPS) with an additional burst capacity bucket of 5,000 requests (some accounts are documented at 2,500 RPS / 1,250 burst: verify your own numbers). Exceed it and callers receive HTTP 429 Too Many Requests. AWS is explicit that these are "not hard limits" enforced on a "best-effort basis", and that usage-plan quotas are also best-effort.

Worked example E4: REST vs HTTP on the same traffic. A busy month produces 5,000,000 API calls; the free tier absorbs the first 1,000,000 (first 12 months), leaving 4,000,000 billable. On REST: 4 × $3.50 = $14.00. On HTTP: 4 × $1.00 = $4.00. Without the free tier the totals are $17.50 and $5.00. Feature for feature the answer may still be REST: API keys, per-client quotas, WAF and private endpoints are REST capabilities, but the pricing stem always resolves the same way.

💡 Tip: Use Ctrl + Mouse Wheel to zoom, or click the fullscreen icon

5.3 What the front door actually handles

The exam asks you to identify the responsibilities, not to configure them. In AWS's own list, API Gateway provides the "front door" experience for back-end services: traffic management, authorization and access control, throttling, monitoring, and API version management: in front of Lambda, EC2 instances or any HTTP endpoint.

Responsibility	In scope for you on this exam
Traffic management	Know that 10,000 RPS + 5,000 burst is the per-account, per-Region default
Authorization and access control	Choose the service and the API type: API keys and per-client quotas are REST features
Throttling	Excess requests ⇒ HTTP 429 Too Many Requests; usage-plan quotas are best-effort
Monitoring	Listed as a Gateway responsibility: the exam asks you to recognise it, never to configure it
API version management	Also Gateway-owned; identify it as a front-door capability
Lambda integration	Clients call Gateway, which invokes Lambda, EC2 or any HTTP endpoint

⚠️ API keys are not authentication for your users. On this exam, "API keys + per-client throttling + AWS WAF + private endpoints" is the signature of a REST API; the cheapest Lambda proxy is an HTTP API; and a 429 always means "you exceeded the throttle", never "the API is broken".

6. Serverless analytics on S3: Amazon Athena and AWS Glue

6.1 Amazon Athena: SQL without a cluster

Athena is serverless interactive SQL over data already in Amazon S3: "run ad-hoc queries using ANSI SQL, without the need to aggregate or load the data into Athena". There is no cluster to launch and no load-before-you-can-ask; you point it at S3, define tables (usually via the Glue Data Catalog) and query.

Pricing (as of Oct 2026): $5.00 per terabyte of data scanned, with a 10 MB minimum per query; DDL statements and failed queries cost $0.
Why query design is a cost decision: AWS documents that compression, partitioning and columnar formats can save 30% to 90% per query.
When not to use it: for "lots of joins across lots of very large tables", AWS says choose Amazon Redshift.

Worked example E5: three ways to pay for the same 3 TB.

Query design	Data scanned	Cost (as of Oct 2026)
Raw uncompressed data, full scan	3 TB	3 × $5.00 = $15.00
GZIP at roughly 3:1	~1 TB	$5.00
Columnar + partitioned, query reads 1 of 4 columns	~0.25 TB	$1.25
`CREATE TABLE` (DDL) or a query that fails	:	$0.00
A query that scans only 3 MB	billed at the 10 MB minimum	minimum charge

The pattern to memorise: Athena charges for bytes read, so the cheapest query is the one that reads the least: prune columns, filter on partition keys, compress and convert to Parquet or ORC.

6.2 AWS Glue: the catalog and the ETL engine

AWS Glue is a serverless ETL and data-catalog service. Crawlers scan data stores and write table definitions (schema, partitions, location) into the Glue Data Catalog, which is what lets Athena, Amazon Redshift and Amazon EMR treat S3 objects as tables. ETL jobs run Spark or Python to clean, deduplicate and format.

Pricing (as of Oct 2026): $0.44 per DPU-hour; 1 DPU = 4 vCPU and 16 GB of memory; ETL jobs run with a minimum of 2 DPUs (default 10); crawlers have a 10-minute minimum; the first million objects stored are free in the catalog.

Worked example E6, pricing a crawl-and-clean job. A job runs 15 minutes on 6 DPUs: 6 × (15/60) hour × $0.44 = 6 × 0.25 × 0.44 = $0.66. Double the DPU count and the price doubles, DPUs are the unit, wall-clock time is the multiplier.

The three-service sentence the exam keeps testing: Glue catalogs and prepares the data, Athena answers SQL questions about it, QuickSight draws the dashboard. Glue never answers a query, Athena never runs a scheduled Spark transformation, and QuickSight never extracts text from a document.

📚 Did you know? Athena's 10 MB minimum per query exists so that thousands of tiny queries still produce a billable unit, and it also means DDL and failed queries are free. A team that runs a nightly DROP/CREATE cycle pays nothing for the DDL and everything for the SELECT that scans the table, which is why partitioning and column pruning are the real cost controls (Athena pricing page, as of Oct 2026; verify current before use).

7. Streaming, search and BI: Kinesis, Firehose, OpenSearch, QuickSight, EMR, Redshift

7.1 Amazon Kinesis Data Streams vs Amazon Data Firehose

This is the most reliable service-selection pair in the analytics domain.

	Amazon Kinesis Data Streams	Amazon Data Firehose
What it is	An ordered log split into shards	A managed delivery pipe into destinations
Capacity unit	Shard: 1 MB/s and 1,000 records/s writes, 2 MB/s reads	No shards to size or manage
Retention / replay	24 hours by default, up to 365 days: replay is possible	Near-real-time batching, no replay window
Destinations	Your own consumers: Lambda, Kinesis Client Library, Firehose	S3, Amazon Redshift, OpenSearch Service, Splunk, Snowflake
Code	You write the consumer	No consumer code; optional Lambda transform; Parquet/ORC conversion
Naming	,	Renamed from Amazon Kinesis Data Firehose to Amazon Data Firehose in February 2024, old names still appear in older material

Worked example E7: shard sizing. A clickstream emits 4,000 records per second averaging 10 KB each: throughput = 4,000 × 10 KB = 40,000 KB/s = 40 MB/s. Divide by the per-shard write limit: 40 ÷ 1 = 40 shards on bandwidth, and 4,000 ÷ 1,000 = 40 shards on record rate. Both constraints agree, so 40 shards. Had the requirement been "land it in S3 with no consumer code", the answer would be Data Firehose and the shard arithmetic would disappear entirely.

7.2 OpenSearch, QuickSight, EMR, Redshift

Service	One-exam-depth job	Key numbers (as of Oct 2026)
Amazon OpenSearch Service	Managed domains (= clusters) for log analytics, real-time application monitoring and clickstream analysis, plus search and OpenSearch Dashboards (the Kibana successor)	Renamed from Elasticsearch Service on 8 September 2021; instance prices vary by deployment: see the pricing page
Amazon Quick Sight	BI: visualizations and dashboards over S3, Athena, Redshift and SaaS data; SPICE is the in-memory engine versus direct query	Reader $3 per user per month; 10 GB of SPICE with every provisioned Author (extra $0.38/GB-month); a session is a 30-minute period
Amazon EMR	Managed Hadoop / Spark / Hive / Presto clusters on EC2 (or EKS or serverless) for big-data processing	Per-second billing with a one-minute minimum; you pay EC2 + EBS + the EMR uplift
Amazon Redshift	Managed petabyte-scale data warehouse: columnar, massively parallel SQL for complex joins and BI	Provisioned from $0.543 per hour; Serverless from $1.50 per hour; Spectrum queries S3 at $5 per TB

The three-way tie-breaker: ad-hoc SQL on S3, no cluster ⇒ Athena · complex joins across very large tables, a persistent warehouse ⇒ Redshift · run Spark/Hadoop yourself, open-source libraries required ⇒ EMR. Executive dashboards on top of any of them ⇒ Quick Sight; logs and search UI ⇒ OpenSearch Service.

Question in the stem	Athena	Redshift	EMR
What you operate	Nothing: serverless endpoint	Managed warehouse (or Serverless)	A cluster of open-source frameworks
Billing unit (as of Oct 2026)	$5.00 per TB scanned	From $0.543/h provisioned, $1.50/h Serverless	EC2 + EBS + EMR uplift, per second, 1-minute minimum
Best at	One-off questions over S3	Persistent BI, lots of joins, petabyte scale	Spark/Hive/Presto jobs you already run
Worst at	Repeated heavy joins across huge tables	Paying for idle capacity on a tiny query	Occasional ad-hoc SQL (cluster startup + cost)
Catalog dependency	Reads the Glue Data Catalog	Spectrum reads S3 via catalog	Reads Hive metastore / Glue catalog

Worked example E9: QuickSight subscription arithmetic. Five analysts need view access and two authors create dashboards (as of Oct 2026): 5 Readers × $3 per user per month = $15.00. Each provisioned Author includes 10 GB of SPICE; if the two authors together need 30 GB: 20 GB beyond the included 20 GB: the extra costs 20 × $0.38 = $7.60 per month. Note also that a session is a 30-minute period, so a Reader who opens five dashboards inside one half-hour window still counts as one session's worth of access.

💡 Tip: Use Ctrl + Mouse Wheel to zoom, or click the fullscreen icon

7.3 Amazon AppFlow: the lookalike that is out of scope

Amazon AppFlow moves data between SaaS applications and AWS stores, which sounds exactly like this lesson's topic, and it is explicitly on the out-of-scope list. So are Amazon MSK (managed Kafka), AWS Clean Rooms, AWS Data Exchange and Amazon DataZone in analytics, plus AWS AppFabric and Amazon Simple Workflow Service (SWF) under application integration. Out of scope does not mean nonexistent: it means they can only ever be a distractor, never the correct answer.

7.4 The analytics decision tree

💡 Tip: Use Ctrl + Mouse Wheel to zoom, or click the fullscreen icon

8. Every remaining in-scope service, one line each

8.1 The machine learning suite (Task 3.7)

AWS groups these as purpose-built AI services, pre-trained APIs you call with your data, requiring no ML expertise, versus Amazon SageMaker AI, where you build, train and deploy your own models.

Service	One line for the exam
Amazon Rekognition	Computer vision: label, detect and analyse images and video
Amazon Comprehend	Natural language processing: sentiment, entities, key phrases and PII in text
Amazon Textract	Extracts text, handwriting and data from scanned documents into structured fields
Amazon Transcribe	Speech to text: call recordings, subtitles
Amazon Polly	Text to lifelike speech
Amazon Translate	Neural machine translation between languages
Amazon Lex	Build conversational interfaces (chatbots and voice agents): the technology that powers Amazon Connect
Amazon SageMaker AI	Fully managed ML: "build, train, and deploy ML models into a production-ready hosted environment"
Amazon Q	Generative-AI powered assistants (Q Business, Q Developer); AWS documents them as "powered by Amazon Bedrock"

Tiering rule: "no ML experience, just call a pre-trained API" ⇒ Rekognition / Comprehend / Textract / Transcribe / Translate / Polly / Lex · "build, train and deploy my own model" ⇒ SageMaker AI · "a generative-AI assistant for employees or developers" ⇒ Amazon Q. Never swap the data type: images ⇒ Rekognition, documents ⇒ Textract, text/NLP ⇒ Comprehend, audio ⇒ Transcribe and Polly. Out-of-scope distractors here: Amazon Personalize, Amazon Fraud Detector, Amazon Lookout for Metrics and AWS Panorama.

8.2 Business applications, developer tools, IoT and the rest (Task 3.8)

Category	In-scope service	One line for the exam
Business application	Amazon Connect	Cloud contact center: voice, chat and tasks with intelligent routing and real-time agent performance tracking; pay only for usage, no per-seat fee
Business application	Amazon SES	Scalable email from within any application: transactional, marketing and mass mail, via API or SMTP
Developer tools	AWS CLI	Free command line for AWS services: the develop and deploy side of task 3.8
Developer tools	AWS CodeBuild	"Compiles your source code, runs unit tests, and produces artifacts": billed per build minute, 100 free minutes per month
Developer tools	AWS CodePipeline	Managed release workflow: model the steps, run the stages: V1 $1.00 per pipeline per month, V2 $0.002 per action-minute
Developer tools	AWS X-Ray	Distributed tracing: segments → traces → a service graph; trace data retained 30 days
IoT	AWS IoT Core	Managed gateway that connects devices over MQTT or HTTP: the manage IoT devices answer
Frontend	AWS Amplify	Create and deploy frontend web and mobile applications
End-user computing	Amazon AppStream 2.0	Stream desktop applications to any browser or device without rewriting code
End-user computing	Amazon WorkSpaces	Fully managed, cloud-native persistent virtual desktops (VDI / DaaS)
End-user computing	Amazon WorkSpaces Secure Browser	Hosted browser where "no HTML, document object model (DOM), or sensitive company data is transmitted to the local machine"
Cloud financial management	AWS Marketplace	"Curated digital catalog" to find, buy, deploy and manage third-party software, data and services: hourly, monthly, annual and BYOL, one bill, Private Marketplace, entitlements via License Manager
Customer enablement	AWS Support	The customer-enablement bullet of task 3.8; plan tiers belong to Domain 4

The end-user-computing stem, verbatim: the exam asks for "the services that can present the output of virtual machines (VMs) on end-user machines"; that is AppStream 2.0 (applications), WorkSpaces (full desktops) and WorkSpaces Secure Browser (browser only).

Naming traps straight from the in-scope list: the guide spells it "Amazon Quick Sight" (with a space) and "Amazon SageMaker AI" (never bare "SageMaker"). AWS's Connect documentation page also now uses the label "Amazon Connect Customer" for the contact-center product: tying that wording to exam phrasing is this research's inference, not a verified exam statement.

8.3 Summary table; this lesson's full service surface

Category	In-scope services covered here	Depth expected
Application integration	Amazon SQS · Amazon SNS · Amazon EventBridge · AWS Step Functions	Choose by pattern; know guarantees and limits
Networking (API)	Amazon API Gateway	REST vs HTTP vs WebSocket, throttling
Analytics	Amazon Athena · Amazon EMR · AWS Glue · Amazon Kinesis · Amazon OpenSearch Service · Amazon Quick Sight · Amazon Redshift	Choose by job-to-be-done; know the billing unit
Machine learning	Amazon Comprehend · Amazon Lex · Amazon Polly · Amazon Q · Amazon Rekognition · Amazon SageMaker AI · Amazon Textract · Amazon Transcribe · Amazon Translate	One line each: data type in, data type out
Business applications	Amazon Connect · Amazon SES	Contact center vs email at scale
Developer tools	AWS CLI · AWS CodeBuild · AWS CodePipeline · AWS X-Ray	Compile, release, trace
Frontend / IoT	AWS Amplify · AWS IoT Core	Deploy frontends; manage devices
End-user computing	Amazon AppStream 2.0 · Amazon WorkSpaces · Amazon WorkSpaces Secure Browser	Present VM output on end-user machines
Cloud financial mgmt / enablement	AWS Marketplace · AWS Support	Third-party catalog; support options
Owned by other lessons	Compute, Storage, Databases, Networking and Content Delivery, Security and Identity, Management and Governance, Migration, Billing and pricing	Do not re-study them here

8.4 Stem → service cheat sheet

If the stem says …	The answer is …
"Can arrive twice and out of order; scale consumers independently"	Amazon SQS (standard)
"Exactly once, in strict order"	Amazon SQS (FIFO)
"One alert to email, SMS and three queues"	Amazon SNS
"SaaS vendor events plus a nightly cron"	Amazon EventBridge / EventBridge Scheduler
"Human approval that waits three days"	AWS Step Functions (Standard)
"Throttle an API, API keys, per-client quotas, WAF"	Amazon API Gateway (REST)
"SQL over S3, pay per query"	Amazon Athena
"Populate a metadata catalog from S3"	AWS Glue
"Stream to S3 with no consumer code"	Amazon Data Firehose
"Ordered stream with replay and shards"	Amazon Kinesis Data Streams
"Executive dashboards"	Amazon Quick Sight
"Sentiment of support tickets"	Amazon Comprehend
"Extract fields from scanned forms"	Amazon Textract
"Build, train and deploy my own model"	Amazon SageMaker AI
"Compile and run unit tests in CI"	AWS CodeBuild
"Present VM output on end-user machines"	AppStream 2.0 / WorkSpaces / Secure Browser

Out-of-scope services that look in-scope (never the correct answer):

Amazon AppFlow
AWS Clean Rooms
AWS Data Exchange
Amazon DataZone
Amazon MSK
AWS AppFabric
Amazon Simple Workflow Service
Amazon WorkDocs
AWS CodeDeploy
AWS CodeArtifact
Amazon CodeGuru
AWS Application Composer
AWS CloudShell
AWS Device Farm
Amazon Personalize
Amazon Fraud Detector
Amazon Lookout for Metrics
AWS Panorama
all AWS Elemental Media Services
Amazon GameLift
AWS RoboMaker
AWS IoT Greengrass
AWS IoT Device Defender
📚 Did you know? Two services in this lesson were renamed rather than replaced: Amazon Data Firehose dropped "Kinesis Data" from its name in February 2024, and Amazon OpenSearch Service was renamed from Elasticsearch Service on 8 September 2021. Older questions, blogs and third-party tutorials still use both old names, so a stem mentioning "Elasticsearch Service" or "Kinesis Data Firehose" is testing the same in-scope service (accessed Oct 2026).

2026 Updates (as of October 2026)

What changed in the guide itself. Every bullet below comes from the current CLF-C02 in-scope and out-of-scope service lists and the launch-era Version 1.0 CLF-C02 guide, all accessed 2026-10. Prices and limits still drift, so re-check before you sit the exam.

The service list was rebuilt. In-scope entries fell 128 → 111 while the explicit negative list grew 11 → 55 out-of-scope entries: for Domain 3 the biggest "what changed" is that AWS now writes down most of the distractors this lesson used to warn you about from memory.

Three re-spellings make old names wrong: Amazon SageMaker → Amazon SageMaker AI, Amazon WorkSpaces Web → Amazon WorkSpaces Secure Browser, and the in-scope list prints the BI service as "Amazon Quick Sight" (two words) while the product name is Amazon QuickSight: whether the guide's spacing is intentional could not be verified.

The Machine Learning row moved from "build" to "use": Amazon Q was added, Amazon Kendra was removed from the in-scope list, and Amazon Bedrock appears on neither the in-scope nor the out-of-scope list (both lists checked 2026-10), so do not memorise Bedrock as in-scope, and treat whether generative-AI content is examined as unverified.

Developer Tools collapsed from 11 in-scope entries to 4, AWS CLI, AWS CodeBuild, AWS CodePipeline, AWS X-Ray, exactly the four covered in section 8.2 of this lesson; anything still listing CodeStar, Cloud9 or a longer Developer Tools roster is quoting a superseded guide.

Formerly in-scope integration and analytics names are now distractors: Amazon MSK and AWS Data Exchange moved from the in-scope list to the out-of-scope list, AWS CodeDeploy, AWS CodeArtifact and AWS IoT Greengrass are out of scope too, and AWS AppSync was removed from the in-scope list; AWS Wavelength is explicitly out of scope while AWS Local Zones appears on neither list.

Category rename: Customer Engagement → Customer Enablement, which now contains only AWS Support, so AWS Activate, AWS IQ and AWS Managed Services are no longer in-scope answers anywhere on this exam.

Sources: AWS Certified Cloud Practitioner (CLF-C02) in-scope services list, out-of-scope services list and exam guide, accessed 2026-10; baseline = launch-era Version 1.0 CLF-C02 exam guide (accessed 2026-10).

📚 Did you know? The largest 2026 change to this lesson is not a price at all: the in-scope list shrank from 128 to 111 entries while the explicit out-of-scope list grew from 11 to 55, and the Developer Tools category alone went 11 → 4. Notes built on the launch-era Version 1.0 guide are therefore quoting a superseded document, which is why every "never the correct answer" list in this lesson was re-checked against the live lists (accessed 2026-10).

Real-World Case Studies

AWS publishes what these patterns look like in production. Every figure below is customer- or AWS-claimed and unaudited, quoted with its source so you can check it: the examinable point is the pattern (which integration service did the work, which analytics engine was chosen, which number moved), not the marketing.

Case A, Capital One: serverless and application integration at bank scale

Element	Detail
Industry / context	Fortune 100 bank, regulated: exit eight on-premises data centers and rebuild for real-time, ML-driven products
AWS services named	30+ services including Amazon Connect, Amazon SES, AWS Lambda, Amazon ECS and Fargate, AWS SAM, AWS Step Functions and AWS Glue
Headline outcomes (AWS-published, customer-claimed)	80% of roughly 2,000 applications built cloud-native; development-environment lead time 3 months → minutes; disaster-recovery test time −70%; incident resolution and transaction errors −50%; more than a third of applications serverless
Integration detail	One application cut cost by 90% on AWS Lambda, and checks ran 80% faster with AWS Step Functions
Physical tail	103 tons of copper and steel recycled from the eight data centers it exited
Source	aws.amazon.com/solutions/case-studies/capital-one-all-in-on-aws and …/capital-one-lambda-ecs-case-study (accessed Oct 2026)

Exam lesson: this is what task 3.8 looks like in production: Step Functions coordinating the steps, Lambda running the steps, SES sending the mail, Connect taking the calls, Glue preparing the data. The exam does not ask you to reproduce the numbers; it asks you to name the service that does each job.

Case B, Paytm: Graviton economics applied to EMR analytics

Element	Detail
Industry / context	A payments platform with a large Amazon EC2 and Amazon EMR analytics estate on AWS
AWS services named	Amazon EMR (analytics), Amazon EC2 on Graviton: with a MySQL database upgrade that had to happen first
Headline outcomes (AWS-published, customer-claimed)	−35% compute cost, 60% of EC2 running on Graviton, and 30 to 35% savings on Amazon EMR
Analytics lesson	EMR bills the instances it runs on, so the compute purchasing decision (Graviton price-performance) is part of the analytics cost conversation
Source	aws.amazon.com/solutions/case-studies/paytm-graviton-case-study (accessed Oct 2026)

Exam lesson: Amazon EMR has no flat standalone fee in this story; it bills a per-second rate (one-minute minimum) on top of EC2 and EBS, so a 30 to 35% EMR saving is achieved by changing the instances underneath, not by flipping an EMR setting. Watch the trap: Graviton is a price-performance choice that needs compatibility work (Paytm had to upgrade its database first), not free money.

Value pattern	Evidence	Underlying principle
Managed beats DIY	Capital One: dev environment 3 months → minutes; −90% cost on one app	The Well-Architected answer is least operational overhead
Integration services carry the load	Capital One: Step Functions, Lambda, SES, Connect, Glue	Task 3.8's services are production plumbing, not lab exercises
Analytics cost follows compute cost	Paytm: 30 to 35% EMR savings via Graviton	EMR = per-second instance time + EMR uplift; right-size the engine
Every percentage is a customer result	"−90%", "30 to 35%", "−70%"	Never write "AWS guarantees X%"

Case C, Amazon Prime Video: integration at a global live-sports peak

Element	Detail
Industry / context	Live sports streaming, where AWS quotes its own delivery team: "With live sporting events, reliability and low latency are absolutely critical because every lost second negatively impacts viewers"
AWS services named	six AWS Regions, Amazon CloudFront, Amazon DynamoDB, Amazon EC2, AWS Elemental MediaTailor, multi-AZ and Regional failover
Headline outcomes (AWS-published, customer-claimed)	11 NFL games streamed to 18.4 million fans in 224 countries and territories during the 2017 NFL regular season; 300,000+ clients polling per ad break; DynamoDB partitions doubled from the console at ad breaks
Integration lesson	Hundreds of thousands of polling clients per ad break is exactly the traffic this exam's push/pull split exists to remove: SQS long polling (0 to 20 s) and SNS push are the in-scope answers to "stop hammering an endpoint"
Exam caveat	AWS Elemental / MediaTailor sits in the Media Services, which are explicitly out of scope: name only the in-scope services the stem asks for
Source	aws.amazon.com/solutions/case-studies/amazon-prime-video (accessed Oct 2026)

Exam lesson: the figure to carry into the exam room is not 18.4 million, it is 300,000+ clients polling per ad break. Any stem about many clients repeatedly checking for new state is a push-versus-pull question: SNS pushes, SQS consumers poll (and long-poll up to 20 seconds precisely to cut that traffic), and EventBridge matches events with nobody polling at all. The second lesson is elasticity: DynamoDB partitions were doubled from the console during the game, not after a procurement cycle, that is the Domain 1 agility story told through an integration service.

Case D, NASA JPL: telemetry at analytics scale

Element	Detail
Industry / context	NASA's Jet Propulsion Laboratory (JPL) and the Perseverance rover: AWS's Public Sector Blog (2021-02-22) calls it "the first planetary NASA mission, with mission-critical communication and transfer of telemetry data in the cloud"
AWS services named	EC2 Auto Scaling mixing Spot, On-Demand and Capacity Reservations; the 2012-era NASA story on AWS also names SWF, CloudFront and RDS
Headline outcomes (AWS-published, customer-claimed)	~4.4 TB of downlinked data daily, generating up to 70 TB of final data products; Spot at "up to a 90 percent discount compared to Amazon EC2 On-Demand pricing"; NASA separately reported "almost a million dollars in cost savings each year" (AWS News Blog, 2012-06-11)
Analytics lesson	Ingest → process → data products is the pipeline shape taught here with Kinesis Data Streams, Data Firehose, Glue and Athena; the case study answers only the compute half of the story
Scope trap	Amazon Simple Workflow Service (SWF), the orchestration service in NASA's 2012 story, is now explicitly out of scope; on this exam the workflow answer is always AWS Step Functions
Source	aws.amazon.com/solutions/case-studies/nasa-jpl-spot-case-study; aws.amazon.com/blogs/publicsector (2021-02-22); aws.amazon.com/blogs/aws (2012-06-11): all accessed Oct 2026

Exam lesson: two separable questions hide inside this story. "How do you process terabytes a day?" → the analytics stack: capture the ordered stream (Kinesis Data Streams), land it without consumer code (Data Firehose), catalog and clean it (Glue), ask questions (Athena), warehouse it (Redshift), chart it (Quick Sight). "How do you pay for the compute underneath?" → pricing models: Spot for fault-tolerant, interruptible batch (documented up to 90% off On-Demand) with On-Demand and Capacity Reservations for the rest. Never merge the halves: Spot never answers an analytics-selection stem, and Athena never answers a pricing-model stem.

Case E, NFL: a fan-facing analytics product

Element	Detail
Industry / context	The NFL's fan-facing digital product, built on AWS
Headline outcomes (AWS-published, customer-claimed)	1 million+ fans; updates every 3 minutes; 500 million+ data points per season
Analytics lesson	A season of 500 M+ data points refreshed every three minutes is a requirement shape: ordered capture, near-real-time delivery, then a presentable view
Source	aws.amazon.com/solutions/case-studies/nfl-case-study (accessed Oct 2026)

Exam lesson: read the stem's cadence, never its marketing. "Ordered, replayable, updates every few minutes" ⇒ Amazon Kinesis Data Streams (shards, 24 hours default retention up to 365 days); "land it in S3 with no consumer code" ⇒ Amazon Data Firehose; "charts and dashboards for a million fans" ⇒ Amazon Quick Sight; "SQL questions over what landed in S3" ⇒ Amazon Athena. The case-study page is the source for the numbers only: the service selection is yours, and it must come from the in-scope list.

What cases C, E add

Value pattern	Evidence	Underlying principle
Push beats polling at scale	Prime Video: 300,000+ clients polling per ad break	SQS long polling (0 to 20 s) and SNS push exist to cut exactly that traffic
The analytics pipeline shape repeats	NASA JPL: 4.4 TB/day → up to 70 TB/day of data products; NFL: 500 M+ data points per season	Capture → land → catalog → query → present: Kinesis, Firehose, Glue, Athena, Quick Sight
Elasticity is a console action	Prime Video: DynamoDB partitions doubled mid-game	Managed services scale without a procurement cycle (Domain 1 agility)
Pricing model ≠ service choice	NASA JPL: Spot up to 90% off On-Demand	Choose the in-scope service first, then the purchase option
Out-of-scope names live inside real stories	AWS Elemental / MediaTailor and SWF both appear in AWS's own case studies	A case study never licenses an out-of-scope answer

📚 Did you know? AWS's own Prime Video case study reports 300,000+ clients polling per ad break, precisely the behaviour SQS long polling (wait up to 20 seconds instead of re-polling immediately) and SNS's push model are documented to remove, so the page doubles as AWS's own justification for this exam's push-versus-pull split (case study accessed Oct 2026; limits as of Oct 2026, verify current before use).
📚 Did you know? NASA JPL's page quotes up to a 90 percent discount for Spot versus On-Demand, while its data pipeline runs ~4.4 TB of downlinks a day: one story, two exam domains: a Domain 4 pricing-model answer (Spot for fault-tolerant batch) and a Domain 3 analytics-selection answer (Kinesis, Firehose, Glue, Athena, Redshift), and the 2012 orchestration service it used, SWF, is now out of scope (case study and blogs accessed Oct 2026).
📚 Did you know? Capital One's page contains both halves of this lesson at once: the integration half (Step Functions 80% faster checks, one application −90% on Lambda) and the business-application half; it is one of AWS's published users of Amazon Connect and Amazon SES. AWS also documents the migration's physical tail: 103 tons of copper and steel recycled from the eight data centers it exited (case study accessed Oct 2026).

⚠️Warning

How to read case-study numbers on exam day: every figure above is customer-claimed or AWS-published and unaudited: never a guarantee, and "up to" is a ceiling, never an average. Attribute the source and access date ("Capital One case study, accessed Oct 2026"), not "AWS proves". A case never licenses an out-of-scope answer: you are asked to select the in-scope service, not to reproduce the marketing figure.