AWS Learning Material Certification Cheatsheets
AIP-C01 Domain 1

Task 1.3: Implement data validation and processing pipelines for FM consumption

Foundation Model Integration, Data Management, and Compliance · 4,803 words · source: Certified-Generative-AI-Developer-Professional-AIP-C01/domain-1/task-1-3-cheat-sheet.md

Domain 1: Foundation Model Integration, Data Management, and Compliance

Task Statement 1.3: Implement data validation and processing pipelines for FM consumption

Domain 1 is 31% of scored exam content. Task 1.3 focuses on preparing data before it reaches a foundation model (FM): validate quality, process text/image/audio/tabular inputs, format requests for the target model interface, and enrich or normalize inputs so model responses are more accurate and consistent.

Use this sheet for scenario questions that ask how to build an ingestion or preprocessing pipeline, which AWS service fits a data type, how to validate source data, or how to format inputs for Amazon Bedrock or SageMaker AI inference.

Official Study Path

Use these in order:

  1. AWS Skill Builder: Exam Prep Plan: AWS Certified Generative AI Developer - Professional (AIP-C01)
  2. AWS Skill Builder: Domain 1 Review: AWS Certified Generative AI Developer - Professional
  3. AWS Skill Builder: Domain 1 Practice and AWS SimuLearn scenario practice
  4. AWS Exam Guide: Domain 1, Task Statement 1.3
  5. AWS Docs: AWS Glue Data Quality, Amazon SageMaker Data Wrangler, Amazon SageMaker Processing, Amazon Bedrock Converse API, Amazon Transcribe, Amazon Textract, and Amazon Comprehend

Exam Objective Map

You should be able to:

Official skill What to know for the exam
Skill 1.3.1: Create comprehensive data validation workflows to ensure data meets quality standards for FM consumption Validate completeness, freshness, schema, allowed values, duplicates, PII, quality scores, and failed records before the FM receives data
Skill 1.3.2: Create data processing workflows to handle complex data types including text, image, audio, and tabular data Choose modality-specific services and orchestration patterns for OCR, speech-to-text, entity extraction, feature processing, document parsing, and multimodal prompts
Skill 1.3.3: Format input data for FM inference according to model-specific requirements Structure Amazon Bedrock JSON requests, Converse messages, content blocks, system/user prompts, inference parameters, tool/guardrail fields, and SageMaker endpoint payloads
Skill 1.3.4: Enhance input data quality to improve FM response quality and consistency Normalize, clean, enrich, label, extract entities, redact sensitive data, standardize formats, and repair malformed inputs before inference

One-Minute Mental Model

Task 1.3 is about the pipeline between raw data and FM inference.

Raw input
  -> validation
  -> modality-specific processing
  -> normalization/enrichment
  -> model-specific request formatting
  -> FM inference
  -> output validation and logging

Exam shortcut:

Structured/tabular quality checks -> AWS Glue Data Quality.
Interactive data prep and feature engineering -> SageMaker Data Wrangler in SageMaker Canvas.
Managed custom preprocessing jobs -> SageMaker Processing.
Audio to text -> Amazon Transcribe.
Document OCR/forms/tables/signatures -> Amazon Textract.
Text entities/key phrases/sentiment/PII -> Amazon Comprehend.
Multimodal model request -> Amazon Bedrock Converse content blocks, model-specific media rules.
Simple transformation/glue logic -> Lambda.
Multi-step pipeline with retries and branching -> Step Functions.
Quality metrics and alerts -> CloudWatch.

What Task 1.3 Is Really Testing

The exam is not just asking whether you can call a model. It is testing whether you know how to make the model input trustworthy, usable, and shaped for the target interface.

Exam area What that means in practice
Data quality Validate that input data is complete, current, correct, deduplicated, and fit for the intended FM task
Data type handling Use specialized processing for text, documents, images, audio, video, and tabular data
Interface formatting Match Amazon Bedrock, SageMaker AI, or model-specific payload schemas exactly
Input enrichment Extract useful structure before inference, such as entities, tables, labels, OCR text, timestamps, or metadata
Safety and compliance Detect or redact PII, enforce access control, encrypt data, and avoid leaking sensitive content into logs
Reliability Quarantine bad data, retry transient failures, branch by data type, and record pipeline outcomes
Observability Emit metrics for validation failures, processing latency, token usage, malformed requests, and model errors

Data Quality Dimensions To Recognize

Dimension Meaning for FM consumption Example validation
Completeness Required fields, documents, or metadata are present Every support case has case_id, created_at, body, and customer_tier
Validity Values match expected type, format, range, or enum Dates parse as ISO 8601, language code is valid, JSON schema passes
Consistency Same entity or field uses the same representation Product IDs and customer IDs use one canonical format
Accuracy Data reflects the source of truth CRM record matches latest authoritative system
Freshness Data is recent enough for the use case Policy document was updated in the last approved refresh window
Uniqueness Duplicate records or documents are removed or handled Same PDF is not embedded twice in a processing batch
Integrity Relationships between records are valid Foreign key or reference ID resolves to a known source record
Relevance Data is useful for the prompt or retrieval task Irrelevant boilerplate, signatures, and navigation text are removed
Safety Sensitive or disallowed content is identified PII, PHI, secrets, offensive text, or unsupported file types are detected

Validation Workflow Pattern

Use this default pattern for exam scenarios about reliable data pipelines.

Data source
  -> landing zone in S3
  -> schema and metadata validation
  -> content quality validation
  -> security and PII checks
  -> quarantine invalid records
  -> transform valid records
  -> write curated input
  -> invoke FM or index into downstream system
  -> publish metrics and audit events
Step AWS services to recognize What to validate
Ingest S3, AppFlow, Kinesis, MSK, SQS, EventBridge, DataSync File type, source identity, checksum, event metadata
Catalog or profile AWS Glue Data Catalog, Glue crawlers, Athena, SageMaker Data Wrangler Schema, column types, distributions, missing values
Data quality rules AWS Glue Data Quality, Lambda, SageMaker Processing DQDL rules, custom rules, data quality score, failed rows
Modality processing Textract, Transcribe, Comprehend, Bedrock multimodal models OCR output, transcript quality, entities, language, metadata
Transformation AWS Glue ETL, SageMaker Processing, Lambda, Step Functions Normalized text, converted formats, chunk boundaries, JSON validity
Security checks Macie, Comprehend PII, IAM, KMS, Secrets Manager Sensitive data, access scope, encryption, secret exposure
Orchestration Step Functions, EventBridge, SQS Retries, branching, dead-letter or quarantine path
Monitoring CloudWatch metrics/logs, CloudTrail, X-Ray Failure rate, latency, data quality score, request errors

AWS Glue Data Quality

AWS Glue Data Quality is a strong answer when the scenario involves validating datasets in the Glue Data Catalog or enforcing data quality checks inside ETL jobs.

Feature Exam relevance
Data Quality Definition Language (DQDL) Define rules for quality checks in a consistent, versionable language
Data quality score Summarize rule pass/fail health for a dataset
Data Catalog entry point Evaluate quality of cataloged datasets already in a data lake
ETL job entry point Proactively evaluate and filter bad data before loading curated data
Failed record identification in ETL jobs Quarantine or fix the records that caused rule failures
EventBridge and CloudWatch integration Trigger workflows and monitor data quality outcomes
ML-based anomaly detection in ETL jobs Detect hard-to-find quality issues over time

Exam cues:

If the question says... Choose...
"Validate tabular source data before sending records to an FM" AWS Glue Data Quality rules in an ETL job
"Monitor quality of cataloged lake data over time" AWS Glue Data Quality on the Data Catalog
"Reject or quarantine records with missing required fields" Glue ETL with Data Quality transform or custom Lambda validation
"Business users need readable data quality rules" DQDL rulesets and data quality score

Exam trap: AWS Glue Data Quality validates data quality. It does not replace modality-specific extraction tools such as Textract for documents or Transcribe for audio.

SageMaker Data Wrangler And SageMaker Processing

Use SageMaker data tools when the scenario is closer to ML-grade data preparation, transformation, feature engineering, or scalable custom preprocessing.

Service or feature Best use
SageMaker Data Wrangler in SageMaker Canvas Interactive, low-code data preparation, analysis, transformation, featurization, and data quality insights
Data Wrangler data flows Repeatable prep steps that can be exported or integrated with ML workflows
SageMaker Processing Fully managed preprocessing, postprocessing, feature engineering, and evaluation jobs using built-in or custom containers
SageMaker Processing with S3 Run scripts over S3 input and write processed output back to S3
CloudWatch for Processing jobs Monitor job metrics and logs

Decision cues:

Requirement Better answer
"Analyst needs to explore, clean, and visualize a dataset with little code" SageMaker Data Wrangler
"Need a repeatable Python/Spark preprocessing job on managed infrastructure" SageMaker Processing
"Need custom libraries or custom container preprocessing" SageMaker Processing
"Need to prepare training or inference data as part of a SageMaker workflow" SageMaker Processing or Data Wrangler exported into SageMaker Pipelines

Processing By Data Type

Text

Common processing tasks:

Need AWS services and pattern
Detect language, entities, key phrases, sentiment, syntax, or PII Amazon Comprehend
Normalize casing, whitespace, HTML, markdown, encoding, or delimiters Lambda, AWS Glue ETL, SageMaker Processing
Split long text for model context windows Lambda, SageMaker Processing, Bedrock preprocessing prompt, later retrieval pipeline
Redact sensitive terms Amazon Comprehend PII plus custom Lambda redaction
Convert unstructured text into structured JSON Amazon Bedrock with schema instructions plus output validation

Good text input is specific, clean, scoped, and formatted. Bad text input is duplicated, noisy, irrelevant, full of boilerplate, or too large for the model context.

Documents And Images

Use a document/image pipeline when the input is a scanned PDF, form, invoice, claim, ID image, chart, screenshot, or photo.

Need AWS services and pattern
Extract raw text from documents Amazon Textract text extraction or Bedrock multimodal model where appropriate
Extract forms and key-value pairs Amazon Textract AnalyzeDocument with forms
Extract tables Amazon Textract table extraction
Ask targeted questions about document fields Amazon Textract queries or Bedrock multimodal prompt with schema
Detect signatures or layout Amazon Textract document analysis
Extract structured fields from images/documents with an FM Amazon Bedrock multimodal model with JSON schema in the user prompt

Exam trap: For deterministic document extraction from business forms, Textract is often the better preprocessing service. For reasoning over visual content or flexible multimodal understanding, use a Bedrock multimodal model if the selected model supports the media type.

Audio

Use Amazon Transcribe when the FM needs text from speech.

Need AWS services and pattern
Convert recorded media in S3 to text Amazon Transcribe batch transcription
Convert live audio to text Amazon Transcribe streaming transcription
Identify speakers Amazon Transcribe speaker partitioning / diarization features
Improve domain-specific terms Custom vocabulary or language customization
Protect sensitive audio transcript content Transcribe content filtering/PII features where applicable, then Comprehend or Lambda redaction
Summarize a meeting or call Transcribe -> transcript cleanup -> Bedrock summarization

Tabular Data

Use tabular processing when the input is CSV, JSON records, data lake tables, operational database exports, or features.

Need AWS services and pattern
Validate schema and quality AWS Glue Data Quality
Transform large datasets AWS Glue ETL, EMR, SageMaker Processing
Query and sample data Athena, Glue Data Catalog
Convert records to prompt-ready summaries Glue/SageMaker Processing/Lambda -> structured prompt context
Detect anomalies or missing values Glue Data Quality, Data Wrangler insights, custom validation

Do not dump large raw tables into prompts. Summarize, filter, aggregate, or retrieve only the rows and columns needed for the task.

Multimodal Inputs

Multimodal processing means the FM receives or reasons over more than one input type, such as text plus image, document plus OCR text, video plus prompt, or audio plus metadata.

Need Exam-level design
Understand an image or document directly Use a Bedrock model that supports the required modality
Improve reliability of document extraction Combine OCR text from Textract with the original image/document in a multimodal prompt
Reference several files in one prompt Label each file with a text content block, then include the media block
Extract structured fields Put the media before the user prompt and include a JSON schema
Reduce randomness for extraction Use low temperature, often 0, when the task is deterministic

AWS docs for Amazon Nova multimodal prompting emphasize that the user text prompt should appear after the media content, and detailed task/output-format instructions should be in the user prompt.

Formatting Input For Amazon Bedrock

Amazon Bedrock request formatting is a central Task 1.3 theme.

Interface Use when
Converse API You want a consistent message-based interface across supported Bedrock chat models
ConverseStream API You need streaming responses for interactive experiences
InvokeModel You need direct model-specific native request format
Prompt Management ARN You want to invoke a managed prompt version and supply prompt variables
SageMaker endpoint You deploy or invoke a model through SageMaker AI and must match that endpoint's expected payload

Bedrock Converse Request Shape

Recognize the core fields:

Field Purpose
modelId The model, inference profile, provisioned throughput, custom model, or prompt resource
messages Ordered conversation turns with role and content
system Optional system instructions, when allowed by the request mode
inferenceConfig Common parameters such as maxTokens, temperature, topP, and stop sequences
additionalModelRequestFields Model-specific parameters outside the common Converse set
guardrailConfig Guardrail configuration for content safety
toolConfig Tool/function-calling definitions
promptVariables Values for variables when invoking Prompt Management prompts
requestMetadata Metadata useful for logging, tracing, or routing

Minimal example:

{
  "modelId": "model-or-inference-profile-id",
  "messages": [
    {
      "role": "user",
      "content": [
        {
          "text": "Summarize this support case in JSON with fields: issue, urgency, next_action."
        }
      ]
    }
  ],
  "inferenceConfig": {
    "maxTokens": 500,
    "temperature": 0
  }
}

Content Blocks

In Converse, message content is an array of content blocks. AWS docs define content blocks for types such as text, document, image, audio, video, tool use, tool result, citations content, search result, cache point, reasoning content, and guardrail content. Support is model-specific, so always match the target model's documented capabilities.

Content block type Exam cue
text Plain prompt, instructions, extracted text, labels, or structured context
document Document input for supported models
image Image input for supported vision-capable models
audio Audio input for supported models or workflows
video Video input for supported multimodal models
toolUse / toolResult Tool-calling conversations
guardContent Apply guardrail assessment to selected content blocks
citationsContent Preserve traceability from response to source content

Multimodal pattern:

{
  "role": "user",
  "content": [
    {
      "text": "Invoice image:"
    },
    {
      "image": {
        "format": "png",
        "source": {
          "bytes": "..."
        }
      }
    },
    {
      "text": "Extract invoice_number, vendor, total, and due_date as valid JSON."
    }
  ]
}

Exam trap: A syntactically valid JSON request can still be wrong if the model does not support that content type, file format, prompt field, parameter, or context size.

Formatting Input For SageMaker AI Endpoints

SageMaker endpoints are flexible, so the "right" payload depends on the deployed container or model.

Requirement What to do
Custom or fine-tuned model behind an endpoint Match the endpoint container's request schema exactly
Real-time low-latency inference Use endpoint payload format expected by the model server
Batch or large preprocessing before endpoint call Use SageMaker Processing, batch transform, or a pipeline
Multiple model versions Use SageMaker Model Registry, endpoint variants, or deployment pipeline controls
Structured downstream integration Validate JSON schema before invoking and validate output after response

Exam shortcut:

Bedrock managed FM -> Bedrock API schema.
SageMaker hosted model -> endpoint/container schema.
Dialog app -> conversation message format.
Downstream system needs JSON -> schema prompt + deterministic settings + output validation.

Input Quality Enhancement Techniques

Technique Why it improves FM output
Normalize text Reduces inconsistent formatting and irrelevant noise
Remove duplicates Prevents repeated context from biasing or bloating prompts
Extract entities Gives the FM explicit structure such as people, dates, products, or locations
Detect language Routes to the right prompt, model, translation step, or reviewer
Redact PII Reduces privacy exposure and compliance risk
Add metadata Helps grounding, filtering, traceability, and source attribution
Convert media to text Makes audio, scanned documents, and images usable in text-only or hybrid workflows
Summarize long inputs Keeps prompts within context limits while preserving relevant signal
Validate JSON Prevents malformed requests and downstream integration failures
Use schemas Makes extraction and model responses more consistent
Use low temperature for extraction Reduces variability when the desired output is deterministic
Store rejected records Enables repair and audit instead of silent data loss

Common Pipeline Architectures

Real-Time Text Enrichment

Client
  -> API Gateway
  -> Lambda validation and normalization
  -> Comprehend PII/entities if needed
  -> Bedrock Converse request formatter
  -> Bedrock model
  -> output schema validation
  -> response

Use for chat, case summarization, classification, routing, and short user-generated text.

Batch Document Processing

Document upload to S3
  -> EventBridge or S3 event
  -> Step Functions workflow
  -> Textract OCR/forms/tables
  -> validation and cleanup Lambda
  -> Bedrock extraction/summarization
  -> store curated JSON in S3/DynamoDB
  -> CloudWatch metrics and alerts

Use for claims, invoices, contracts, forms, policies, and knowledge ingestion.

Audio Summarization

Audio file in S3
  -> Transcribe batch job
  -> transcript cleanup and speaker formatting
  -> Comprehend or Lambda PII redaction
  -> Bedrock summarization or action-item extraction
  -> store transcript, summary, and metadata

Use for meetings, calls, contact center recordings, voicemail, and interviews.

Tabular Data Preparation

Raw table or files
  -> Glue Data Catalog
  -> Glue Data Quality rules
  -> Glue ETL or SageMaker Processing transform
  -> curated dataset in S3
  -> prompt context generation or model endpoint input
  -> CloudWatch data quality metrics

Use for datasets, customer records, product catalogs, analytics exports, and data-lake-backed FM workflows.

Multimodal Extraction

Image/document/video/audio
  -> file validation and metadata capture
  -> optional Textract or Transcribe preprocessing
  -> Bedrock multimodal content blocks
  -> schema-constrained extraction prompt
  -> JSON validation and repair flow

Use when the model must reason over media content, not just pre-extracted text.

Service Chooser For Task 1.3

Requirement clue Best-fit service or feature
"Validate data quality with rules" AWS Glue Data Quality
"Author readable data quality rules" DQDL rulesets
"Identify failed rows in an ETL pipeline" AWS Glue Data Quality in AWS Glue ETL jobs
"Interactive data cleaning and visualization" SageMaker Data Wrangler in SageMaker Canvas
"Managed preprocessing job with custom code/container" SageMaker Processing
"Extract text, forms, tables, signatures, layout from documents" Amazon Textract
"Convert speech to text" Amazon Transcribe
"Detect entities, key phrases, language, sentiment, syntax, or PII" Amazon Comprehend
"Lightweight request normalization" AWS Lambda
"Branch by data type and retry failed steps" AWS Step Functions
"Stream records or handle high-volume events" Kinesis, MSK, SQS, EventBridge
"Store raw and curated files" Amazon S3
"Monitor quality and pipeline metrics" Amazon CloudWatch
"Audit AWS API activity" AWS CloudTrail
"Format chat prompts across Bedrock models" Bedrock Converse API
"Send media plus prompt to supported model" Bedrock content blocks with model-specific media support
"Invoke custom model" SageMaker AI endpoint with endpoint-specific payload

Validation And Error Handling Checklist

Failure mode Recommended response
Missing required fields Reject or route to quarantine; emit validation metric
Invalid schema Return clear error or run deterministic repair if safe
Unsupported file type Reject before invoking Textract, Transcribe, or Bedrock
Poor OCR or transcript confidence Route for human review or alternate processing
PII present when not allowed Redact, mask, deny, or use approved secure path
Input too large for context window Chunk, summarize, retrieve only relevant sections, or use batch pipeline
Model-specific field unsupported Use correct API, model parameters, or model-specific adapter
Downstream JSON parse failure Validate, retry with repair prompt, or fail closed
Processing timeout Use async pipeline with Step Functions, SQS, or batch job
Repeated transient service errors Retry with backoff and dead-letter/quarantine after threshold

Metrics To Emit

Professional-level answers include observability.

Metric Why it matters
Records processed Baseline pipeline volume
Validation pass/fail count Measures data quality gate health
Data quality score Summarizes quality rule outcomes
Quarantined records Tracks rejected input and repair workload
Processing latency Shows bottlenecks by step
Service error count Tracks Textract, Transcribe, Bedrock, Glue, or SageMaker failures
Malformed request count Detects request formatting bugs
Output schema failure count Catches model or prompt regressions
Token usage Controls cost and prompt bloat
PII detection count Measures privacy exposure and redaction workload

Security And Compliance Cues

Concern Exam-ready control
Sensitive documents in S3 Encrypt with KMS, restrict bucket access, use IAM least privilege
PII in text Detect with Comprehend PII or purpose-built controls, redact before model call when required
Secrets in prompts Scan/filter input and never log secrets in plain text
Prompt and response logging Apply retention, redaction, encryption, and access controls
Cross-service access Use IAM roles with least privilege
Regulated data Preserve audit trails with CloudTrail and keep processing in approved Regions
Multi-tenant inputs Enforce tenant metadata, access filtering, and separate data paths where needed
Tool or API calls after FM output Validate model-generated parameters before calling downstream systems

Scenario Decision Table

If the question says... Choose...
"The model responses are inconsistent because source data has missing fields and duplicate records" Add a validation and cleansing step with Glue Data Quality or custom validation before FM invocation
"The company needs to process thousands of PDFs and extract tables before summarization" S3-triggered Step Functions workflow with Textract, cleanup Lambda, and Bedrock
"Calls must be summarized after contact center recordings are uploaded" Amazon Transcribe batch to text, then Bedrock summarization
"Live meeting captions are required before downstream GenAI analysis" Amazon Transcribe streaming, then near-real-time processing
"A no-code user wants to explore and clean data before ML workflows" SageMaker Data Wrangler in SageMaker Canvas
"A Python preprocessing script must run over large files in S3 without managing servers" SageMaker Processing
"The app must send a chat-style request to different Bedrock models" Bedrock Converse API message format
"The app sends images and text to an FM" Use a multimodal Bedrock model and correctly ordered content blocks
"The downstream app requires exact JSON" Include schema instructions, use low temperature, validate output, and retry/repair invalid JSON
"The input text contains customer names, account numbers, and addresses" Detect and redact PII before model invocation if not required for the task
"Bad rows should not stop the whole batch" Quarantine failed records and continue processing valid records
"Quality failures need to trigger alerts" Publish CloudWatch metrics and use EventBridge or alarms

Exam Traps

Trap Correct thinking
"Send raw data directly to the FM" Validate, clean, redact, and format first
"One preprocessing method works for all data" Choose modality-specific processing for text, image, audio, document, and tabular data
"Glue Data Quality extracts text from PDFs" Glue validates data quality; Textract extracts document text, forms, and tables
"Transcribe summarizes calls" Transcribe converts speech to text; use Bedrock or another analysis step for summarization
"Comprehend is a general FM" Comprehend is NLP analysis for entities, key phrases, language, sentiment, syntax, PII, and related insights
"Valid JSON means valid model input" The payload must also match the target model, API, content type, and parameter support
"Multimodal prompts ignore content order" For models such as Amazon Nova, place media before the final user prompt and put detailed task instructions in the user prompt
"Temperature should be high for extraction" Deterministic extraction usually uses low temperature, often 0
"Logs are harmless" Prompts, files, transcripts, and responses may contain sensitive data
"Validation ends before inference" Also validate model output if downstream systems require structured data

Study Checklist

Before moving to Task 1.4, make sure you can:

  • Explain why high-quality FM inputs improve response accuracy and consistency.
  • Map validation requirements to AWS Glue Data Quality, Lambda, SageMaker Processing, or Data Wrangler.
  • Choose Textract, Transcribe, Comprehend, or Bedrock multimodal models for the right data type.
  • Describe a Step Functions pipeline for batch document, audio, or tabular preprocessing.
  • Format a basic Amazon Bedrock Converse request with messages, content blocks, and inference parameters.
  • Explain why model-specific input requirements matter for Bedrock and SageMaker endpoints.
  • Identify when to redact PII or quarantine bad data before inference.
  • Name the operational metrics that make a preprocessing pipeline production-ready.

Quick Practice Questions

  1. A company wants to summarize scanned claim forms, but the forms contain tables and signature fields. What should happen before the summarization prompt?

    • Use Amazon Textract to extract document text, forms, tables, layout, or signatures, then normalize the extracted content before invoking Bedrock.

  2. A dataset in a data lake has missing IDs, stale records, and invalid enum values. Which service is the best exam answer for managed data quality checks?

    • AWS Glue Data Quality with DQDL rules and data quality score tracking.

  3. A meeting summary workflow starts with audio files uploaded to S3. Which service converts the files into text?

    • Amazon Transcribe batch transcription.

  4. A Bedrock request includes messages, content blocks, inferenceConfig, and guardrailConfig. Which API is this describing?

    • Amazon Bedrock Converse API.

  5. A custom preprocessing script must run on managed infrastructure and read/write S3 data. Which service fits?

    • Amazon SageMaker Processing.

Sources