AWS Certified Data Engineer Associate (DEA-C01) Exam Guide (2026)

If you can say why a Glue job beats Lambda for a 200 GB nightly load, when Firehose is the wrong stream, and why an Iceberg table on S3 is not the same as a Hive-partitioned prefix, you are in the right exam.

DEA-C01 is the associate data engineering cert. It is not AI Practitioner and it is not Machine Learning Engineer. AWS still lists ML training and inference as out of scope for the job. Using a large language model inside a pipeline is now in scope on the version 1.1 outline.

This guide is for people who will sit the current DEA-C01 blueprint, including the 2026 updates around Iceberg, SageMaker Catalog, vectors, and Unified Studio.

Who this exam is for

Take it if you already move data for a living and you need AWS to stop being a pile of similar service names. AWS describes the target candidate as someone with the equivalent of 2 to 3 years in data engineering and 1 to 2 years of hands-on AWS.

You do not need to train models. You do need to ingest, transform, store, operate, and lock down pipelines. You need SQL you can write, not a dashboard you can present. Drawing business conclusions from the data is out of scope. Language-specific syntax is also out of scope. The exam wants the pipeline idea, not a Python trivia item.

Skip it if you want a vocabulary badge for Bedrock prompts. That is AIF-C01. Skip it if you want training jobs and endpoints. That is MLA-C01. Skip it if the work is only CloudWatch alarms and patch windows. That is CloudOps.

The candidate AWS wants already feels volume, variety, and velocity in the body. A 2 GB CSV and a 2 TB CDC stream are not the same design. The sitting punishes people who pick the service they memorized last week instead of the constraint in the stem.

Exam shape

AWS currently weights four domains:

Domain 1 is a third of the scored items. If you only memorize Redshift sort keys and skip Kinesis fan-out and Glue workflows, the math is against you.

The sitting is 65 questions. AWS scores 50 of them. Fifteen are unscored research items and they are not marked. Pass mark is 720 on a 100 to 1,000 scale. Compensatory scoring applies. You do not need a passing score in every domain.

There are two response types. One correct answer of four. Or two or more correct answers of five or more. Unanswered items score as wrong. There is no extra penalty for a guess.

AWS lists the duration as 130 minutes on the certification page. Use the section table on the score report as a hint, not a second pass mark. Compensatory scoring means a weak Domain 4 can still pass if Domain 1 is clean.

What changed in 2026

The exam code is still DEA-C01. AWS did not replace it. Version 1.1 of the official exam guide published on 12 December 2025 is the 2026 sitting. Study materials from 2024 and early 2025 still cover Glue, Kinesis, Redshift, and Lake Formation. They are thin on the skills AWS added.

New skills on the first-party outline:

• Integrate large language models for data processing
• Manage open table formats such as Apache Iceberg
• Describe vector index types, with HNSW and IVF as the examples
• Create and manage business data catalogs, with Amazon SageMaker Catalog as the example
• Describe vectorization concepts, with an Amazon Bedrock knowledge base as the example
• Use domains, domain units, and projects in SageMaker Unified Studio
• Manage data access through SageMaker Catalog projects
• Describe a governance data framework and data sharing patterns

AWS also added vectors to the recommended general IT knowledge. Skill examples now mention HNSW on Aurora PostgreSQL, MemoryDB for key and value access, and lineage through SageMaker Catalog.

Services added to the official in-scope list include Amazon Aurora, Amazon Q, Amazon Bedrock, Amazon Kendra, AWS Data Exchange, and Amazon S3 Tables. Services dropped from that list include AWS Cloud9, AWS CodeCommit, and AWS Schema Conversion Tool.

The job-task out-of-scope list still says do not perform ML training and inferences. The trap is treating every Bedrock mention as "go fine-tune." On this exam, the model is a processing step or a retrieval index, not a training job.

If your notes stop at Hive partitions, a single Glue Data Catalog, and "Kinesis or Firehose," add a week for Iceberg, S3 Tables, SageMaker Catalog, and Unified Studio projects.

How ingestion choices fail

Domain 1 is not a service list. It is a constraint list. The stem will tell you whether later consumers need the raw records, whether the load is a file drop or a change log, and whether the job can finish before a timeout.

Read the stem for consumers, replay, and latency before you pick a stream service. Firehose is a delivery pipe. It is not a replay log.

Kinesis Data Streams or Amazon MSK when more than one consumer must read the same records, or when you must rewind. Firehose when you only need to land, buffer, and convert into S3. AWS DMS with CDC when the source is a database and you need checkpoints. Glue or EMR when the files are large, the schema will change, or Spark is already the shop language. Lambda is the wrong transform when the object is huge or the runtime will blow the function limit.

A useful drill: write one sentence per story that names the SLA and whether you can replay. If you cannot name those two facts, you are guessing the logo.

LLM-in-the-pipeline on version 1.1 is still an ingest or transform step. The model classifies, extracts, or enriches records. It does not become a SageMaker training job because the word Bedrock appeared.

How to study without wasting a month

Week 1. Ingestion you can replay. Streaming vs batch, then the constraint in the stem. Kinesis Data Streams or MSK when multiple consumers need the same records or you must rewind. Firehose when you only need to land and convert. DMS with CDC when the source is a database and you need checkpoints. Glue streaming or Spark when formats differ and the schema will change. Practice fan-in, fan-out, throttling, S3 event triggers, and EventBridge schedules.

Week 2. Where the data lives. S3 with Athena or Iceberg when analysts query files. S3 Tables when the stem wants an Iceberg table that AWS manages. Redshift when the warehouse must join and serve BI. DynamoDB when the access is a key lookup. Aurora or RDS when an application owns OLTP. Glue Data Catalog is the technical catalog. SageMaker Catalog is the business catalog on the 1.1 outline. Add S3 Lifecycle, versioning, Object Lock, and DynamoDB TTL. Practice sort keys, distribution keys, GSIs, and partition evolution.

Week 3. Operate the pipeline. Pair an orchestrator with a failure. Glue workflows or Step Functions for AWS-native graphs. MWAA when the shop already thinks in Airflow. DataBrew or SageMaker Unified Studio when the work is prepare and profile, not a long Spark job. Athena and Redshift SQL for views and checks. CloudWatch for job metrics, CloudTrail for API history, Logs Insights or Athena on the logs when the stem asks you to find the failure. Write quality rules for empty fields, consistency, and skew. If the option is "open the console and rerun," look for the retry and the alarm first.

Week 4. Who can see which row. IAM roles for the job. Lake Formation when the stem wants column or row limits on S3, Athena, Redshift, or EMR. KMS for keys, including cross-account. Secrets Manager for rotation. Macie when PII must be found. Config when the question is "who changed this." Unified Studio domains, domain units, and projects when the version 1.1 catalog story shows up. Redshift data sharing when two accounts need the same warehouse data without a copy. Encryption in transit is not optional on a compliance stem.

Do not spend the month building a second lake "for the cert." Spend it writing the constraint, then picking the service. That is the exam habit.

Traps that look like easy elimination

Easy banks teach the wrong habit. Real DEA items often have two AWS data services that both "move data."

• Many consumers, replay, or per-shard processing is Kinesis Data Streams or MSK. Firehose does not keep a replay log for those consumers.
• "Land it in S3 and convert the format" is often Firehose or a Glue job. It is not a reason to stand up EMR.
• Lambda is the wrong transform when the file is huge or the job will run past the function timeout. Glue or EMR owns that load.
• Daily batch fails a two-hour SLA. Streaming Glue or an event trigger is the usual fix.
• Frequent key lookups are DynamoDB. Ad hoc SQL on a lake is Athena. Warehouse joins are Redshift. Do not put all three stories in Redshift because it can run SQL.
• Hive-style prefixes are not Iceberg. If the stem wants schema evolution, hidden partitioning, or S3 Tables, pick the open table format.
• Glue Data Catalog answers technical schema and crawlers. SageMaker Catalog answers business context, lineage, and project access on the 1.1 outline.
• Lake Formation is the row and column answer. A bucket policy cannot do that job.
• An LLM or Bedrock knowledge base on this exam is processing or retrieval. It is not "train a new model."
• Firehose does not deduplicate your stream. You track sequence numbers or you design the consumer to be idempotent.
• Cloud9 and CodeCommit are gone from the in-scope list. Do not spend study time there.
• HNSW and IVF show up as index types, not as a reason to stand up a training cluster.

If you can delete the scenario and still pick the answer from the service name alone, the question is too easy. The live exam will not do that.

How this maps to CloudFluently

Study with the live AWS Data Engineer Associate practice exam and the beginner AWS data engineering projects course. Pair those with the version 1.1 skills above: Iceberg, SageMaker Catalog, Unified Studio, and LLM-in-the-pipeline.

Frequently Asked Questions

Is DEA-C01 a new exam code in 2026? No. The code stayed. Version 1.1 of the official guide, published 12 December 2025, is what changed the skills.

Does this exam require machine learning? No. AWS still lists ML training and inferences as out of scope. Bedrock or an LLM on this outline is a processing or retrieval step.

How many questions count? Fifty scored items. Fifteen extra items do not count and are not marked. Pass is 720.

What should you study first? Ingestion constraints, then store choice, then the 1.1 catalog and Iceberg skills. Domain 1 is 34 percent.

Where is the official outline? AWS Certified Data Engineer - Associate (DEA-C01). Official delta: DEA-C01 revisions. Use AWS for task statements and weights. Use this page for how to study and what changed.