www.awsgood.com
The people who get real value from this exam are the ones already on the hook for a broken pipeline at 2am: analytics engineers at 20–200 person companies, consultants who stand up data lakes for SMB clients, and ops folks who inherited a Glue job nobody documents. AWS recommends about a year of hands-on experience with the analytics stack before you sit it, and that guidance is honest for once.
The uncomfortable part in 2026 is DEA-C01. The Data Engineer Associate exam covers ingestion, transformation, cataloging and cost control with far more overlap than the two names suggest, and it costs half as much. If your day job is closer to "move data reliably" than "choose between Redshift Spectrum, Athena and EMR for this query pattern", the associate credential buys you more signal per dollar.
Before you pay for anything, spend five minutes on the official certification page confirming the exam is still listed and bookable in your country and language. AWS folded the old Big Data – Specialty into DAS-C01 back in 2020, and the specialty line-up has been trimmed more than once since. A 2026 study plan that starts with a dead exam is a bad week.
The public exam guide splits DAS-C01 into five domains. Processing and storage together account for 46% of the scored content, which is why people who only memorised service names keep scoring in the 600s.
| Domain | Weight | What it really tests | Perceived difficulty |
|---|---|---|---|
| Collection | 18% | Kinesis Data Streams vs Firehose, MSK, DMS, Snowball Edge, Kinesis Agent | Medium |
| Storage and Data Management | 22% | S3 layout and partitioning, file formats, lifecycle, Lake Formation, Glue Data Catalog | Highest |
| Processing | 24% | Glue/Spark tuning, EMR cluster sizing, Managed Service for Apache Flink, Lambda | High |
| Analysis and Visualization | 18% | Athena, Redshift (provisioned vs Serverless), QuickSight, OpenSearch | Medium |
| Security | 18% | IAM, KMS key policies, Lake Formation grants, VPC endpoints, Macie | Medium-high |
Storage is where most people lose points, not because the services are exotic but because the questions are about consequences: 500,000 objects under 1 MB in a single S3 prefix, then a question about why Athena queries are slow and expensive. If you can explain small-file compaction without looking it up, you're already ahead of the average candidate.
Booking happens through AWS Certification / Pearson VUE. You pick a testing centre or an online proctored slot, and online slots at short notice are the usual reason people reschedule — book two to three weeks out if you need a specific evening.
| Item | DAS-C01 (Specialty) | DEA-C01 (Associate) | SAA-C03 (Associate) |
|---|---|---|---|
| Fee, US region | USD 300 | USD 150 | USD 150 |
| Questions / minutes | 65 / 130 | 65 / 130 | 65 / 130 |
| Passing scaled score | 750 | 720 | 720 |
| Validity | 3 years | 3 years | 3 years |
Fees and taxes vary by region, and a retake costs the same as the first attempt — which is the real argument for not booking before your practice scores stabilise. AWS marks a block of the 65 questions as unscored, so two or three questions that feel like they're from a different exam aren't worth spiralling over mid-slot.
Seventy hours sounds abstract until you put it in weeks. My own split looked like this, and it's roughly what I'd hand to anyone with a full-time job:
| Week | Focus | Hours | Deliverable |
|---|---|---|---|
| 1 | Collection: Streams vs Firehose vs MSK | 12 | One working ingestion path into S3 |
| 2 | Storage: partitioning, formats, catalog | 14 | Parquet table with partition projection |
| 3 | Processing: Glue jobs, EMR sizing | 14 | A Spark job you actually debugged |
| 4 | Analysis: Athena tuning, Redshift modes | 12 | Same query, two engines, cost compared |
| 5 | Security: KMS, Lake Formation, endpoints | 10 | A 403 you fixed on purpose |
| 6 | Practice exams and weak spots | 8 | Two timed runs above 80% |
Theory alone won't carry you through the Processing domain. Spin up a tiny lab and break it:
# us-east-1, one shard, delete it when you're done
aws kinesis create-stream --stream-name das-lab --shard-count 1 --region us-east-1
aws kinesis describe-stream-summary --stream-name das-lab --region us-east-1
--query 'StreamDescriptionSummary.[StreamStatus,OpenShardCount]' --output text
ACTIVE 1
Then point a Firehose delivery stream at S3 with a 128 MB / 300 s buffer, and query the output with a projected table instead of a crawler:
```sql
CREATE EXTERNAL TABLE events (
event_time timestamp,
event_type string,
payload string
)
PARTITIONED BY (dt string)
STORED AS PARQUET
LOCATION 's3://das-lab-111122223333/events/'
TBLPROPERTIES (
'projection.enabled' = 'true',
'projection.dt.type' = 'date',
'projection.dt.range' = '2025-01-01,NOW',
'projection.dt.format' = 'yyyy-MM-dd',
'storage.location.template' = 's3://das-lab-111122223333/events/dt=${dt}/'
);
```bash
aws kinesis delete-stream --stream-name das-lab --region us-east-1 --enforce-consumer-deletion
Budget intuition matters here: Athena list price is USD 5 per TB scanned in us-east-1, so a single unpartitioned `SELECT *` over 18 GB of raw JSON shows up as a line item you'll remember. Converting to Parquet with the right partitions typically drops scanned bytes by an order of magnitude, and that exact trade-off is a recurring exam theme.
## Hard-won lesson: the 403 and the crawler that ate 40 minutes
Last November we moved an SMB client's clickstream off a self-managed Kafka box into Firehose → S3 → Athena. The Glue Crawler finished in 92 seconds, created two tables, and every query against them failed:
COLUMN_NOT_FOUND: line 1:15: Column 'event_type' cannot be resolved
My first assumption was a permissions problem, which was wrong and cost me twenty minutes. I pulled the table definition:
```bash
aws glue get-table --database-name clicks --name firehose_out
--query 'Table.StorageDescriptor.[Columns,SerdeInfo]' --region us-east-1
The output showed a single column named `col0` and a `LazySimpleSerDe` with `TextInputFormat` — a CSV-shaped table over JSON files. Firehose writes objects with no `.json` extension, so the default classifier had no idea what it was looking at and guessed wrong. Three Athena queries in a row failed the same way before I stopped blaming IAM.
The fix was to stop crawling. We defined the table by hand with partition projection (the DDL above), pointed the SerDe at `org.openx.data.jsonserde.JsonSerDe`, and the next query returned rows in 1.4 seconds while scanning 240 MB instead of the 18 GB the crawler's table had been pulling. Verification was CloudTrail plus `EXPLAIN` — no mysterious grants, just a classifier guessing. If I'd checked `SerdeInfo` first, the whole detour would have been five minutes.
That pattern — crawler guesses, query fails, engineer blames permissions — is worth rehearsing. It's the sort of thing the Storage and Data Management domain quietly tests.
## Exam technique, traps, and when NOT to sit this exam
With 65 questions in 130 minutes you get two minutes each, and scenario questions in Processing routinely eat four. My rule: first pass at 90 seconds per question, flag anything longer, and leave 25 minutes for the flagged pile. Multi-select questions state how many answers to pick — picking two when three are correct scores zero, so count before you click.
The traps repeat. Kinesis Data Streams limits (1 MB/s or 1,000 records/s per shard) versus Firehose's buffer-and-batch model. Redshift provisioned versus Serverless when the workload is spiky and small. Lake Formation column-level grants that require the table to be registered, not just an S3 bucket policy. Anything asking for "the most cost-effective" answer usually wants partitioning plus compression, not a bigger cluster.
**When not to sit it:** if this would be your first AWS certification, start with SAA-C03 — the vocabulary is a prerequisite. If your work is EC2, networking and IAM for