Two things about COF-C03 that catch people out.
The passing score is 750 out of 1000, not 75 out of 100. It's a scaled score, so 750 is roughly 75% accuracy but not exactly.
And not every question counts. Some are unscored beta items mixed into the 100, and you can't tell which. There's no question you can safely coast on.
The format: 100 questions, 115 minutes, $175, valid 2 years, Pearson VUE. That's 69 seconds per question.
Where the points actually are
Six domains, and they are not close to even.
| Domain | Weight |
|---|---|
| 1. Cloud Data Platform & Architecture | 25% |
| 2. Account Access & Security | 20% |
| 3. Data Transformation | 20% |
| 4. Data Loading & Unloading | 10% |
| 5. Data Protection & Data Sharing | 10% |
| 6. Performance & Tuning | 15% |
Architecture and Security are 45% between them. If you're short on time, that's where it goes.
Domain 1: Architecture (25%)
The biggest domain, and most of it comes down to knowing which of the three layers does what. Storage, Compute, Cloud Services. Questions phrased as "which layer handles this operation" show up constantly.
What you need cold:
- Micro-partitions: 50-500 MB, columnar, immutable, managed automatically. The immutability is the part people forget, and it's why an UPDATE rewrites partitions.
- The three caches: Result Cache, Metadata Cache, Warehouse Cache. Know where each lives and what invalidates it. This is a favorite.
- Edition differences: Standard, Enterprise, Business Critical, VPS. Which feature starts at which tier.
- Data Marketplace vs Data Exchange, which are easy to blur together.
Domain 2: Account Access & Security (20%)
Mostly RBAC. Know the hierarchy and what each default role can actually do:
ACCOUNTADMIN
├── SECURITYADMIN manages object grants account-wide
│ └── USERADMIN creates users and roles
└── SYSADMIN owns databases and warehouses
PUBLIC not in the tree above.
Automatically granted to every user.
PUBLIC sitting outside that hierarchy is itself testable. It's a pseudo-role every user gets by default, which is why anything granted to PUBLIC is effectively granted to everyone.
The exam also likes asking which role should perform a given task, and the intended answer is usually "the least privileged one that can", not ACCOUNTADMIN.
Also in scope: network policies (IP allowlist and blocklist, and what they apply to), MFA with Duo, and SSO with SCIM provisioning.
Domain 3: Data Transformation (20%)
- SQL UDF vs JavaScript UDF vs External Function, and when each is the right pick
- Stored procedure execution context across JavaScript, Snowflake Scripting, and Python
- Standard View vs Secure View vs Materialized View. Secure Views block the optimizer from certain pushdowns, which is a tradeoff the exam expects you to know.
- Streams plus Tasks as the CDC pattern
Domain 6: Performance & Tuning (15%)
Small domain, but this is where the "choose the best option" questions get genuinely hard.
Scale up for one slow query. Scale out for many concurrent queries. Getting this backwards is the single most common mistake, and it comes up more than once.
Also here: reading a Query Profile (spilling, pruning rate, join order), when clustering keys are worth it (tables over 1 TB, high-cardinality columns), and Resource Monitor notify and suspend actions.
Domains 4 and 5 (10% each)
Ten questions each. Know the surface, don't go deep.
For loading: COPY INTO options (ON_ERROR, PURGE, FORCE, MATCH_BY_COLUMN_NAME), Internal vs External stages, Snowpipe auto-ingest via SQS or REST.
For protection: Time Travel is 0 to 90 days depending on edition, Fail-Safe is a fixed 7 days and only Snowflake can access it. That distinction is worth a guaranteed point. Plus Zero-Copy Clone and the Provider / Consumer / Reader Account model for sharing.
Try one
In which scenario is scaling out (a multi-cluster warehouse) most effective?
A. You want to speed up a single very large query
B. You want to eliminate query queuing caused by many concurrent users
C. You want to reduce storage costs
D. You want to extend the Time Travel retention period
Answer
B. Multi-cluster warehouses handle rising concurrency. When queries start queuing, you add clusters to process more of them in parallel. For a single slow query you scale up instead, meaning a larger warehouse size.
A is the trap, and it catches people who read "multi-cluster" as "more power".
We keep 9,900 practice questions across all 11 SnowPro certifications, each with an explanation on every option. Free to work through: nicheelab.com/en/snowflake-exam.
Four tactics for the day
69 seconds per question means you flag and move. Don't burn four minutes on one item.
Multiple response questions tell you how many to pick ("Select 2"). Pick exactly that many.
On "most appropriate" questions, throw out options that are technically correct but not best practice. Being right isn't the bar.
Watch for negation. "Which is NOT correct" appears often enough that misreading one is a realistic way to lose a point.
After you pass
Core is the prerequisite for all five Advanced tracks: Data Engineer, Administrator, Architect, Data Scientist, Security Engineer. Renewal is a separate Recertification exam ($175, 60 questions, 90 minutes). Let it lapse and you sit the full thing again, so start about 6 months out.
We cover all 11 SnowPro exams if you're continuing past Core.
The full breakdown, with a 6-week study plan, is free to read.
Top comments (0)