Appendix
The Numbers
Every exact figure the exam can ask, on one page. Don't memorize this early: the concepts in Parts 1 to 7 make most of these derivable. Cram this page in the final 48 hours, when raw recall is what's left to optimize.
The Numbers · Final Cheat Sheet
Every Figure Worth Memorizing
Storage and data
| Fact | Number |
| Micro-partition size (uncompressed) | 50–500 MB |
| Max size of a VARIANT, ARRAY or OBJECT value | 128 MB (16 MB before the 2025_03 behavior change) |
| Recommended load file size (compressed) | 100–250 MB |
| COPY load metadata retained per table | 64 days |
| Snowpipe load history retained | 14 days |
| Query result cache lifetime | 24 h, re-extended on reuse, max 31 days |
Protection and retention
| Fact | Number |
| Time Travel default | 1 day (all editions) |
| Time Travel range, Standard edition | 0–1 day |
| Time Travel range, Enterprise+ | 0–90 days (permanent tables) |
| Time Travel max for transient and temporary tables | 1 day |
| Fail-safe (permanent tables, non-configurable) | 7 days |
| Fail-safe for transient and temporary tables | 0 days |
Compute and cost
| Fact | Number |
| Warehouse sizes | XS = 1 credit/h, doubling each step (S 2, M 4, L 8, XL 16, ... 6XL 512) |
| Minimum billing per resume | 60 seconds, then per-second |
| Max clusters in a multi-cluster warehouse | 10 in Snowsight. Through SQL: 300 for XS to Medium, 160 for Large, 80 for XL, 40 for 2XL, 20 for 3XL, 10 for 4XL and up |
| Multi-cluster availability | Enterprise edition and above |
| Cloud services billed only above | 10% of daily warehouse compute credits |
| Snowpark-optimized warehouse memory | 16x a standard node |
| Serverless features billed per second | Snowpipe, tasks (serverless), auto-clustering, MV maintenance, SOS, QAS, replication |
| ACCOUNT_USAGE retention / latency | 365 days / ~45 min to 3 h |
Access, security, editions
| Fact | Number / value |
| System roles hierarchy | ORGADMIN · ACCOUNTADMIN ← SECURITYADMIN ← USERADMIN, ACCOUNTADMIN ← SYSADMIN ← PUBLIC (bottom) |
| Column-level + row access policies, and multi-cluster | Enterprise+ |
| Tri-Secret Secure, private connectivity, HIPAA/PCI | Business Critical+ |
| Failover / failback for business continuity | Business Critical+ |
| Dedicated isolated deployment | Virtual Private Snowflake (VPS) |
| Periodic rekeying of encryption keys | Enterprise+, yearly when enabled |
Loading and sharing
| Fact | Number / value |
| Snowpipe typical latency | ~1 minute after file arrival |
| Snowpipe Streaming latency | Seconds (row-level, no files) |
| Default unload file size (COPY INTO location) | ~16 MB per file, MAX_FILE_SIZE up to 5 GB on cloud storage |
| Shares: who pays compute | Consumer (except reader accounts, where the provider pays) |
| Objects shareable in a share | Tables, external tables, secure views, secure materialized views, secure UDFs |
| Cross-region / cross-cloud sharing | Via listings with auto-fulfillment (replication under the hood) |
Commonly confused pairs
The exam's favorite distractors are the near-twin of the right answer. These are the pairs worth one deliberate pass:
| This | vs that | The line between them |
PUT | COPY INTO <table> | PUT moves a local file into an internal stage (client-side command); COPY moves staged files into a table (needs a warehouse) |
GET | COPY INTO @stage | GET downloads staged files to the client; COPY INTO a stage unloads table data into new files on the stage |
LIST @stage | SELECT FROM DIRECTORY(@stage) | LIST is a metadata command you can't filter with SQL; DIRECTORY is a queryable table, but only after being enabled and refreshed on the stage |
UNDROP | CREATE ... CLONE ... BEFORE(STATEMENT => ...) | UNDROP resurrects an object that was dropped; the clone recovers the pre-DML state of an object that still exists |
| Owner's rights | Caller's rights | Default procedures run as the owning role (controlled elevation); EXECUTE AS CALLER runs as the invoker and is the only mode that reads caller session state |
| Time Travel | Fail-safe | Self-service, configurable 0 to 90 days vs Support-only, fixed 7 days, permanent tables only |
| Scale up | Scale out | Bigger warehouse fixes spilling (one heavy query); more clusters fix queuing (many concurrent queries) |
| Clustering key | Search Optimization | Physical reorder serving range predicates vs an access path serving equality lookups; both fix "partitions scanned ≈ total", for opposite filter shapes |
The Last 48 Hours
Re-take the seven practice sets and log misses by part. Anything under 70%: re-read that part's Exam Focus page only, not the whole part. The night before, read this appendix twice and stop. On exam day: 100 questions, 115 minutes, so bank time on the easy recall questions to spend on multi-selects. Eliminate the one clearly wrong option first in every "choose two." Never leave a blank; there's no penalty. And when a scenario feels ambiguous, re-read the stem for the constraint you skipped: edition, table type, or who owns the object. That constraint is usually the whole question.
Memory aid
"1-7-90, 64, 128, 10." One day of default Time Travel, seven of Fail-safe, ninety max. Sixty-four days of COPY metadata. 128 MB per VARIANT value, while 16 MB keeps its two other jobs: a micro-partition's typical compressed size and the default unload file. Ten clusters in Snowsight and a 10% cloud services threshold. Six numbers unlock a dozen questions.