Snowflake

Snowflake tutorial — the data warehouse, not the paper craft.

The practical Snowflake guide for data engineers: what the platform actually gives you, which features matter in production, and how to keep the bill under control.
Start the Snowflake course See dbt on Snowflake

The architecture that made Snowflake different

  • Storage. Columnar micro-partitions (~16 MB compressed) on the cloud provider's object storage. You never touch files.
  • Compute. Virtual warehouses — independent MPP clusters (XS through 6XL) that spin up in seconds and auto-suspend.
  • Services layer. Metadata, security, optimization, transactions. Shared across all warehouses.
  • Why it matters. ETL, BI and ad-hoc each get their own warehouse. Zero resource contention.

The SQL and the extras

SQL is ANSI + QUALIFY, MATCH_RECOGNIZE, PIVOT/UNPIVOT, semi-structured (VARIANT, FLATTEN) and generous window function support. Beyond SQL:

  • Streams — change tracking on a table. The foundation of CDC pipelines.
  • Tasks — scheduled SQL. Chain with streams for near-real-time transforms.
  • Dynamic Tables — declarative incremental materializations. dbt-like, native.
  • Snowpark — Python/Scala/Java UDFs and DataFrame API running inside Snowflake compute.
  • Iceberg Tables — read/write Iceberg on your own S3 bucket. No data lock-in.

Performance: micro-partitions, clustering, pruning

Snowflake auto-prunes micro-partitions when your WHERE clause filters on a column with good clustering. If a large table is filtered heavily on one column (say, event_date), define a clustering key on it:

ALTER TABLE events CLUSTER BY (event_date, tenant_id);

Check pruning with SYSTEM$CLUSTERING_INFORMATION. If average_depth creeps up, re-cluster. Don't cluster tiny tables — it costs more than it saves.

Cost tuning (the part that gets you promoted)

  • Auto-suspend 60s. Default is 10 min. That's ~9 minutes of idle spend per query burst.
  • Right-size warehouses. Bigger warehouse = shorter runtime, roughly linear cost — but only if the query actually parallelizes.
  • Result cache. Free. Enable it, and prefer identical queries where you can.
  • Query profile. Read every slow query's profile. Spilling to remote storage = wrong warehouse size or bad join order.
  • Materialized views + search optimization are expensive — use surgically, not by default.
  • Monitor. Set resource monitors. Alert on daily credit spend, not just monthly.

Governance and access

RBAC hierarchy: warehouses, databases, schemas and objects. Use role hierarchy (SYSADMIN → ANALYTICS_ADMIN → ANALYST), never grant directly to users. Row-access policies for tenant isolation. Column masking for PII. Object tagging + Access History for audit.

The learning path

DataForge's Snowflake course walks through the whole platform with real datasets — architecture, performance, streams/tasks, Snowpark, cost tuning and interview-grade query optimization. Pair it with our dbt tutorial (dbt on Snowflake is the market default) and SQL for data engineering.

FAQ

What is Snowflake?
Snowflake is a cloud data warehouse that separates storage from compute. You store data once in columnar micro-partitions on cloud object storage, then spin up independent compute clusters (virtual warehouses) to query it. Each team can size its own warehouse; workloads never fight for resources.
How does Snowflake pricing actually work?
Two lines: storage (~$23/TB/month compressed) and compute (credits per second per running warehouse, ~$2–$4/credit depending on edition and cloud). 90% of surprise bills come from warehouses left running or serverless features (Snowpipe, tasks, materialized views) firing more than expected. Auto-suspend at 60s and monitor credit usage weekly.
Snowflake vs BigQuery vs Databricks in 2026?
Snowflake wins on ease of use, governance and BI workloads. BigQuery wins on ad-hoc analytics and Google-native stacks (serverless, no cluster to size). Databricks wins on ML + heavy Spark + lakehouse. All three read Apache Iceberg now, so 'lock-in on data' is less of a factor than it was.
Do I need to learn Snowflake if I know SQL?
Yes — the SQL is standard, but the warehouse model, micro-partitions, clustering, streams, tasks, Snowpark, dynamic tables, and cost tuning are specific to Snowflake and show up in every senior interview at Snowflake shops. It's 1–2 weeks of focused learning.
How do I get free hands-on Snowflake practice?
Snowflake offers a 30-day, $400-credit trial — enough to build a full end-to-end project. Combine with Snowflake's sample datasets (SNOWFLAKE_SAMPLE_DATA) so you're not blocked on loading data. Suspend warehouses aggressively.

Ready to start?

7 days free. Then less than a coffee per month.

Start the Snowflake course