Snowflake - The Complete Masterclass
Learn Snowflake through guided lessons, assignments, and quizzes, with coverage spanning warehouse fundamentals and more advanced platform topics.
A data warehouse stores cleaned, structured data in tables optimized for fast SQL analytics and reporting. A data lake stores large volumes of raw data in any format (tables, logs, JSON, images) cheaply in object storage, and imposes structure only when the data is read. A lakehouse puts warehouse-style features (tables, transactions, fast SQL) on top of lake storage using open table formats, aiming to give one copy of data for both BI and data science.
These three terms appear in almost every data engineering job posting and cloud vendor pitch, often with little explanation. The easiest way to understand them is through the problem each one was designed to solve.
The problem it solved: business users needed consistent, fast answers from data spread across many operational systems.
A warehouse holds structured, modeled data, typically organized as fact and dimension tables. Data is cleaned and conformed before or during loading, so the schema is defined up front. This is called schema-on-write.
Strengths
Weaknesses
Examples: Snowflake, Google BigQuery, Amazon Redshift, Azure Synapse, and traditional on-premises warehouses.
The problem it solved: companies began collecting huge volumes of varied data (clickstreams, sensor data, logs, documents) that didn't fit neatly into warehouse tables, and was too expensive to store there.
A data lake stores raw files in cheap object storage such as Amazon S3, Azure Data Lake Storage or Google Cloud Storage, in their original formats. Structure is applied only when someone reads and processes the data. This is called schema-on-read.
Strengths
Weaknesses
The problem it solved: many companies ended up running both a lake (for data science) and a warehouse (for BI), copying data between them. That meant double cost, delays and two versions of the truth.
The breakthrough was open table formats, mainly Delta Lake, Apache Iceberg and Apache Hudi. They add a metadata layer on top of files in object storage, which provides:
With these, one copy of data in cheap storage can serve both BI dashboards and machine learning. Databricks popularized the term "lakehouse" around 2020. Since then most major platforms have moved in this direction. Microsoft Fabric is built on Delta tables, and Snowflake and BigQuery both support Iceberg tables. The line between "warehouse" and "lakehouse" products is blurring.
| Data warehouse | Data lake | Lakehouse | |
|---|---|---|---|
| Data types | Structured | Anything | Anything, with table structure where needed |
| Schema | On write | On read | Enforced on tables, flexible for raw zones |
| Storage cost | Higher (historically) | Lowest | Low (object storage) |
| Main users | Analysts, BI | Data scientists, engineers | Both |
| SQL performance | Excellent | Poor without extra tools | Good to excellent |
| Reliability of updates | High | Low (historically) | High (via table formats) |
| Main risk | Rigidity, cost | Data swamp | Complexity; still a maturing ecosystem |
Many teams organize a lake or lakehouse in layers, often called the medallion architecture:
The transformations between layers are often written in SQL with dbt, or in Spark for heavy processing. Whether the gold layer lives in a "warehouse" or in lakehouse tables matters less than having clear layers, tests and ownership.
For a small or mid-sized company in Vietnam or elsewhere starting its first data platform, a single cloud warehouse with ELT and dbt is usually the fastest path to value. Lakehouse architecture pays off as data volume, variety and ML ambitions grow.
Analysts mostly need to know where trusted, modeled data lives and how to query it. Data engineers are expected to understand all three patterns, object storage, table formats and the trade-offs between them. Our data engineering roadmap shows where these skills fit, and Data Analyst vs Data Engineer compares the two paths.
Data Warehouse Fundamentals (DataCamp) covers warehouse concepts and modeling from the ground up. For a hands-on start with one of the most widely used cloud platforms, Snowflake: The Complete Masterclass (Udemy) walks through loading, querying and managing data. Browse all Data Warehousing & Modeling courses.
Is a data lake cheaper than a data warehouse? Storage in a lake is usually much cheaper. Total cost also depends on compute, tooling and the engineering time needed to keep a lake usable, so a warehouse can still be cheaper overall for small, structured workloads.
Is Snowflake a data warehouse or a lakehouse? It started as a cloud data warehouse and has added lakehouse capabilities such as Iceberg table support. Most major platforms now offer a mix, so focus on the capabilities you need rather than the label.
Do I need a lakehouse for machine learning? Not necessarily. Plenty of ML on tabular business data runs directly from a warehouse. A lakehouse helps most when ML needs large volumes of raw or unstructured data.
What is a data swamp? A data lake that has lost its usefulness: data dumped without documentation, ownership or quality checks, so nobody can find or trust anything in it.
Warehouses give structured, fast, trustworthy analytics. Lakes give cheap storage for any data at scale. Lakehouses use open table formats to combine the two on one copy of data. For most teams, the right answer depends on data variety, scale and ML needs, and a well-governed simple setup beats an impressive architecture that nobody trusts.