What is a lakehouse? DuckLake and Parquet for small firms
What a lakehouse is, how DuckLake and Parquet differ from a traditional data warehouse, and why small-business data doesn't need a big cluster to be fast.
By The DataxLand team· · 6 min read
When small business owners look into data warehouses, they run into a forest of terms: data warehouse, data lake, lakehouse, Iceberg, Parquet, cluster. Most of the material is written for companies with hundreds of terabytes and a team of data engineers. The usual takeaway is "this isn't for me".
In fact, the idea behind a lakehouse is fairly simple, and it suits small businesses better than you might think. This post explains in plain language what a lakehouse is, what Parquet and DuckLake are, and why the data of most small and mid-sized businesses runs fine on a single machine, without an expensive server cluster.
Three ways to store data for analysis
| Approach | Where the data lives | Strengths | Weaknesses |
|---|---|---|---|
| Data warehouse | Inside an analytical database system | Fast SQL queries, transactions, permissions | Data is "locked" in the vendor's format; cost per server or per amount scanned |
| Data lake | Files on cheap storage (object storage) | Cheap, open, holds anything | No transactions, hard to know which table is the right version, easily becomes a file "swamp" |
| Lakehouse | Open files on object storage + a catalog that manages tables | Cheap and open like a lake, with transactions and versions like a warehouse | Needs a catalog management layer |
In short: lakehouse = data in open file formats, plus a ledger that records which files belong to which table at which version. That ledger is what turns a folder of files into a warehouse with transactions.
Parquet: the lakehouse file format
Apache Parquet describes itself as an "open source, column-oriented data file format designed for efficient data storage and retrieval", with compression and encoding schemes for large data (parquet.apache.org, accessed 16 Nov 2026).
What does "column-oriented" mean for a shop? Picture an orders file with 30 columns. The question "revenue by month" only needs two columns: date and amount. With a row-based file (like CSV or Excel), the machine has to read all 30 columns of every row. With Parquet, it reads only the two columns it needs. Values in the same column are also alike (same type, many repeats such as channel names), so they compress very well.
| CSV / Excel | Parquet | |
|---|---|---|
| Human-readable | Yes | No (needs a tool) |
| Stores data types (dates, numbers) | Not reliably | Yes, per column |
| Querying a few columns over many rows | Must read everything | Reads only the needed columns |
| Size | Large | Much smaller thanks to columnar compression |
| Tools that can read it | Everywhere | Most modern data tools |
DuckLake: the catalog is an ordinary SQL database
Well-known lakehouse formats keep their ledger (metadata) as many JSON or Avro files next to the data. DuckLake takes a different approach: the metadata lives in a SQL database. The official site describes DuckLake as a lakehouse format built on SQL, with a catalog served by an ACID-compliant SQL database such as PostgreSQL, SQLite, MySQL or DuckDB, while the data is plain Parquet files on disk or object storage (ducklake.select, accessed 16 Nov 2026).
According to the same site, DuckLake has:
- Multi-table ACID transactions: a write that updates five tables either succeeds for all five, or none of them change.
- Snapshots and time travel: read a table exactly as it was at an earlier version.
- An MIT license for the specification and the DuckDB extension.
The first two features sound technical, but they solve exactly two pains of file-based reporting: "someone is halfway through editing the file" and "nobody knows which data September's numbers were calculated from".
Why you don't need a big cluster
Big data systems split work across many machines because the data does not fit on one. Small-business data usually does.
Illustrative example (assumed): a shop with 500 orders a day, an average of 3 product lines per order, keeping 2 years of data.
500 orders × 365 days × 2 years = 365,000 orders
365,000 orders × 3 lines = 1,095,000 product linesAbout a million rows — right at the 1,048,576-row limit of an Excel worksheet (Microsoft), but tiny for an analytical engine. DuckDB, the engine that usually comes with DuckLake, describes itself as an analytical SQL database system that runs "in-process" (embedded in the application, no separate server), reads and writes Parquet on local disk or object storage, under an MIT license (duckdb.org, accessed 16 Nov 2026). An ordinary server handles a few million rows like this fast enough for a dashboard.
The cost consequence: you pay for storage (compressed Parquet files) and for compute time when it actually runs, instead of keeping a cluster running day and night.
When a lightweight lakehouse is not enough
To be fair, this approach has limits:
- Hundreds of continuous concurrent writers (for example, a real-time transaction system) — that is the job of a transactional database, not an analytical warehouse.
- Extremely large event data, such as click logs from an app with millions of users a day.
- Second-level latency requirements for every query on freshly written data.
Most management reporting in a small business — revenue, profit, inventory, receivables and payables, daily ads — falls into none of these cases.
How DataxLand does it
Every DataxLand workspace has its own lakehouse-style warehouse: Parquet files on object storage, a table catalog dedicated to the workspace, and storage keys used only for that workspace. Each build writes to the run's build warehouse; dashboards, Si and MCP only read the published version, so you never see numbers from a build that is still in progress. See Architecture.
You don't set up a catalog, choose an engine or manage files. You connect sources (Excel files, marketplaces, databases), Si proposes how to read your data for you to approve, and DataxLand builds the warehouse. The cloud version keeps the warehouse and its backups in Vietnam (data residency). When you need to take your data elsewhere, the export includes the standardized tables and metric tables as Parquet (Export and delete data).
Frequently asked questions
Can a lakehouse replace Excel?
It doesn't replace it; it sits behind it. Excel is still where you look, experiment and calculate quickly. The lakehouse holds the correct copy of the data, with history, so every report reads from one source.
Does Parquet lock my data into a vendor?
Parquet is an open format that many popular analytics tools can read. That is one reason a lakehouse is less locked in than a warehouse that uses an internal format.
Does a small business need to hire a data engineer to use a lakehouse?
If you build it yourself, yes: someone has to choose the tools, write the loading pipelines, the checks, and run it. If you use a platform that packages the lakehouse inside, you work with sources, metrics and dashboards, not with files directly.
What is time travel good for in reporting?
It answers "which data was last month's number calculated from". When someone asks why the report sent last week differs from today's, you can compare the two versions instead of guessing.
Read next: Outgrowing Excel: when you need a data warehouse · What is a semantic layer · 5 sales data errors to check