Vendors describe these three as a progression, as if a warehouse were last decade’s answer, the lake a mistake and the lakehouse the destination. That story is tidy and mostly marketing. They are three different bets about where to pay the cost of organising data: before you store it, after you store it, or with a layer of rules in between. I will argue that for most Indian mid-size companies the first bet is still the right one.
The warehouse: organise first, store second
A data warehouse holds structured, cleaned, modelled tables. Data is shaped to a defined schema as it is loaded, which is sometimes called schema-on-write. Analysts query it with SQL and get predictable answers and predictable speed, because the engine knows exactly what the columns are and has optimised the storage for it. Governance is comparatively simple: you grant access to tables and columns, and the definitions of revenue, customer and order sit in one agreed place.
The price is upfront effort. Someone has to decide what the tables look like before data arrives, and awkward data such as images, free text, audio or deeply nested JSON does not fit comfortably. Historically, storage and compute were also tied together, so growth meant paying for both. Modern cloud warehouses have separated these, which has softened the cost argument considerably.
The lake: store first, organise later
A data lake is cheap object storage, files in folders, holding raw data in whatever format it arrived: CSV, JSON, Parquet, logs, images. Structure is applied when someone reads it, called schema-on-read. The attraction is flexibility and low storage cost. Data scientists can get at raw material, including data nobody yet knows how to use, and machine learning teams like having it.
The well-known failure is the “data swamp”. Without enforced structure, ownership and cataloguing, nobody trusts the files. Two folders hold two slightly different versions of the same extract. A failed job leaves half-written files. Plain files also lack transactions: if one process is writing while another reads, the reader may see a partial result, and updating or deleting individual rows, which privacy requests sometimes demand, is awkward.
The lakehouse: files plus a rulebook
A lakehouse keeps the lake’s cheap open files but adds a table layer on top. Open table formats such as Delta Lake, Apache Iceberg and Apache Hudi keep metadata that records which files make up a table at a given moment. That gives transactions (a write either fully happens or does not), schema enforcement, row-level updates and deletes, and the ability to query an earlier version of a table. SQL engines and ML tools can then read the same copy of the data.
The promise is one copy of data serving both reporting and data science, without exporting from a lake into a separate warehouse. Reality is more uneven. The table formats are maturing quickly, but you take on operational choices: which format, which catalogue, which compute engine, how to compact small files, how to manage permissions across all of them. Those are engineering decisions that a managed warehouse makes for you.
Side by side
| Question | Warehouse | Lake | Lakehouse |
|---|---|---|---|
| Typical data | Structured, modelled | Raw, any format | Both, with table layer |
| When structure is applied | On load | On read | Table layer enforces it |
| Reliable updates and deletes | Yes | Difficult | Yes, via table format |
| Team skill needed | SQL, modelling | Data engineering | SQL plus data engineering |
| Main risk | Rigidity, cost at scale | Swamp, no trust | Complexity of the stack |
Where I land, and where I would change my mind
If your data is mostly orders, payments, customers, tickets and web events, and your main consumers are analysts and dashboards, start with a managed cloud warehouse. A team of five to fifteen people does not have spare capacity to run table maintenance and catalogue configuration, and the business questions are not waiting. Most of the “lake” benefits can be had by dropping raw files into object storage as a landing area and loading from there, which is an ordinary and sensible pattern, not a lake strategy.
I would shift towards a lake or lakehouse when one or more of these becomes true: large volumes of semi-structured or unstructured data that must be kept (clickstream at serious scale, sensor data, documents); an active machine learning group that needs raw access and reproducible training data; a wish to avoid lock-in to one vendor’s storage format; or storage costs for very large histories that become uncomfortable in a warehouse. The decision follows the workload, not the fashion.
One last point that matters more than the label. Whichever you choose, the hard problems are the same: agreed metric definitions, ownership of each dataset, quality checks, access control and a catalogue people actually use. A lakehouse with no owners is a swamp with better branding, and a warehouse with clear owners is worth more than either.
Leave a Reply