Warehouse or lakehouse is one of the few architecture questions where the vendor marketing has genuinely muddied the water. Warehouses have added semi-structured support; lakehouses have added transactions and governance. The two categories have converged far enough that for a large share of mid-sized teams either choice would work fine, and the deliberation costs more than the difference.
Here is how we actually run this decision with clients.
First: are you sure you need either one?
If your data lives in one production Postgres instance, totals in the low hundreds of gigabytes, and your analytics needs are dashboards plus the occasional ad-hoc query, a read replica and a scheduled transformation job will serve you well past the point most teams assume. We have watched organisations spend two quarters standing up a platform to serve a workload a replica handled comfortably.
The signal that you have outgrown that setup is not data volume. It is the moment analytical queries start affecting production performance, or the moment you need to join data from three or more systems that have no common key.
The cheapest data platform is the one you did not need to build. Deferring the decision by six months costs far less than picking wrong and migrating.
Question one: what shape is your data?
If the overwhelming majority of what you analyse arrives as rows from operational databases, SaaS APIs and event streams with stable schemas, a warehouse is a straightforward fit. The tooling, the SQL ergonomics, and the analyst hiring pool all point the same way.
If a meaningful share of your value sits in images, audio, documents, or raw sensor output that you want to keep in original form and process with Python rather than SQL, the lakehouse story becomes real rather than theoretical. Open table formats let the same files serve both a SQL engine and a training job without a second copy.
Question two: who actually queries it?
This one is decisive more often than the technical criteria and gets discussed least. A team of SQL-fluent analysts with no platform engineers will be productive on a managed warehouse on day one and will struggle with cluster configuration, file compaction and partition tuning.
A team with strong Python and data engineering capacity gets real value from a lakehouse's flexibility and cost profile. Choosing a platform your team cannot operate is the most expensive mistake available here, and no benchmark will warn you about it.
Question three: what does the cost curve look like?
Warehouses generally cost more per terabyte stored and less per hour of engineering. Lakehouses invert that — object storage is cheap, and the savings get spent on the people keeping the platform healthy.
- Under roughly 10 TB with steady query patterns, warehouse pricing is rarely the constraint.
- Above that, or with heavy reprocessing, separating storage from compute starts to matter financially.
- Model the engineering time honestly. A platform engineer costs more annually than a lot of storage.
Question four: how strict is your governance regime?
Under GDPR, HIPAA or a comparable framework, row- and column-level access control, lineage and auditable deletion are requirements rather than nice-to-haves. Warehouses have had mature answers here for years. Lakehouse governance has improved substantially but still needs deliberate assembly. If you have a compliance deadline, weigh that gap seriously.
What we usually recommend
For most mid-sized teams: start with a managed warehouse and ELT. Land raw data first, transform inside the warehouse, and keep the transformations in version control from day one. It is the fastest route to answering business questions, and answered questions are the only real output of a data platform.
Add lakehouse components later, when a specific workload justifies it — unstructured data, machine learning at scale, or a storage bill that has become a genuine line item. Keeping raw data in open formats from the beginning is what makes that later step an addition instead of a migration.
- Land raw, transform downstream — never overwrite the source of truth.
- Version-control every transformation from the first commit.
- Store raw data in open formats so today's choice is not permanent.
- Pick the platform your team can operate on a Tuesday afternoon.
