Few debates in data infrastructure generate more heat than "warehouse versus lakehouse." Vendors on each side have spent years arguing the other is obsolete. Meanwhile, most enterprises have quietly arrived at a more practical conclusion: they need both, and the real question is how to divide the work.
Apache Iceberg is a big reason why. Open table formats have made it possible to store data once, in open files on object storage, and let multiple engines read and write it. That's a genuine shift. But it doesn't mean every workload should run directly against the lake, any more than the arrival of the data lake a decade ago made the warehouse disappear.
What Iceberg changed, and what it didn't
Iceberg brought warehouse-like reliability to data lake storage: ACID transactions, schema evolution without rewriting data, time travel and partition management. Just as important, it's open. Spark, Trino, Flink, Snowflake, Databricks and many others can all work with the same tables. That breaks the lock-in that came from storing data in one vendor's proprietary format.
What Iceberg didn't change is the physics of query performance. Reading columnar files from object storage, planning queries across thousands of files, and serving hundreds of concurrent users with sub-second response times are hard problems. A table format defines how data is organized. It doesn't make every engine that reads it fast, predictable or cheap to run.
That gap is why the most effective architectures we see treat the lakehouse and the warehouse as complementary layers, each doing what it does best.
A practical way to divide the work
Think about workloads, not platforms. For each one, ask three questions. How predictable is it? How many people or systems hit it at once? How fast does the answer need to come back?
| Workload | Best home | Why |
|---|---|---|
| Raw and historical data, landing zones, archives | Lakehouse (Iceberg) | Lowest-cost storage, open to every engine, scales without limit |
| Data engineering, large transformations, ML feature prep | Lakehouse | Batch-oriented, throughput matters more than latency |
| Data science exploration across many domains | Lakehouse, with a fast ad hoc SQL engine | Broad and unpredictable, needs access to everything |
| Executive dashboards and operational BI | Warehouse | High concurrency, sub-second response, predictable cost |
| Regulatory, financial and other mission-critical reporting | Warehouse | Governed, consistent and repeatable, with strict SLAs |
| Customer-facing and embedded analytics | Warehouse | Thousands of concurrent users, latency is part of the product |
| AI agents and copilots on curated business data | Warehouse | High query volume, needs governance and workload isolation |
In practice, that usually looks like a medallion architecture. Raw (bronze) and refined (silver) data live in the lakehouse. The curated, business-ready "gold" layer, the data that dashboards, applications, reports and agents depend on every day, is served from a high-performance warehouse.
Why the gold layer still belongs in a warehouse
The gold layer is where the business actually consumes data. It's where concurrency spikes at 9 a.m. on Monday, where the CFO's dashboard has to load instantly, and where an SLA miss becomes a phone call. Those workloads reward an engine built for them:
- Concurrency. A warehouse like Yellowbrick is designed to serve many simultaneous queries without performance falling off a cliff.
- Predictable latency. Data stored in an optimized native format, close to compute, delivers consistent response times that object-storage scans struggle to match.
- Workload management. Fine-grained control over who gets which resources, so critical reporting always comes first.
- Predictable cost. Steady, heavily used workloads are often far cheaper on dedicated capacity than on metered, per-query compute.
- Deployment choice. Yellowbrick runs on its own appliances, on Kubernetes in your private cloud and in every major public cloud. The gold layer can sit wherever your security and sovereignty requirements say it should.
And because Yellowbrick is PostgreSQL-compatible, the BI, ETL and data science tools your teams already use connect without new drivers or rewrites.
Where Floe fits
That leaves one question many customers are asking: what about fast, interactive SQL directly on the lakehouse itself?
That's the problem Floe is built to solve. Floe is a separate initiative led by Yellowbrick founder Neil Carson and a team of senior Yellowbrick engineers. It's built on the same engine heritage, and it brings data warehouse-grade SQL to the lakehouse, for both human analysts and AI agents.
Floe is designed for highly interactive, ad hoc queries over Iceberg data: the unpredictable, cross-domain questions that increasingly come from agents rather than people. The team is also building Floecat, a "catalog of catalogs" that unifies metadata across lakehouse catalogs to improve query planning.
The division of labor is simple:
- Yellowbrick runs today's business-critical warehouse workloads: governed, high-concurrency and predictable, on the infrastructure you choose.
- Floe brings that same SQL engineering to open lakehouse data for ad hoc and agent-driven exploration.
Together they reflect a view we've held for a long time. Enterprises shouldn't have to choose between openness and performance. They should be able to put each workload where it runs best.
Questions to ask before you move a workload
Before migrating any workload from the warehouse to the lake, or the other way, it's worth answering:
- How many concurrent users or systems will query this, at peak?
- What response time does the business actually need?
- Is the query pattern predictable, or exploratory?
- What happens if performance degrades for an hour?
- What will it cost at full production volume, not just in a pilot?
- Where must this data physically reside for security, sovereignty or regulatory reasons?
If the answers point to high concurrency, tight latency, strict governance or predictable cost, that workload belongs in the warehouse. If they point to broad, exploratory, throughput-oriented work, the lakehouse is the natural home.
The bottom line
Iceberg is a real step forward, and open table formats are here to stay. But openness at the storage layer doesn't remove the need for performance at the serving layer. The winning architecture isn't warehouse or lakehouse. It's an open data foundation with the right engine on top of each workload.
Planning your lakehouse strategy? Talk to Yellowbrick about where the warehouse fits, and follow Floe for what's next on the lakehouse.