Lakshya

Data Platform & Analytics · Chapter 9 of 10

Databricks, Snowflake, and running both

Named in 66% of enterprise postings in a reader-supplied sample, and 30% name both. The question is not which one.

3 min read1 diagramAll 10 chapters

These two dominate enterprise data platform postings by name rather than by category, which is unusual — most job descriptions name capabilities. Understanding what each actually is, and why so many organisations run both, is worth more than a preference.

WHERE EACH STARTED, AND WHERE THEY MET DATABRICKS began: Spark compute for engineers and ML notebooks, files, Python/Scala SNOWFLAKE began: a SQL warehouse for analysts tables, SQL, zero-ops both now offer SQL · notebooks · ML · streaming open table formats · governance So the honest answer to “which is better” is that the question is out of date. What still differs is the centre of gravity: Databricks assumes engineers who write code and want control; Snowflake assumes analysts who write SQL and want none. Both scale; both cost real money. 66% of enterprise postings in the reader-supplied sample name one of them. 30% name BOTH. Running both is not indecision — it is usually an acquisition, or two org units that each chose well for their own workload. The interview question is how you make them coexist, not which you would pick. Source: 38 data-engineering postings supplied by a reader, August 2026. Directional, not representative.
They started at opposite ends and converged. Databricks began as Spark compute for engineers; Snowflake began as a zero-operations SQL warehouse for analysts. Both now do SQL, notebooks, ML, streaming, open table formats and governance — so “which is better” is an out-of-date question.

What genuinely still differs

DatabricksSnowflake
Centre of gravityEngineers who write codeAnalysts who write SQL
Native unitFiles in object storage, Delta or Iceberg tablesTables in a managed service
Compute modelClusters you size and tune, plus serverlessVirtual warehouses that auto-suspend
Where it is strongestML, streaming, large unstructured transformationConcurrency, SQL analytics, sharing, ease of operation
Where it costs youCluster tuning is real work, and idle clusters are real moneyRunaway queries are cheap to write and expensive to run
GovernanceUnity CatalogHorizon / native RBAC and masking
The honest weaknessComplexity — many ways to do the same thingEscaping SQL for heavy engineering work is less natural

Why organisations end up with both

  • Acquisition. The most common reason, and it is nobody's design.
  • Two org units each chose well. The ML team picked Databricks and the finance analytics team picked Snowflake, and both were right for their workload.
  • Migration that stalled at 70%. The value was front-loaded and the last 30% was never funded, so both run forever — the same pattern as any migration.
  • Deliberate separation of engineering compute from analyst-facing serving, which is a legitimate architecture and not a mistake.

What an interviewer is actually asking

Not which you prefer. They are asking whether you can make two expensive platforms coexist without duplicating the data or the definitions — because the failure mode is two copies of the customer table with two different definitions of active, and every number in the business then depending on which one you queried.

The answer that lands: one storage layer in an open table format, one place where conformed entities are defined, and the two engines reading it rather than each owning a copy. Iceberg and Delta with interoperability make this genuinely possible now, which was not true three years ago. Then be honest that governance across both is the hard part, because each has its own catalogue and they do not reconcile themselves.

What to actually learn

  • Pick one and go deep rather than claiming both shallowly, and be explicit about which. Interviewers can tell within two questions.
  • Learn the cost model of whichever you pick, because that is where enterprise conversations go. Cluster sizing and auto-termination for Databricks; warehouse sizing, auto-suspend and query cost attribution for Snowflake.
  • Understand open table formats — Iceberg and Delta — because they are what makes coexistence possible and they appear in 36% of postings.
  • Know the governance surface of your chosen platform, since governance is named in 68% of postings in the same sample.
← Serving the AI workloadStatistics and experimentation, without the degree →