Databricks started as managed Spark, and if that’s still your mental model, it’s a decade stale. Today it’s a full lakehouse: pipelines, a SQL warehouse, notebooks, model training and serving, all running on open table formats in your own object storage. It’s also, per the Hacker News chatter about its Series K at over $100 billion, one of the most valuable private software companies ever. For a data team the live question is a narrower one: do you need this much platform?
My answer up front
If your team writes Python, builds pipelines and trains models, Databricks earns its complexity and probably consolidates two or three other tools. If your team writes SQL and builds dashboards, it’s a lot of surface area and two bills for a job a plain warehouse does more simply. Most of the regret I’ve seen in either direction comes from misreading which team you are.
The two-bill problem
Understand this before any calculator: Databricks charges in DBUs (Databricks Units) for the platform, and your cloud provider separately charges for the machines and storage underneath. Any comparison that quotes only the DBU rate is showing you roughly half the cost.
The published starting rates by workload: jobs compute from $0.15 per DBU, higher tiers at $0.22 and $0.40 for SQL and serverless options, and model serving from $0.07 per DBU beyond free usage. Committed-use contracts bring the rates down meaningfully, and there’s a free trial plus the Free Edition, which replaced the retired Community Edition in 2025 as the no-cost way to learn the platform. The rate spread is the point: the same hour of compute costs very different amounts depending on whether it runs as a scheduled job or an interactive notebook, and idle interactive clusters are the classic first-invoice shock. Job-ify everything you can, set aggressive auto-termination on the rest, and the platform prices fairly.
What you’re actually buying
Three things justify the platform tax. First, one system for the whole lifecycle: the pipeline that lands the data, the SQL that models it, the notebook that explores it and the model that serves predictions all share one catalogue and one permission model. Second, Delta Lake and Unity Catalog keep your data in open formats on storage you own, which makes this the rare big platform where the exit is genuinely credible: your tables outlive your contract. Third, the engineering ceiling: when a workload gets weird (streaming, huge joins, custom ML), you’re on Spark with full control rather than against a SQL-only wall.
A realistic cost picture
Sketch a mid-size setup: nightly pipelines on jobs compute, a SQL warehouse for the BI layer, and a couple of exploratory notebooks. The jobs run at the friendliest rate, the SQL sits in the middle, and the notebooks, at the top rate with humans forgetting to close them, are where budgets leak. The pattern from teams that run this well is boring and effective: everything scheduled goes through jobs, every interactive cluster gets a termination timer measured in minutes, and someone reviews the workspace’s spend view weekly for the first quarter. On top of the DBUs, remember the cloud line: the VMs, the object storage and the networking land on your AWS, Azure or GCP invoice, and a fair comparison against any warehouse adds the two together.
One more honest note on the entry price: the Free Edition is for learning, not for production, and the trial still runs on your cloud account’s meter for infrastructure. Budget the evaluation like a small project, not a free lunch.
Where teams stumble
Three patterns repeat. Analysts hired for SQL being handed notebooks and clusters they never asked for: solve it by giving them the SQL warehouse and hiding the rest. Unity Catalog adopted late, after permissions have already sprawled: turn it on from day one, migrating governance retroactively is miserable everywhere. And treating the platform as mandatory for everything: it’s perfectly sane to run Databricks for engineering and ML while a lighter BI tool like Metabase serves the dashboard crowd.
What the users say
The review volume is enormous and warm: 4.6 from 1,363 reviews on G2 and 4.6 from 1,219 on Gartner Peer Insights, more Gartner reviews than any other tool in our directory. The G2 spread deserves a second look: 79% five-star and zero percent at three stars or below, the cleanest distribution among our twelve tools. Across 1,363 reviews that’s remarkable, and whatever discount you apply to review-site enthusiasm, the floor it implies is real. Wikipedia attention tells the same story, about 35,000 article views a month, the highest we track. The consistent gripes are the learning curve outside the SQL editor and bills that need active management, both of which match my read above.
Databricks or Snowflake, the eternal question
They’ve converged from opposite directions, and for pure SQL analytics they’re now closer than either admits. The real differences left: Snowflake is simpler to operate and stronger on governed data sharing; Databricks is stronger for engineering-heavy and ML workloads and keeps data in open formats by default. Cost comparisons between them are workload theatre unless you run your own jobs on both, because a Snowflake credit and a DBU aren’t the same unit of anything. If you genuinely sit in the middle, my tiebreaker is your hiring plan: staff of SQL analysts, Snowflake; staff of data engineers, Databricks.
Verdict
The most complete data platform you can buy, priced and shaped for teams that will use most of it. Go in with workloads jobbed, auto-termination on, and a clear reason you need more than a warehouse. If that reason is real, little else competes; if it isn’t, you’ll pay a platform premium to run dashboards.
