Reference · Chapter 29
Databricks vs Snowflake: cost, workloads, when to move
Reviewed by TechFabric engineers. How we write this guide →
Databricks and Snowflake both run SQL analytics well, and for dashboards and ad hoc queries the people using them will notice little difference. Databricks is the better fit once the work includes long-running transformation, model training, agents or applications that read and write the same data; Snowflake is the better fit for spiky analytical SQL that runs hard in office hours and sleeps at night. If the reason for comparing them is a bill, fix the workload before choosing a platform, because moving an unchanged workload mostly moves the invoice.
The question usually arrives dressed as architecture. Somebody's Snowflake bill has grown faster than the business did, a vendor has explained that a lakehouse would be cheaper, and a meeting that started about money is now about platforms. This chapter separates the two, and argues against migrating in the cases where we'd expect a migration to disappoint.
The short version
Snowflake is a warehouse that has grown outward into pipelines, machine learning and Postgres. Databricks is a processing platform that has grown inward toward warehouse-grade SQL. If your work is dashboards and queries, the warehouse fits. If it has become hours of transformation, model training and applications built on the data, the platform fits. Both now read and write Apache Iceberg, so the choice is less permanent than it was. A bill is driven by the shape of the workload, and the logo on the invoice changes less than people hope.
Databricks vs Snowflake at a glance #
As of October 2026, checked against both vendors' documentation:
The two platforms, side by side
| Databricks | Snowflake | |
|---|---|---|
| Where it started | Spark processing, grown inward toward warehouse SQL | A cloud data warehouse, grown outward into pipelines and ML |
| Unit on the bill | DBUs per second, priced by compute type | Credits per second, with a 60-second minimum each time a warehouse starts |
| Where the data sits | Your own cloud storage, in Delta or Iceberg tables | Snowflake-managed storage, or Iceberg tables on an external volume you own |
| Catalog and governance | Unity Catalog | Snowflake Horizon Catalog |
| Iceberg interchange | Managed Iceberg tables, read and written by outside engines through Unity Catalog's Iceberg REST catalog; Snowflake-managed Iceberg tables read through catalog federation | Snowflake-managed Iceberg tables, plus externally managed ones through a catalog integration, writable when the remote catalog is an Iceberg REST catalog |
| SQL analytics | SQL warehouses, serverless or classic | Virtual warehouses |
| Long-running transformation and Spark | Native, the original workload | Snowpark, or a separate engine |
| ML, models and agents | MLflow, Model Serving, Agent Bricks, AI Functions | Snowflake ML and Cortex AI, including Cortex Agents |
| Questions in plain English | Genie Agents, with Genie One for business users | Cortex Analyst over semantic views, and Snowflake CoWork (formerly Snowflake Intelligence) |
| Applications on the data | Databricks Apps | Streamlit in Snowflake |
| Postgres for applications | Lakebase: serverless, scales to zero, copy-on-write branches | Snowflake Postgres (generally available since 24 February 2026): a dedicated VM per instance, forks are full copies |
| Best fit | Mixed work: pipelines, ML, applications and SQL on one copy of the data | Spiky analytical SQL that queries hard in office hours |
Sources for the Snowflake column: virtual warehouses, Iceberg tables, Cortex Analyst and Snowflake Postgres. For Databricks: Iceberg in Unity Catalog (last updated 15 October 2026) and Lakebase.
For SQL analytics the two nearly are rivals feature for feature. What the table can't show is what your bill is paying for, and that's where we'd start any conversation about moving.
Is Databricks cheaper than Snowflake? #
Sometimes, and rarely for the reason people expect. The instinct is that the heavy aggregation is the expensive part, and we'd test that instinct first, because it's often wrong.
An engineer whose firm builds these pipelines put the pattern plainly on Hacker News in September 2025: "the primary driver of Snowflake costs is not the compute for aggregation, but the compute required for lots of reads/scans." Their fix for one customer was to keep hourly aggregates in Snowflake and push the results into a store built for serving, so the dashboards stopped hitting the warehouse. That "slashed their Snowflake bill by ~25%, which was worth millions to them."
That fix didn't involve leaving Snowflake. It came from noticing that user-facing dashboard traffic and analytical aggregation are different workloads, and that paying warehouse prices to serve a dashboard was the defect.
Snowflake is good, and it is not a relational database #
The fairest assessment of Snowflake we've read came from an engineer who spent months trying to reduce a bill and came away more impressed than when he started. He describes gaining "respect as in 'this is really great' as well as respect as in 'I need to be on guard here or I'm going to get hurt.'"
His diagnosis of why teams overspend is the useful part. "I think my biggest misconception at the outset was thinking of Snowflake like it's a relational database. It's not." There are no b-tree indexes. There are clustering keys, which colocate data so queries can prune micropartitions, and a well clustered table filtered on its clustering key is fast. Join on something else and you'll pay for it, unless you enable search optimization, which costs more again.
A team applying Postgres instincts to Snowflake will produce a bill that looks like a platform problem and is really a modelling problem. Migrating that team to Databricks carries the same instincts onto a platform with its own ways to be expensive.
Where the two genuinely differ #
Set cost aside and the real difference, in our view, is the shape of the work.
Snowflake's compute model is built for spiky analytical SQL and it's very good at it. You pay per second with a 60-second minimum when a warehouse starts, auto-suspend and auto-resume are on by default, and a suspended warehouse costs nothing. For a business that queries hard in office hours and barely at night, that model is close to ideal.
It fits less comfortably when the work stops being queries. Long-running transformation, Spark jobs and training runs hold compute for hours, and the thing that made short queries cheap does nothing for you there. Databricks was built for that workload first, and it's where we'd make the case for it without needing a cost story.
The second difference, and the one we'd weigh most for a team building products on its data, is what happens past the dashboard. When the interesting work becomes software (an operator console, an agent with permissions, a service that writes back), a lakehouse with an operational database beside it does something a warehouse was never meant to do.
Lakebase versus Snowflake Postgres #
Both vendors now sell a managed Postgres beside the analytics platform, and on a feature list the two look alike: the Postgres wire protocol, backups, high availability, point-in-time recovery. Underneath they're different machines, and the difference decides what development, testing and a bad Tuesday in production cost you.
Snowflake Postgres runs, in Snowflake's words, "a Postgres database server on a dedicated virtual machine managed by Snowflake." You pick an instance size up front and pay for it while it runs. Lakebase separates storage from compute, so compute autoscales with load and suspends when nobody is connected.
Two managed Postgres services, beside two analytics platforms
| Databricks Lakebase | Snowflake Postgres | |
|---|---|---|
| Compute | Serverless, autoscaling, storage separate from compute | A dedicated VM per instance, sized when you create it |
| When nobody is using it | Suspends idle compute on its own and wakes on the next connection | Bills until somebody suspends it; storage keeps billing while suspended, and resuming typically takes 3 to 5 minutes |
| Changing size | Scales within the range you set, with no maintenance event | A new configuration "requires a failover maintenance" |
| A copy of production | A copy-on-write branch, ready in seconds and holding no data of its own until something changes | A fork, which "performs a full copy of all of the origin data", in a time that grows with the database |
| What the copy costs | Only the pages it changes, and compute while it runs | A full instance, priced on the fork's own plan, storage and high availability |
| Recovering from a bad write | Branch at a timestamp or an exact LSN and repair with ordinary SQL | Fork from the last base backup with the write-ahead log replayed, within a 10-day window |
| Governance | Registered in Unity Catalog, with synced tables flowing to and from Delta | Snowflake's own roles and policies, with a catalog integration that exposes Postgres tables to Snowflake as read-only Iceberg tables |
Why it costs less. Most of a database's life isn't peak load. It's nights, weekends, staging environments nobody has opened since Thursday, and the copy an engineer made to test a migration. A dedicated VM bills compute for all of that, and every extra environment is another VM with its own storage. On Lakebase the idle hours cost nothing, and a branch for every engineer or every pull request costs only what it writes. The more environments a team runs, the wider that gap gets.
Why development is faster. A branch is a full, writable copy of production data in seconds, so every pull request can run its migrations against a copy of production before anyone merges. With a full-copy fork you ration those copies, because each one takes time to build and bills as a whole instance.
Why troubleshooting is easier. When something breaks in production, you branch from the moment before it broke and read the data as it was, without touching the live database. When an UPDATE loses its WHERE clause, recovery is a branch at a timestamp and a repair query, where a fork means a restore and a log replay.
Preetham Reddy, TechFabric's founder, compares the two in Lakebase vs Snowflake Postgres: branching is the difference, and the branch, scale-to-zero and recovery timings we measured live are in Lakebase, with the architecture.
Two details decide whether the branching argument survives a real team. Put an expiry on every throwaway branch and make "branch per pull request, deleted on merge" part of the pipeline, because a permanent branch is billed differently from an expiring one. And remember what a branch doesn't carry: logical replication slots and subscriptions aren't inherited, so a branch that has to publish or subscribe needs them set up again.
The migration that relocates the bill #
What to try before you move #
Three steps, in order, and none of them needs a migration. This is the sequence we'd run before agreeing that a migration is the answer.
Find out where the money goes. Snowflake's account usage views will tell you, and we'd expect the answer to be more concentrated than anybody guesses: a few dashboards, or one badly clustered join. Start with credits by warehouse:
-- Snowflake: credits by warehouse over the last 30 days
SELECT warehouse_name, SUM(credits_used) AS credits
FROM snowflake.account_usage.warehouse_metering_history
WHERE start_time >= DATEADD(day, -30, CURRENT_TIMESTAMP())
GROUP BY warehouse_name
ORDER BY credits DESC;Then attribute the top warehouse's credits to whoever runs the queries. QUERY_ATTRIBUTION_HISTORY carries per-query credits; it leaves out warehouse idle time and queries of about 100 ms or less, and lags by up to eight hours, so read it as a ranking rather than a reconciliation:
-- Snowflake: who spends the top warehouse's credits, last 30 days
SELECT user_name,
query_tag,
COUNT(*) AS queries,
SUM(credits_attributed_compute) AS credits
FROM snowflake.account_usage.query_attribution_history
WHERE warehouse_name = 'BI_WH' -- your top warehouse from the query above
AND start_time >= DATEADD(day, -30, CURRENT_TIMESTAMP())
GROUP BY user_name, query_tag
ORDER BY credits DESC
LIMIT 20;If the top of that list is a BI tool's service account, you've found serving traffic paying warehouse prices.
Move the read-heavy serving off the warehouse. This is the change we'd expect to pay back first. Dashboards that scan all day belong somewhere built for serving, and the aggregation can stay exactly where it is.
Then look at what remains. If it's spiky analytical SQL, you have a healthy warehouse workload and no reason to move. If it's hours of transformation, training runs and applications waiting to be built, you've found a real platform argument, and a better one than the bill was.
Open table formats make that split cheaper than it used to be. Databricks exposes managed Iceberg tables to outside engines, Snowflake included, through Unity Catalog's Iceberg REST catalog, and Snowflake can write to externally managed Iceberg tables when the remote catalog is an Iceberg REST catalog. In the other direction, Unity Catalog reads Snowflake-managed Iceberg tables as foreign tables through catalog federation, read-only. We'd design any split around a shared Iceberg table with one writer, never a nightly copy.
How to migrate from Snowflake to Databricks #
When the exercise above does point at Databricks, we'd move consumers before data. Lakehouse Federation for Snowflake (page last updated 13 October 2026) gives two ways to read Snowflake from Unity Catalog. Query federation pushes each query down to Snowflake over JDBC, which suits a proof of concept and reporting that can tolerate Snowflake doing the work. Catalog federation runs the query against the table files on Databricks compute, which is the mode Databricks positions for incremental migration and long hybrid periods. Either way, a dashboard can be repointed at a Unity Catalog name while the table still lives in Snowflake, its numbers proved, and the storage swapped underneath later.
For the conversion itself, Lakebridge lists Snowflake as a source for both its Analyzer and its transpilers (including dbt repointing). The method is the same as for any warehouse, and Migration has it end to end: the inventory before the plan, the commands, reconciliation as evidence, and the federation connection's 90-day OAuth refresh token that will stop a federated catalog mid-project if nobody writes the date down.
What changes for a Snowflake estate is where the hard part sits. The SQL converts well. The gravity around it doesn't: tasks and streams that become Lakeflow pipelines or jobs, warehouses someone sized two years ago, and BI tools pointed at a specific Snowflake role, whose access has to be rebuilt as Unity Catalog grants before anyone switches.
How to decide #
We'd name the workloads and cost each one where it runs today. Separate the SQL that serves dashboards from the transformation that runs for hours, from the training, from the applications somebody wants to build next year.
Then ask which platform each would sit on if you were starting fresh, and whether the split is worth two bills. That exercise decides itself more often than it looks like it will, and it decides against migrating more often than a migration vendor's website will tell you.
| What the workload is | Where we'd run it | What would change our mind |
|---|---|---|
| Dashboards and ad hoc SQL, busy in office hours | Leave it on Snowflake | A second consumer of the same data, such as ML or an application, that would otherwise need a copy |
| Dashboard serving that scans all day | Off the warehouse, into a serving store, on either platform | Nothing; this is a modelling fix, not a platform choice |
| Transformation that holds compute for hours | Databricks, Lakeflow pipelines or jobs | It's a handful of short SQL transforms that already finish in minutes |
| Model training, agents, applications writing back | Databricks, with Lakebase for the operational data | The team has no plans past reporting |
| A mix of all four | One platform for the data, Iceberg tables shared where a second engine must read them | Two contracts are cheaper than one migration, which happens |
When it does decide the other way, the reasons are specific, and specific reasons are what keep a project alive in month four when it gets hard.
What's popular