Request a 30-minute demo

Our product expert will guide you through our demo to show you how to automate testing for every part of your workflow.

See data diffing in real time
Data stack integration
Discuss pricing and features
Get answers to all your questions
By providing this information, you agree to be kept informed about Datafold's products and services.
Submit your credentials
Schedule date and time
for the demo
Get a 30-minute demo
and see datafold in action

How Qventus moved 1,300 dbt models from Snowflake to Databricks two months early, with zero disruption to hospitals

Highlights

  • ~1,300 dbt models moved from Snowflake to Databricks across three distinct codebases, while Qventus kept shipping on the Snowflake codebase
  • ~10x less engineering effort: roughly 30 person-months of planned engineering replaced by one part-time Qventus engineer per codebase reviewing Datafold’s output
  • Fully off Snowflake two months ahead of the contract deadline, with no disruption to the hospitals running on the migrated workloads
  • One platform, broader adoption: active users querying data on Databricks up 32% in the four months since cutover

Qventus planned to spend 3-4 dedicated engineers over three quarters to get off Snowflake, and its team doubted even that would beat the contract deadline. Instead, Qventus finished two months ahead of the deadline with one part-time engineer per codebase reviewing Datafold’s work. This is how.

About Qventus

Qventus uses AI to intelligently automate operations across care settings to remove the administrative burden from care teams and help hospitals deliver best-in-class patient care. Qventus’ AI teammates plug into EHR systems and use machine learning and behavioral science to predict length of stay, free up bed capacity, optimize OR utilization, and reduce surgery cancellations. Backed by more than $200M in funding through Series D, Qventus partners with more than 150 leading hospitals across the country.

Why Qventus moved to Databricks

Qventus chose Databricks because its workload was never a warehouse workload. EHR ingestion, per-client derived pipelines, ML models, and now agents all needed one substrate. Snowflake had become a second platform with a second copy of the data, a second governance model, and a second bill.

The data was scattered. Raw EHR feeds landed in one place, transformed models lived in Snowflake, ML training and feature work happened on Spark somewhere else, and product and services teams built their own objects outside version control to fill the gaps. Every new client meant re-plumbing that path by hand. Moving faster carried real overhead: keeping two platforms in sync, reconciling definitions across them, and paying people to move data between engines instead of building on it.

A few things drove the decision on engineering grounds:

  • Fewer hops. Every hop between the lake, the warehouse, and the ML environment was a networking bill and a copy to keep consistent. Collapsing to one platform removed most of that traffic.
  • Notebooks against the real tables. Data science, analytics engineering, and data engineering needed to run transforms, feature engineering, and model training on the same storage without staging data out first, not an export-and-analyze loop.
  • Raw data landed once and kept. Qventus wanted to reprocess history whenever a model, a metric definition, or a client config changed. A warehouse-first design pushes teams to curate early and discard the raw layer. The lakehouse keeps Bronze intact and lets Silver and Gold evolve on top of it.
  • Governance on one surface. Unity Catalog gave Qventus one permission model, one lineage graph, and one audit trail across tables, models, functions, and agent tools.

An approaching Snowflake contract end date set the window to make the move.

The migration nobody wanted to staff

Deciding to move was the easy part. Leaving Snowflake entirely meant migrating roughly 1,300 dbt models spread across three very different codebases:

  • Qventus’ core dbt project, 600+ models built on a structured foundational data model
  • Its inpatient workloads, ~330 models built over time without a shared foundational data model
  • A legacy environment of ~330 models on an older version of the platform

And it had to happen under real constraints. The Snowflake contract had a fixed end date, so the deadline was not negotiable. As a fast-growing company, Qventus could not pause development for the better part of a year, so the migration had to track a moving target. And because its models run on protected health information (PHI) pulled from hospital EHR systems, even small discrepancies at cutover could erode clinical and customer trust. Validation was not optional.

Doing it in-house meant standing up an LLM translation pipeline, a validation harness, and a rollout process from scratch, then running all three for the better part of a year with no guarantee of finishing in time.

How the migration worked

Datafold delivered the migration as a Databricks Delivery Provider partner, giving Qventus a single path from Snowflake to Databricks. The work ran on the Datafold Migration Agent (DMA): an autonomous migration system that ingests an entire dbt project, plans the work, translates the code into Databricks-native dbt, executes every model on the target platform, and proves equivalence with cross-database data diffs, iterating on its own output until the data matches. The same system carried all three codebases from start to finish.

“Migrating Qventus from Snowflake to Databricks required both speed and absolute precision. Datafold provided an end-to-end solution: their automated code conversion accelerated the dbt migration, while their data diffing replaced manual spot-checks with systematic, column-level validation. This combined approach allowed us to migrate production healthcare workloads at scale with total confidence in data consistency.”

Adithya Nanduri, Solutions Architect, Databricks

DMA preserved the structure of each dbt project. Models, sources, the references between them, and the project layout all stayed the same. Working with full project context, including the dependency graph between models, the agent rewrote only what had to change between platforms (SQL dialect, function semantics, platform-specific edge cases), then compiled and ran every model on Databricks, as a full rebuild and, for incremental models, as an incremental load, to prove the translation worked before delivery.

Because Qventus kept developing on Snowflake right up to each cutover, Datafold also kept the migrated code in sync, re-translating and re-validating changes on a weekly cadence. Ahead of the final inpatient cutovers, Datafold ran a full drift analysis across the project, classifying more than 2,300 differences across 470+ files and automatically retranslating or preserving each one, so Qventus never had to pause development for the migration.

Validation happened in two stages.

For every translated model, DMA ran a cross-database data diff comparing the Databricks output against its Snowflake counterpart, column by column and value by value, not just row counts or schemas. When it found a mismatch, the agent diagnosed the root cause and corrected its own translation, iterating until the outputs converged. The agent escalated to Qventus only the questions that needed real business context. Each model shipped with its data diff attached, an audit trail Qventus could inspect rather than a translation it had to take on faith.

Then, ahead of each cutover, Datafold stood up continuous data diff monitors comparing both platforms running side by side in production, so Qventus could cut over knowing exactly where the two systems agreed and where they did not.

From a 20-model proof to the full exit

The engagement started small and grew as Qventus saw what the system could do.

Proof of concept. Datafold migrated ~20 dbt models end to end, translated, executed on Databricks, and validated with data diffs, to show Qventus what a fully automated migration would look like on its own code. That gave leadership the confidence to commit to the full migration.

The core project. Datafold’s first full assignment was Qventus’ most important codebase: the 600+ models built on its foundational data model. DMA translated 95% of the project with essentially no human input. The remaining 5% needed business context that only Qventus could provide. From there, the work was proving correctness: validating all 600+ models against production data, with Qventus’ team reviewing and signing off on each batch, while Datafold kept the translated code current with an actively developed Snowflake codebase. Qventus cut the hospital customers running on that codebase over to Databricks in mid-2026.

The inpatient workloads. With the core project validated, Qventus pointed Datafold at the harder problem. The inpatient project, ~330 models, was built without a foundational data model, a fundamentally different architecture that made translation more complex. DMA translated the full project in roughly two months. Qventus then cut its hospital clients over to Databricks one at a time, on its own schedule, with Datafold provisioning data diff monitors for each client’s environment so every cutover was backed by a live comparison against Snowflake. As Qventus onboarded new hospitals mid-migration, Datafold added them to the same validation pipeline. Qventus cut its final inpatient clients over ahead of the December 2026 Snowflake deadline.

The legacy environment. By the last codebase, ~330 models on an older version of the platform, the process was well established. Once Qventus provisioned access, Datafold delivered the translated project within about two weeks. Qventus reviewed it, Datafold fixed the issues flagged in review, and the pull request was merged. The legacy environment cut over to Databricks at the end of September 2026. Monitors followed the same pattern as the other two codebases.

The results

Qventus initially scoped the exit at 3-4 dedicated engineers for three quarters. With Datafold, it took roughly one part-time Qventus engineer per codebase, reviewing translated code and validation results rather than building and running the migration itself. Qventus redirected that capacity to building new products and shipping them to customers.

Qventus is fully off Snowflake, two months ahead of its contract deadline, an outcome its team once doubted was possible. Roughly 1,300 dbt models now run on Databricks, validated against their Snowflake counterparts. The hospitals relying on those models saw no disruption. Models tuned for Databricks compute now deliver faster dbt builds and query times.

The move also changed who uses the platform. The number of active people querying data on Databricks is up 32% in the four months since cutover. One platform with notebooks, SQL, and governed tables in the same place turned out to be something more of the company wanted to use.

Platform management overhead on the data platform team is down about 5%, with Databricks support absorbing work that previously landed on Qventus engineers: cluster tuning, upgrade handling, and the troubleshooting that used to consume platform time.

Conclusion

Qventus set out to leave Snowflake with a plan it doubted would finish on time. It finished two months early, with a fraction of the engineering it had budgeted, and with no disruption to the hospitals running on the migrated workloads. Raw data now lands once and stays, governance lives on one surface, and ingestion, transformation, ML, and agents run on the same platform. The engineers who were slated to spend three quarters on the migration spent them building product instead.

About Datafold

Datafold delivers the fastest migrations to modern data platforms, at a fixed price and on a guaranteed timeline, through end-to-end AI automation. Customers get a finished, validated migration without having to staff one.

Challenge
Snowflake Snowflake
~1,300 dbt models
Databricks Databricks
dbt dbt
Outcome
Fully off Snowflake ahead of the contract deadline
2 months early
~10x less effort
32% more active users
Migrated ~1,300 dbt models across three codebases with zero disruption to hospitals