How to find and export your Databricks data for AI training

01

What is Databricks data, and why do AI labs and data vendors buy it?

What a Databricks workspace records

Databricks is a cloud platform that combines a data lake, which keeps raw files cheaply, with a warehouse, which keeps organized tables. That combination is called a lakehouse. A company's workspace is its own Databricks environment, and the tables in it are usually stored in the Delta format, which keeps the data in compressed files together with a log of every change made to them.

Lakehouses often hold the largest and rawest data a company keeps. That includes streams of events from software, logs from machines and sensors and the outputs of data pipelines (automated jobs that move and clean data), kept next to the curated tables built from them. Unity Catalog is Databricks' system for listing those tables and controlling who may see each one.

Why raw and modeled history teaches an AI model something

An AI model learns by studying a very large number of examples and finding the patterns in them, and those examples are called training data. Raw logs show exactly what happened and when, and the curated tables built from them show what the business concluded. Studying both lets a model see how real operations turn into recorded results, which public text on the internet almost never shows.

AI labs now want to train AI agents, software that carries out a job step by step, and an agent learns a job from records of that job being done. A data vendor can turn anonymized event histories into worked examples that an agent studies, or into a practice copy of an operation in which the agent is scored on each attempt. The full mechanism is in AI training data: why AI companies want your company's data.

How Databricks fits into a dataset

A Databricks workspace is usually one of several sources in a company's dataset, next to email, documents and tickets. Companies are paid $100K to $2M or more per dataset, according to micro1, the data vendor we typically refer companies to. The price covers the whole dataset, not each source on its own, and what your data is worth explains what moves it.

02

Where does Databricks show how much data you have?

How to read your storage total

The steps below are run by a developer or data engineer in a SQL warehouse or a notebook, the two places in Databricks where SQL commands are typed. They follow the guide to reviewing Delta table details.

1. Make sure you have Unity Catalog access to the tables you want to measure. The command works in any workspace that has Delta tables (it also works on Apache Iceberg tables). Reading the real usage of the underlying cloud storage requires access to the cloud account behind the workspace.
2. List the tables in each schema, the folder inside a catalog that groups related tables:

   SHOW TABLES IN my_catalog.my_schema;

3. For each table, run the command below and note sizeInBytes, numFiles, createdAt and location:

   DESCRIBE DETAIL my_catalog.my_schema.orders;

4. Add up the sizeInBytes values and divide by 1,073,741,824 to get GB. The result describes only the newest version of each table, so the folder that holds it in cloud storage can be bigger.

What DESCRIBE DETAIL returns

Why the number can fall short

The command reports on one table at a time, so a table you forget is missing from the total. Older data files that Delta keeps so you can look back in time stay in cloud storage until they are removed, which is why the folder is larger than sizeInBytes. Report the table sizes and not the size of the storage path.

03

What should you know about your Databricks data before you apply?

The size in GB

Write down the sum of sizeInBytes across the tables in GB. Each table's data files are compressed, so the same data written out as CSV would be larger. Don't report the size of the whole storage path, because that also contains older file versions kept for time travel.

The years of history

The years of history come from the earliest event date in the data, or from the table's createdAt value when the tables carry no dates.

The number of seats

The seat count is the number of users or accounts that the tables describe. The people who log in to Databricks are not seats.

What you don't need to do first

You don't need to export anything to apply, and you don't need to clean or scrub the records. The data vendor anonymizes them before anyone else sees them, and nothing moves until a lawyer-reviewed license is signed. You only need the size, the years and the count for the first conversation. The application, the intro call and each later step are laid out in how to sell your company's data.

04

How do you export your full Databricks history?

The steps with a deep clone

Each step is run by a developer, with the administrator's help for the first. Databricks describes the command in its guide to cloning tables.

1. Ask a workspace administrator for an external location that points at a bucket or container you control, and ask that it come with permission to create tables in it.
2. Clone each table into that location with a deep clone, which brings across its current data and its metadata:

   CREATE TABLE my_catalog.export.orders
   DEEP CLONE my_catalog.my_schema.orders
   LOCATION 's3://my-export-bucket/orders';

3. Move the files out of that path with your cloud provider's tools. On Amazon S3, for example, aws s3 sync copies the folder to a machine:

   aws s3 sync s3://my-export-bucket/orders /data/export/orders

Why you clone a managed table instead of copying its files

A managed table is one whose files Databricks itself controls, as opposed to an external table, whose files sit in a location you chose. Databricks does not support reading a managed table's files directly by path, except in compatibility mode, because that would bypass the access controls of Unity Catalog. Cloning the table to an external location is the supported way to get its files out. See the guide to Unity Catalog managed tables.

What a clone includes and leaves out

A deep clone copies the table's current data and its metadata, but it does not copy the Delta table history, the table description, any custom commit notes or the Unity Catalog tags. Earlier states of the table are therefore not part of the export.

A dropped managed table is deleted from storage after a recovery period, and UNDROP TABLE brings it back until then. The default is 7 days, and the period can now be changed at the catalog or schema level to 0 hours or to between 7 and 30 days, as a Public Preview feature.

How long an export takes

Databricks publishes no time, because it depends on table size and the compute you run the clone on. Each table is a separate statement, so a large workspace is exported as a series of clones.

Sources: Databricks documentation, Review Delta Lake table details with DESCRIBE DETAIL, Unity Catalog managed tables, Clone a table and the CLONE syntax reference. Steps checked against these pages in November 2026.

05

What stays private when Databricks data is licensed?

You choose what is in scope

A table reaches the export location only if you clone it, so a catalog or schema holding sensitive records can stay in the workspace. The permissions that Unity Catalog applies to your account also limit what you can clone. The data vendor deletes records too sensitive to clean safely instead of anonymizing them.

Names are replaced before a lab sees them

Anonymization means replacing every detail that identifies a person or a company, such as a name or an email address, with a stand-in that is the same every time it appears. The records still link to each other and the dates still run in order, so a lab can learn from them, but they no longer say who anyone is. The diagram shows one set of records before and after.

The method, including what is deleted outright and how you review samples, is in how company data is anonymized for AI training.

Ownership and payment

Licensing gives the buyer limited rights to a copy, while the company keeps ownership of the originals. The terms that matter, such as exclusivity, permitted use and deletion, are covered in data licensing. micro1 pays in full within 30 days of accepting the data, usually as a single payment.