The Python library for data science and AI that runs where your data lives.
In-database data science & AI¶
VAST Orbit lets data scientists, analysts, and ML engineers run a complete workflow — preparing, exploring, analyzing, and modeling data — directly inside the VAST AI OS, using the pandas- and scikit-learn-style API they already know. Instead of copying data out to a notebook, VAST Orbit pushes the work down to where the data sits and brings only the answers back, so the same code runs on a kilobyte or a petabyte, against a table or a Parquet file, without a rewrite.
What VAST Orbit brings to your business¶
Most data science works backwards: the data is huge and stays put, while the tools are small and insist everything be copied to them. VAST Orbit removes the copy — preparation, analytics, and scoring all run inside VAST — and that single change pays off differently for everyone who touches the data.
Your tools and your syntax, at any scale.Keep the pandas and scikit-learn habits you already have. VastFrame behaves
like a DataFrame and the models follow the scikit-learn API, so the notebook you
write against a sample becomes production code against the full table — no
rewrite, no new language, and no waiting on an export before you can think.
Because the work runs in VAST, the size of the data stops being your problem: the same few lines hold whether you are working through a megabyte or a petabyte.
One governed surface instead of pipeline sprawl.Since the work executes where the data lives, there are no extracts to copy and no side pipelines to babysit. Sensitive data stays inside VAST under one set of controls, lineage stays intact, and the platform you already run becomes the single place analysts, models, and agents all work.
That means far less to secure, move, and reconcile — and far fewer moving parts between a question and its answer.
Answers sooner, on more of your data, for less.Every copy is time lost, compute and storage paid for twice, and another place data can leak. Removing it means questions are answered sooner, against more of the estate at once, at a fraction of the infrastructure cost.
A central data strategy stops being an architecture diagram and becomes something teams use day to day — on a single table or every asset across the business.
Key features¶
Preparation, analytics, and scoring execute inside VAST. Your data never leaves the platform, so work stays fast, governed, and secure.
VastFrame behaves like a pandas DataFrame and the models follow the
scikit-learn API, so notebook code becomes production code unchanged.
One federated query joins VAST tables, data-lake files, and external databases — with no ETL pipeline to build first.
VAST’s flash-native, disaggregated architecture keeps queries interactive from gigabytes to exabytes, so workflows grow with your data.
Train on intelligently sampled data, then deploy the model for in-database inference that scores billions of rows where they live.
A built-in MCP server and the %%ai and %%sql notebook magics let
assistants and LLMs query and reason over your VAST data directly.
What you can do with it¶
Clean, shape, and understand data at any scale.Cleaning and shaping data is usually the slowest part of a project, and it only gets slower as the data grows. With VAST Orbit you handle missing values, remove duplicates, normalize, encode, and engineer features using familiar calls that execute inside VAST — so a petabyte table feels like a megabyte one and nothing is pulled into Python first.
When it is time to look before you leap, charts and statistical summaries are generated from intelligent samples, letting you profile distributions, correlations, and outliers across billions of rows in seconds rather than waiting on an export.
Join everything, in one query, close to the data.Real questions rarely live in a single table. Because VAST Orbit reaches your data through one federated query engine, one piece of Python can join a VAST table with Parquet files in the data lake and a customer record in PostgreSQL — all in the same query and all executed next to the data.
That turns the idea of a central data layer from an architecture diagram into something an analyst can use day to day: ad-hoc analysis across the whole estate, with no ETL job to schedule and no warehouse to load first.
Train flexibly, then score billions of rows in place.VAST Orbit supports a hybrid workflow that matches how teams really work:
train quickly with the embedded algorithms or bring a scikit-learn model you
trained locally, then deploy it for inference that runs as SQL inside VAST.
Scoring happens where the data is, so you can refresh predictions on live data continuously — fraud scores, churn risk, demand forecasts — instead of shuttling features and results between systems and serving infrastructure.
Make your data a first-class tool for AI.Modern AI is only as good as its access to context. VAST Orbit exposes your
data and its own analytics as tools an LLM can call through an MCP server, and
the %%ai magic lets you ask a question in plain language and get back a
working, schema-aware query.
The result is a single governed platform where analysts, models, and agents all operate on the same data, with no copies in between — the foundation for AI applications you can actually trust in production.
Use cases¶
The same library powers very different problems, because they all come down to running analytics and models close to large, fast-moving data.
Score every transaction against historical patterns in place and catch fraud in real time — without exporting sensitive financial data.
Combine CRM, billing, and clickstream into one view and predict churn where the data already sits, then refresh scores continuously.
Optimize coverage and capacity over billions of call-detail records, joining live traffic with historical patterns in a single query.
Monitor sensor streams and forecast failures across live and historical readings together, scoring devices where their data lands.
Price risk and flag anomalies across policy, claims, and external data sets at once, with full lineage and no data leaving the platform.
Explore genomic and clinical data at petabyte scale with familiar Python, keeping regulated data governed inside VAST throughout.
A quick look¶
Connect once, then prepare, join, visualize, and model — every step executing in VAST:
import vastorbit as vo
# Connect to VAST
vo.new_connection({
"host": "vast-cluster.example.com",
"port": 8080,
"catalog": "vast_catalog",
"schema": "analytics",
})
# A VastFrame is a handle to data in VAST — nothing is pulled into Python
customers = vo.VastFrame("vast_catalog.crm.customers")
# Prepare the data, in place
customers = customers.fillna({"income": 0, "age": customers["age"].avg()})
customers["income_norm"] = customers.normalize("income")
# Explore with intelligent sampling
customers["age"].hist(nbins=20)
# Join a VAST table with data-lake files — one query, executed in VAST
transactions = vo.VastFrame("hive.default.transactions")
enriched = customers.join(transactions, on="customer_id", how="inner")
# Train a model, then score billions of rows in the database
from vastorbit.machine_learning.vast import RandomForestClassifier
model = RandomForestClassifier(n_estimators = 4)
model.fit(enriched, ["age", "tenure"], "churn")
predictions = model.predict(enriched) # runs inside VAST
Built on the VAST AI OS¶
VAST Orbit is only as capable as the foundation beneath it. The VAST AI OS unifies storage, database, and compute into one consistent system, so every asset — transactional tables, data-lake files, streaming events, and vector embeddings — lives in one place and speaks one language. When the infrastructure is that consistent, everything becomes possible: there is no copy step to slow you down, no second system to reconcile, and no scale ceiling to design around.
VAST Orbit turns that consistency into a single queryable surface for Python, where
one VastFrame reaches a table or a file, a gigabyte or an exabyte, all the same
way. It works with VAST 4.5 and later. Learn more about the foundation it builds
on at the VAST AI OS.
Today that single query runs on Trino; VAST’s own query engine is on the way and will
become the default. Because you work through one VastFrame API, your code stays
exactly the same when the engine underneath it changes.
Installation¶
Install the beta with pip:
pip install vastorbit
VAST Orbit needs Python 3.12+, network access to your VAST cluster, and VAST 4.5 or later. Version 0.1.x is a beta; a production-ready 1.0.0 is on the way. See Getting Started for full setup and Connection for connection and authentication options.
Note
Placeholders such as vast-cluster.example.com and vast_catalog stand in
for your own cluster host and catalog names throughout the docs.
Explore the documentation¶
Install VAST Orbit and run your first in-database query in minutes.
Connect to VAST, authenticate, and reach every catalog.
Master VastFrame and the in-database operations behind it.
Train models and deploy them for inference at production scale.
See the visualizations VAST Orbit can build from sampled data.
The library by the numbers — measured live from the source.
Follow end-to-end tutorials across analytics and ML workflows.
Meet the people behind VAST Orbit and the story that shaped it.
Note
VAST Orbit brings Python data science to the VAST AI OS — query anywhere, analyze everything, and build AI at any scale, all with in-database execution and zero data movement.