How It
Works
Use the expressive power of SQL along with the most advanced machine learning algorithms and pretrained models in a high performance database.
PostgresML indexes application data alongside computed features like embedding vectors with your machine learning models all together in a seamless shared memory space. This eliminates network calls, process boundaries, data duplication and other unnecessary complexity, which makes it more reliable, efficient, fast and simple.
Our serverless platform is built around our custom Postgres pooler PgCat, which allows you to scale your inference layer to millions of predictions per second across multiple GPU accelerated machines.
Try Now For FreeFast
Up to 40x performance improvement over traditional microservice architectures. Shared memory between data and models in a single process eliminates network calls, process boundaries and data duplication.
Comprehensive
PostgresML can download state-of-the-art open source models from HuggingFace, or train your own end to end. It also supports many algorithms like Torch, Tensorflow, XGBoost, LightGBM, and all the classical ones in Scikit.
Scalable
Scale your inference layer to millions of predictions per second in our GPU accelerated cloud. Start for free with our serverless option, or use dedicated Postgres replicas.
Simple
PostgresML is your application and vector database, model store, feature store, inference server and ML deployment pipeline, with support for all major languages and application frameworks as clients.
Your Data
Efficiently load data or features from upstream sources with Postgres replication, or connect your application directly through our custom load balancer, PgCat.
Highly Available
Abstract everything behind a single connection string with smart query routing, sharding, and enterprise-grade managed infrastructure.
How to Use It
Train
SELECT pgml.train(
'Sales Forecasting',
task => 'regression',
relation_name => 'historical_sales',
y_column_name => 'next_week_sales',
algorithm => 'xgboost'
);
Learn About Training
Deploy
SELECT pgml.deploy(
'Sales Forecasting',
strategy => 'best_score',
algorithm => 'xgboost'
);
Learn About Deployments
Predict
SELECT pgml.predict(
'Sales Forecasting',
ARRAY[
last_week_sales,
week_of_year
]
) AS prediction
ORDER BY prediction DESC;
Learn About Predictions
What Adopters Say
"Absolutely brilliant" on PgCat, our sharded PostgreSQL proxy.
"The improvement is quite remarkable" on PostgresML v2.0.
"Bleeding edge stuff in a matter of minutes."
"I'm itching to apply it everywhere I set my foot!"
"The simplicity and ergonomics are really exciting."


Start Now
With Machine Learning

Discover All Features
Fastest Inference
PostgresML eliminates separation between your model server and datastore, minimizing latency and computation costs. You can even generate embeddings on the fly in queries. Our benchmarks show a 8x-40x improvement over Python HTTP microservices.
OverviewAll Your Favorite Algorithms
Whether you need a simple linear regression or extreme gradient boosting, PostgresML includes support for all classification and regression algorithms in Scikit Learn, XGBoost, LightGBM and pre-trained deep learning models from Hugging Face.
AlgorithmsHyperparameter Search
Use either grid or random searches with cross validation on your training data to discover the most important knobs to tweak on your favorite algorithm, with best practices automatically enforced for testing model quality before deploying to production.
HyperparametersOnline & Offline
Predictions are served via a standard Postgres connection to ensure that your core apps can always access both your data and your models in real time. Pure SQL workflows also enable batch predictions to cache results in native Postgres tables.
Scheduled Training
Automate workflows to retain models periodically, taking advantage of auto scaling with dedicated resources like GPU's to optimize training times while alleviating loads on the primary database.
Open Source
PostgresML is entirely open source, including the ML libraries we support and the underlying database, Postgres (of course).
Open SourceManaged Deployments
Use either grid or random searches with cross validation on your training data to discover the most important knobs to tweak on your favorite algorithm, with the best practices enforced for testing modal quality before deploying to production.
Fast Vector Operations
Vector operations make working with learning embeddings a snap, for things like nearest neighbor searches or other similarity comparisons. Further optimized with BLAS for maximum performance.
Instant Visualizations
You can build common data visualizations to detect outliers, bimodal distributions, feature correlations, and much more.
We Have the Perfect
Plan For You
Serverless
$0
Starting at
For startups
Free, without cache acceleration
-
$0.25/hr per GB GPU cache
-
Multi GPU burst capability
-
Cache your models on the GPU
-
Instant scalability up to 256GB GPU
-
Scale further with advanced sharding functionality
Dedicated
$0.60/hr
Starting at
For orgs of any size
Dedicated cluster with fixed hardware
-
Choose CPU, RAM or GPU resources
-
Horizontally scalable inference with replicas
-
High availability for your production applications
-
Multiple users
-
Multiple databases
-
Automated Backups
-
Metrics
-
Logs
Enterprise
--
Custom Pricing
For orgs with custom needs
Your hardware, your way
-
Customized hardware
-
Solution Architecture support
-
Private VPC/On-prem deployments
-
Access Control Lists
-
Single Sign-on