Githubnumalogic
Background
Numalogic is a collection of ML models and algorithms for operation data analytics and AIOps. At Intuit, we use Numalogic at scale for continuous real-time data enrichment including anomaly scoring. We assign an anomaly score (ML inference) to any time-series datum/event/message we receive on our streaming platform (say, Kafka). 95% of our data sets are time-series, and we have a complex flowchart to execute ML inference on our high throughput sources. We run multiple models on the same datum, say a model that is sensitive towards +ve sentiments, another more tuned towards -ve sentiments, and another optimized for neutral sentiments. We also have a couple of ML models trained for the same data source to provide more accurate scores based on the data density in our model store. An ensemble of models is required because some composite keys in the data tend to be less dense than others, e.g., forgot-password interaction is less frequent than a status check interaction. At runtime, for each datum that arrives, models are picked based on a conditional forwarding filter set on the data density. ML engineers need to worry about only their inference container; they do not have to worry about data movement and quality assurance.
Numalogic realtime training
For an always-on ML platform, the key requirement is the ability to train or retrain models automatically based on the incoming messages. The composite key built at per message runtime looks for a matching model, and if the model turns out to be stale or missing, an automatic retriggering is applied. The conditional forwarding feature of the platform improves the development velocity of the ML developer when they have to make a decision whether to forward the result further or drop it after a trigger request.
Key Features
- Ease of use: simple and efficient tools for predictive data analytics
- Reusability: all the functionalities can be re-used in various contexts
- Model selection: easy to compare, validate, fine-tune and choose the model that works best with each data set
- Data processing: readily available feature extraction, scaling, transforming and normalization tools
- Extensibility: adding your own functions or extending over the existing capabilities
- Model Storage: out-of-the-box support for MLFlow and support for other model ML lifecycle management tools
Use Cases
- Deployment failure detection
- System failure detection for node failures or crashes
- Fraud detection
- Network intrusion detection
- Forecasting on time series data
Getting Started
For set-up information and running your first pipeline using numalogic, please see our getting started guide.
Installation
Numalogic requires Python 3.8 or higher.
Prerequisites
Numalogic needs PyTorch and PyTorch Lightning to work. But since these packages are platform dependendent, they are not included in the numalogic package itself. Kindly install them first.
Numalogic supports pytorch versions 2.0.0 and above.
numalogic can be installed using pip.
pip install numalogic
If using mlflow for model registry, install using:
pip install numalogic[mlflow]
Build locally
- Install Poetry:
curl -sSL https://install.python-poetry.org | python3 - - To activate virtual env:
poetry shell - To install dependencies:
If extra dependencies are needed:poetry install --with dev,torchpoetry install --all-extras - To run unit tests:
make test - To format code style using black and ruff:
make lint - Setup pre-commit hooks:
pre-commit install
Contributing
We would love contributions in the numalogic project in one of the following (but not limited to) areas:
- Adding new time series anomaly detection models
- Making it easier to add user's custom models
- Support for additional model registry frameworks
For contribution guildelines please refer here.
