Let's build a Real Time ML System
Dec 02, 2024
Today I want to show you how to design a real time ML system that can predict crypto prices 5 minutes into the future, using
-
Python and Rust to build our micro-services.
-
Apache Kafka for real time data streaming
-
LLMs for sentiment extraction from unstructured data,
-
Kubernetes for a production-ready deployment
Let’s start!
Our goal 🎯
Let's build an ML system that can predict crypto prices 5 minutes into the future.
Attention 🚨
Predicting short-term crypto price movements is very hard.
If it wasn’t everyone would be rich. And, as far as I know, not everyone is rich.
I am not rich for example.
To predict this target we want to use ML, and for that we need reliable data sources that can help us generate the best predictions we can.
In this case we will use 2 real-time data streams
-
Crypto trade data
-
Market news about blockchain, or any economic factor that can possibly impact the target metric we want to predict.
System design 📐
Let's break down our system into 4 types pipelines
-
Feature pipelines (2 of them)
-
Training pipeline
-
Inference pipeline
-
Monitoring pipeline
Let's go one by one.
1. Feature Pipelines 💾
Our 2 feature pipelines transform raw data into reusable ML model features, and save them in our Feature Store.
Feature pipeline for technical indicators
-
Ingests market trades from a websocket API like Kraken or Binance,
-
Aggregates them into 1-minute windows
-
Engineers technical indicators to capture momentum in the market, and
-
Pushes them to the Feature Store
Feature pipeline for market sentiment
-
Scrapes news by pooling a website like coinbase,
-
Parses the raw text into a structure sentiment score using an LLM, and
-
Pushes them to the Feature Store
2. Training pipeline 🏋🏽
We can use a boosting tree model like XGBoost to uncover any patterns between the
-
current technical indicators, and
-
market senttiment
and
-
the price of BTC or ETH 5 minutes into the future (target metric)
The final model is pushed to the model registry, so it can be loaded by our next pipeline: the inference pipeline.
3. Inference pipeline 🔮
Finally, we need to generate and make fresh predictions accessible. This is what the inference pipeline does.
To serve the final predictions, while making them easily accessible to our next pipeline (the monitoring pipeline) I recommend you split this step into 2 services:
Prediction generator
-
Loads the model from the registry and
-
Continuously reads the latest features from the Feature Store,
-
Generate a new prediction and
-
Pushes this prediction into a DB, like an Elastic Search index.
Prediction API
This is a lightweight service that
-
Receives incoming requests from client apps,
-
Finds the predictions in the Elastic Search index, and
-
Returns the prediction to the client app.
Bonus 🎁
If latency is critical, you can build a websocket API, instead of a REST API, to avoid reading and writing to Elastic Search.
4. Monitoring pipeline 🔎
There are lots of things you can monitor here. The most fundamental one is the error of your predictions.
To monitor these you need to build a streaming service that
-
Listens to incoming price data
-
Loads the predictions from the DB
-
Computes the error and
-
Pushes the error into another Elastic Search index
These errors are plotted on a dashboard using Kibana, and trigger alerts to your Slack or Discord team.
BOOM!
Wanna build this system with me?
On December 2nd, 213 brave students and myself will start building this system, in my course Building a Real Time ML System. Together.
Live. You and me. Step by step.
It will take us at least 4 weeks, and 40 hours of live coding sessions, to go from idea to a fully working system, that we will deploy to Kubernetes.
Along the way you will learn
-
Universal MLOps design principles
-
Tons of Python tricks
-
Feature engineering in real time
-
LLMs engineer market signals from unstructured data
-
Some Rust magic
-
.. and more
Talk to you next week,
Wish you a great weekend,
Pau
