Drift Detection with a Low Memory Footprint for ML Models on Kafka Streams

Drift is a known and widespread problem for Machine Learning models, doubly so with Data Streaming. In Radicalbit we devised a solution to this issue for models leveraging Kafka streams for input and output data. The advanced Machine Learning models monitoring we developed for our MLOps platform allows us to detect drift in a Kafka topic fed into the model. The main advantage of our solution lies in its very low memory footprint: while such a feature is important for any computing solution, it is especially valuable in situations where hundreds of messages per second are received, like ours. Furthermore, we can combine this solution with our alerting system: when drift is detected, a new dedicated Kafka topic is populated; this is associated with a webhook that sends alerts to a channel of choice, e.g. Slack, Telegram bot, etc.