ScaR: Towards a Real-Time Recommender Framework Following the Microservices Architecture

ScaR is an open-source real-time recommender framework designed around the microservices architecture, aiming for both scalability and flexibility in handling dynamic, high-volume data streams. The core idea was to move beyond traditional monolithic recommender systems—where updating models or adapting to new contexts can be prohibitively slow—and instead decompose the system into loosely coupled services with simple REST interfaces. Each microservice in ScaR (covering data modification, recommendation logic, customization, evaluation, and client-facing proxying) runs independently and can be scaled or replaced as needed. Apache Solr serves as the backbone for both data storage and search, leveraging its near-real-time visibility and efficient indexing to provide prompt recommendations immediately after user events occur.

A key feature is that ScaR embraces frequent incremental data updates, rather than relying on batch recalculation—a pain point in matrix factorization methods or other model-based approaches. Solr’s configurable commit strategies (hard/soft commits and even real-time get) allow data to be visible for recommendation within seconds, without costly downtime.

We evaluated ScaR’s scalability using a Foursquare benchmark, simulating hundreds of thousands of real-world recommendation requests under varying deployment scenarios. In a single-machine deployment, the system processed up to 161 recommendations per second before bottlenecking. Simply adding additional microservice replicas and distributing them across two servers increased throughput by up to 83%, reaching 295 requests/sec with consistent response times below 200 ms for even the most complex hybrid recommendation strategies. These results show that the microservices approach delivers tangible horizontal scalability for real-time recommender workloads.

Practically, ScaR has already seen production use in commercial environments, but the framework is also fully open for research adaptation. Ongoing work explores alternative backend technologies (such as graph databases), dynamic hybridization of methods, and further containerization for rapid deployment. The real value here is the modularity: each recommendation component can be swapped, retrained, or reconfigured independently, making ScaR a flexible base for both experimentation and scalable deployment in demanding, constantly evolving environments.

@inproceedings{lacic2015scar,
  title = {ScaR: Towards a Real-Time Recommender Framework Following the Microservices Architecture},
  author = {Laci\'{c}, E. and Traub, M. and Kowald, D. and Lex, E.},
  year = {2015},
  booktitle = {Proceedings of the Workshop on Large Scale Recommender Systems at the 9th ACM Conference on Recommender Systems},
}