OpenAdServer – A self-hosted ad server in Python, aiming to replace Revive" (github.com)

OpenAdServer is an open-source, self‑hosted ad serving platform built in Python (FastAPI) that combines production-ready ad stack features with ML-powered CTR prediction to offer a middle ground between simplistic banner servers and Google‑scale complexity. It supports DeepFM/Logistic Regression/FMx CTR models (PyTorch), real‑time eCPM ranking, multiple bid types (CPM/CPC/CPA/oCPM), targeting (geo/device/demographics/interests), frequency capping, budget pacing, and common ad formats (banner/native/video/interstitial). The system is deployable via Docker Compose or Kubernetes, exposes a simple API (/api/v1/ad/request, /api/v1/event/track, campaign/creative APIs), ships with training/evaluation scripts (AUC ~0.72 example), and includes observability (Prometheus + Grafana). Licensing is Apache 2.0 with no revenue share—ideal for SMBs, game studios, researchers and teams that want data ownership and full customization. Technically notable: the request pipeline is low‑latency (<10ms P99 for serving; <5ms for model inference), supports model hot‑swapping without downtime, automatic sparse/dense feature engineering, and an architecture using PostgreSQL for campaigns, Redis for cache, and PyTorch for online predictions. Benchmarks on modest hardware show thousands of QPS depending on model complexity (DeepFM ≈1,200 QPS). Roadmap items include OpenRTB, header bidding and multi‑tenant SaaS mode. OpenAdServer positions itself as a modern alternative to Revive and hosted ad networks by combining ML ranking, full control, and easy self‑hosting.