Volga rewrites real-time AI/ML engine in Rust

This reads like a serious infra simplification play: Volga is trying to make real-time ML feel like a single SQL-native system instead of a stitched-together distributed stack.

• –Rust plus DataFusion is a credible way to cut the overhead of JVM-heavy streaming systems, especially for teams that care about latency and operational footprint.
• –The SlateDB-backed remote state story is the most interesting bet here because compute/storage separation could make rescaling and checkpoints far cheaper.
• –Request mode is a strong differentiator: serving point-in-time-correct features from the same dataflow could eliminate an entire feature-store/KV tier.
• –The ML-specific functions (`topk`, `_cate`, `_where`) suggest this is optimized for feature engineering rather than generic analytics.
• –The big unknown is maturity; joins, late events, and state movement are exactly where “clean architecture” meets ugly production reality.