Struggling with real-time geocoding performance on self-hosted data?

Hey everyone, circling back to our previous chat about achieving sub-millisecond IP geocoding latency. That discussion was super helpful for understanding the theoretical limits and external API options. Now, we're trying to achieve similar speeds but with our own self-hosted IP data, and it's proving to be a real beast. The core problem we're facing is needing extremely fast, real-time geocoding lookups for literally millions of requests daily. Relying solely on external APIs, while convenient, is just too costly at our scale and introduces a dependency we'd rather avoid. We've committed to a self-hosted solution, but our current setup isn't consistently hitting that crucial sub-millisecond target for IP lookup performance, which is vital for our user experience and analytics. So far, we've thrown quite a bit at it. We've been using the MaxMind GeoLite2 City database, experimenting with various database engines like PostgreSQL, ClickHouse, and even Redis for storing the IP ranges. We've also implemented aggressive caching layers, both in-memory and using Redis, for frequently accessed IPs. Of course, optimizing database indexes and query patterns has been a constant focus. We've even spun up high-spec cloud VMs to rule out hardware as a primary bottleneck, but the issue persists. Despite all these efforts, we've hit some significant limitations and challenges. While caching definitely helps, cache misses still frequently hit the database, leading to inconsistent latency spikes that push us past our target. Maintaining data accuracy and freshness with self-hosted solutions is an ongoing task, and updates, even incremental ones, can sometimes impact live query performance. We're also constantly battling the trade-off between accuracy (e.g., using more granular datasets) and raw lookup speed. A specific challenge has been efficiently querying IP ranges in traditional relational databases at such a massive scale; it feels like we're always fighting against the inherent structure when trying to optimize for IP lookup performance. So, my specific questions are: How are others in this community achieving sub-millisecond real-time geocoding latency with self-hosted IP datasets? Are there specialized databases, unique data structures, or indexing techniques particularly suited for IP range lookups that we might be overlooking? We're open to exploring anything that could give us that consistent performance boost. Also, any recommendations for optimizing the data update process without impacting live query performance would be a lifesaver. Has anyone successfully built a highly performant, self-hosted real-time geocoding service at scale and cracked this nut? I'd love to hear about your experiences!