mcp server health checks - polling vs event-based patterns

We've been running 6 MCP servers in prod for 3 months and health checking is still a mess. Currently polling every 30s with a lightweight ping tool, but wondering if there's a better pattern. Event-based feels cleaner but adds complexity. Anyone doing this differently? Our failure mode is silent degradation where the server responds but returns stale data, so simple up/down checks aren't enough.