We can no longer blame the rendering engine for dropped frames. With Impeller now the standard, the "shader compilation jank" of the past is gone, yet high-scale production apps still struggle to maintain a consistent 120Hz. The core tension has shifted from the engine to our own architecture: we are building features so complex and data-heavy that we silently overload the UI thread, choke the rasterizer, and blow through frame budgets regardless of how fast the underlying graphics layer is. This roundtable is a "war room" for performance tuning in a post-Impeller world. We will move beyond basic optimization tips and focus on the hard reality of hunting down application-level jank using DevTools and tracing. Bring your toughest performance mysteries.. whether it’s list rendering at scale, main-thread bottlenecks, or custom layout architectures and let’s debate how to achieve true "zero jank" when the default widgets aren't enough.

Every offline-first tutorial gives you a boolean: online or offline. Production gives you a third state, and it's the one that breaks apps — the connection reports "connected," the request hangs for 30 seconds, and your FutureBuilder sits on a spinner forever. That ugly middle is the default at every conference venue, which is exactly where ConnectMachine (22,000+ users, 4.7★) gets used hardest. This is a walkthrough of the offline architecture we actually ship, in three layers. Reads. A cache-first fetcher over Hive with a central TTL policy table — one minute for transcriptions, seven days for the user profile — plus in-flight request de-duplication so a rebuild storm doesn't fire the same call five times. We chose hard-expire when offline: an expired entry throws rather than serving unbounded stale data, and TTLs are generous to compensate. I'll defend that choice and show where it hurt. Connectivity. Why we replaced the online/offline boolean with NetworkStatus { online, unstable, offline }, derived from a latency heartbeat and consecutive-failure counting rather than from the OS's idea of connectivity — and how a broadcast Stream of that status changes what your widgets build. Writes. Durable queues for card scans, recording uploads, and deep links, with a state machine that knows the difference between "retry this" and "only a human can fix this." Plus schema-versioned, user-scoped cache keys so an app update or an account switch doesn't resurrect the wrong data.