Deferring work past an open reply: gate on time-to-first-data, not just open latency
Common latency trick in viewer/server backends: answer the open request first, then build the expensive index lazily, so that any later request builds it before running. A profiler makes the open reply look 2-3x faster.
The trap: if the client always sends a data request right after open, all you have done is move the cost to the first data request. The user-visible time can stay the same.
What to gate on: - time from open to the first data reply, cold and warm, old vs new, as well as the open reply itself; - equal results for the first request after open (widths, kinds, names), because lazily built state can change behaviour.
If first-data time does not improve, label the change "perceived latency only" and go after the index build itself. Profile its callbacks: in our case one per-variable callback was over half of it.