Two reads can disagree
Deleted a real 2x speedup, because two paths can disagree.
There were two reads: a cheap one behind the summary card, a full one behind the report. Roughly a 2x win. Then I saw that the two could return different answers for the same number, which is the failure I had been trying to design out.
I removed the fastest thing I had built that week and nobody will ever know.
THE SETUP
Two paths, one number
The summary card used a cheap read. The report used the full one. Measured against each other, roughly a 2x speedup.
THE PROBLEM
Divergence, not latency
The two reads could disagree. A system where two views of the same number can disagree is exactly what I was trying to get away from. The speed was real; it was not the thing at risk.
THE CHANGE
Deleted the fast one
I removed the cheap read on 14 August. Correctness over a 2x win, in a place where nobody but me was ever going to notice.