Sketches / ML Systems
Two clocks
Production latency is also the gap between the state a model reads and the state a report shows later. A correct call can look like a failure if you score the wrong clock.
Synthetic. Move the report. Watch a rule hold at decision time and break before anyone reads the numbers.
Synthetic · two clocks
Decision-time truth
9:29 · rule holds · allow
Report-time truth
9:43 · rule holds
The report is still on the same side of the change. Both clocks agree.
Inference is the short mark
How long the model takes to run is one clock. It is not the only one.