Right now, we have wt2html and html2wt metrics on a per-request basis. However, in reality, these make it somewhat hard to use for understanding how performance changed over time since they depend on the composition of the requests. Lots of small requests might skew the metrics lower even when there has been no real performance gain and likewise if more large requests enter the mix, we see an upward trend which can give a false sense of performance degradation.
One idea is to track (a) time per KB of output HTML for wt2html requests (b) time per KB of processed input HTML for html2wt requests.
This will insulate metrics from being skewed by input/output sizes as well as the composition of the request set. As long as there are no serious contention effects from high request volumes, these metrics might provide us with a better handle on tracing performance changes over time.
Additionally, by pegging the metrics to HTML size rather than wikitext size, we get a more accurate performance picture since templates and extensions can have a small footprint in input and a large footprint in output.