@fkaelin recently did some more analysis on 3 months of revision html -- details in slack. This analysis confirmed that the 99.9 percentile is still small enough, around 2.3MB.

We'd like to also have maximum values for all pages on at least one large wiki, but all (wikitext) wikis is better.

From @ssastry in slack:

the rule of thumb from what I remember from long back is: parsoid html is probably 7x the size of iinput wt ... and given that 2mb is the wt cap, 14mb html is probably a good ballpark estimate for upper limit, but that said, there are pages (not in article namespace) out there that can generate large blobs of html. Plus, this metric might be different for wikis with languages that have multi-byte encoding of its characters unlike say english.

I added some graphs to https://grafana.wikimedia.org/dashboard/snapshot/xzDtG1FZnWCRRaFBQKw2oxpmzK8EUZec -- the tl;dr is a p50 on output size of ~38kB, a p90 of 270kB, and a p99 of 977kB. There are 9 buckets from 2.2k up to 1Mb, so those match reasonably well with the outputs, although the upper bound is probably not terribly reliable given how close it is to our top bucket limit.

Change #1254205 had a related patch set uploaded (by Ottomata; author: Ottomata):

[operations/puppet@production] Increase the kafka-jumbo maximum message size to 20MB

https://gerrit.wikimedia.org/r/1254205