Investigating what data sources to use for T201656, I considered options for getting data more frequently than what wmf.mediawiki_history would enable. One idea was to use the MW database, another was to use event.mediawiki_revision_create. I wanted to check whether mediawiki_revision_create stores all revisions or not. Comparing the random month of October 2018 against the history table for English Wikipedia, I found 70,719 revisions were missing (or 1.5%). Investigating some of them in the MW database, I did not find a clear pattern, but on the other hand I have not been digging very deeply.
I'm unsure about what the expected reliability of an EventBus table should be, so I thought I'd open this task to discuss it.
Update
After improvements to envoy proxies, we found that we are missing 0.0014% of events. The remaining missing events don't seem possible to fix in the current architecture, we would need some other way to send them more reliably or reconcile them later.