From @ssastry,
The baseline http 500 error rate is usually because of m/w api timeouts, or timeouts on large pages (more common) or other uncommon crashers (which we usually catch, file bugs for and fix).
That periodicity is usually because the same failing pages are regularly crawled (by ex: Google) or edited (for ex: wiki pages that are used as "databases" / lists by bots and editors).
But, as of https://www.mediawiki.org/wiki/Parsoid/Deployments/2018#Wednesday,_Nov._7,_2018_around_1:54_pm_PT:_Deployed_970751a .. that number spiked ... since that baseline bump has persisted till now, we should look at Kibana and see what the http 500 requests are that don't fit the timeout or large page pattern.