Yesterday during asw-c2-eqiad switch replacement in T155875: asw-c2-eqiad reboots & fdb_mac_entry_mc_set() issues we saw a spike in 500s returned by varnish. Most of those seemed to come from mw api, thus dependent services (e.g. restbase) was affected too.
The DB servers behind asw-c2 (db1055, 56, 57, 59) were purposefully not depooled from mediawiki since this failure mode is something mediawiki should be able to handle gracefully.
Related incident: https://wikitech.wikimedia.org/wiki/Incident_documentation/2017-01-26_API_Slowdown