The upstream phabricator[2] might also go away though I don't think that is quite as imminent.

It's perfectly fine to configure cluster.databases with one service, and Phabricator internally (in effect) builds a one-service cluster.databases configuration (by looking at mysql.user, mysql.host, etc) if your configuration is empty. Swapping from mysql.user to cluster.databases causes (or, at least, should cause) no change in behavior -- and is happening behind the scenes anyway -- it just gives you access to the persistent flag.

When:
[ Projects ][ include any of ][ X, Y, Z, ... ]
Take these actions [ only the first time this rule matches ]:
[ Send me an email ]

I suspect you're seeing that because when you reload a page by issuing a "Reload" command in your browser, most (all?) modern browsers interpret that to mean "Reload the page, skipping some/most/all caches". That is, the cache behavior of these sequences differs:

Eventually Phabricator ran out of local sockets (30k limit):

We need an easy way to tell logged-out traffic apart in our caching layer.

I'm happy to take a look at this if someone with access has a chance to figure out where it isn't working on your end -- my expectation is that it's supposed to work, but I could definitely have overlooked something.