All relays are constantly connecting to down relays and failing over and over

Starlight at https://lists.torproject.org/pipermail/tor-relays/2017-December/014001.html points out that when a relay is receiving many extend cells to extend circuits to a relay that's down, it will try over and over to make a new connection to that next relay:

650 ORCONN $5E2975B380F6E908AFB29E0D8D0B967AF73B41C2~WombleNode01 LAUNCHED ID=61580092
650 ORCONN $5E2975B380F6E908AFB29E0D8D0B967AF73B41C2~WombleNode01 FAILED REASON=CONNECTREFUSED NCIRCS=4 ID=61580092
650 ORCONN $5E2975B380F6E908AFB29E0D8D0B967AF73B41C2~WombleNode01 LAUNCHED ID=61580095
650 ORCONN $5E2975B380F6E908AFB29E0D8D0B967AF73B41C2~WombleNode01 FAILED REASON=CONNECTREFUSED NCIRCS=3 ID=61580095
650 ORCONN $5E2975B380F6E908AFB29E0D8D0B967AF73B41C2~WombleNode01 LAUNCHED ID=61580098
[...]

It seems to me that if just a moment ago you tried to connect to a relay and you failed for reason connectrefused, then maybe you already know what the answer is going to be for other circuits that show up in that second.

How long should we cache the answer for? And, are there any anonymity implications to this design change? (We already batch circuit extend requests as they're waiting for the connect request to finish or fail, so I think whatever the anonymity implications are, maybe we already have them?)