IMHO, we need a fix for this. The race condition does happen in the real-world with frequently-fetched ResourceLoader objects. Our only recourse after the race condition is to do a one-off ban in varnish of some complex regex-matched subset of RL URLs. This bit us today in T133765 , where 30-day-long cache objects had race-losing outputs.

The situation we're in today without this fix (or broader design fixes would be good, too, but at least this fix) is basically this:

1. We hope we don't hit the race condition on random deploys
2. If we do, we hope we notice (in this case, a ticket popped up something like 4 days later complaining about JS errors), but don't have any way to really verify it.
3. When we notice it, we have to do a manual varnish ban, which doesn't scale and shouldn't be a routine operation.