@jsha there are also reports of continied 500s in another thread, which might be related.
From @mnordhoff in production:
and with staging from @larryboymi:
EDIT: I missed that the second reproducer hadn't tried again since last weekend so this one may not be useful (unless it's a totally different issue).