It's hard to track down, but we appear to occasionally report back 503
errors that don't get logged.  Any thoughts?

Has anyone experimented with adding a Retry-After: header to the 503
response to handle short term unavailability due to load?