I've attached fixes for bug #2870, that xfer can cause sync_client to
die:
https://bugzilla.andrew.cmu.edu/show_bug.cgi?id=2870
I also discovered that rename has a similar problem, and included a
fix for that as well. As I was analyzing this problem, I came to the
conclusion that the "allow a certain number of errors" code in
sync_client is serious flawed. Under some circumstances, it will
allow the replica to become out of sync with the primary backend.
While it will log these fatal errors, they are mixed in with very
similar non-fatal errors. And more importantly, while sync_client
usually dies when it encounters a fatal error, under some
circumstances it will merely log an error and keep running.
:wes
