This still needs to be analyzed in more depth, but from a quick look at the PG connector it seems that we might not correctly handle the case where the connector is restarted and the WAL with the offset to resume from has been purged in between (snapshotting mode "initial"). We should make sure to detect this case properly and stop the connector with an error message, hinting that a new snapshot must be made.

As a follow-up, we should add a snapshotting mode "when_needed", akin to the one for MySQL, which automatically does a new snapshot in this situation.

This came up in a question on Gitter, at January 30th, 2018, 15:57 (UTC+1).