[Date Prev][Date Next]
(ITS#4703) mirrormode hole
Full_Name: Howard Chu
Submission from: (NULL) (18.104.22.168)
Submitted by: hyc
If the current master commits an update but crashes before the change is
propagated to the mirror, and then new updates are subsequently committed on the
new master, the non-propagated changes will never be picked up, because the new
master's context is newer than everything on the crashed master.
The hole can be plugged by keeping a per-consumer contextCSN in addition to the
overall contextCSN, so that the old state isn't lost.
While we're at it, we should add a consumer option for specifying additional
locations in which to store contextCSNs. This would e.g. solve ITS#4626 by
allowing a consumer in a glued subordinate to keep its parent context up to