Date: | Tue, 25 Sep 2012 16:18:14 -0500 |
Subject: | kadmind per-slave ipropd dumps are wasteful |
From: | Nico Williams <nico@cryptonector.com> |
To: | krb5-bugs@mit.edu |
When iprop falls back on full resyncs kadmind uses a per-slave
dump. This is woefully inefficient and puts unnecessary strain
on the master. Moreover, since the dump will have iprop info in
the header we can tell if an existing dump is good enough to
send to any slave: if the sno/timestamp from the header is in
the ulog then we should send the dump without dumping at all,
thereby significantly cutting down the time needed for a full
resync from dump + xfer + load to xfer + load + iprop.
dump. This is woefully inefficient and puts unnecessary strain
on the master. Moreover, since the dump will have iprop info in
the header we can tell if an existing dump is good enough to
send to any slave: if the sno/timestamp from the header is in
the ulog then we should send the dump without dumping at all,
thereby significantly cutting down the time needed for a full
resync from dump + xfer + load to xfer + load + iprop.