rpc.statd is single-threaded and runs its HA callout synchronously. If
it is too slow then latency accumulates and rpc.statd's backlog grows.
Running a pair of add-client/del-client events with the current code
averages ~0.030s in my test environment. This mean that 1000 clients
reclaiming locks after failover can easily cause 10s of latency. This
could cause rpc.statd to become unresponsive, resulting in a time out
for an rpcinfo-based health check of the status service.
Split the add-client/del-client events out to a standalone
statd_callout executable, written in C, to be used as the HA callout
for rpc.statd. All other functions move to statd_callout_helper.
Now, running a pair of add-client/del-client events in my test
environment averages only ~0.002s. This seems less likely to cause
latency problems.
The standalone statd_callout executable needs to read a configuration
file, which is generated by statd_callout_helper from the "startup"
event. It also needs access to a list of currently assigned public
IPs.
For backward compatibility, during installation a symlink is created
from $CTDB_BASE/statd-callout to the new statd_callout, which is
installed in the helper directory.
Testing this as part of the eventscript unit tests starts to become
even more of a hack than it used to be. However, the dependency on
stubs and the corresponding setup of fake state makes it hard to move
this elsewhere.
Signed-off-by: Martin Schwenke <mschwenke@ddn.com>
Reviewed-by: Amitay Isaacs <amitay@gmail.com>
Autobuild-User(master): Martin Schwenke <martins@samba.org>
Autobuild-Date(master): Tue Jun 25 04:24:57 UTC 2024 on atb-devel-224