Although the API of the Karaf Config MBean looks like it offers synchronous calls, in reality
this is not true. I ran into some issues with this, because I tried to install configurations
for bundles to be picked up (see attached maven test case project to be used with a default running
Karaf 2.2.5 container)
It seems that the config bundle may need quite a long time to finish applying the configuration
changes, but returns BEFORE this is finished in background.
I have searched around, dug through code, but found no way to determine at which time I can
safely assume that a configuration change is finished and I can continue with installing the bundle
that may pick up this configuration (in some cases the default configuration is enough, but more
often I want to override that...). So I finally ended up with polling for the configuration changes,
but even that sometimes fails and I could loop infinitely and get no result (got an empty Map from proplist method).
From the JavaDoc of the MBean I would - as said - expect that I do not have to do that, but could
assume that after a create or propset call returns, the configuration is active. I would have been
a bit happier, if the config MBean offered JMX events that notify about the completion of a
configuration change, but after digging deeper it seems that even that would not be an optimal solution.
From my understanding the underlying implementation is a mix of Karaf maintaining its own config
files inside the etc folder plus updating data inside the OSGi ConfigAdmin Service.
The good thing about the console commands is that they implement a strategy that really
matches the idea of the OSGi ConfigAdmin service (the service always publishes complete configurations
as Dictionary instances). So on the console the config commands maintain state between config:edit
and config:update. Sending off config:update will publish all collected changes as one configuration
The API on the JMX level was more or less directly derived from the way the config console commands
work, but tries to avoid maintaining state - so there's no equivalent to config:edit / config:update.
Instead, propset, propdel and propappend immediately trigger propagation of simple changes to the
configuration. So on JMX level the granularity is not sending of a new Dictionary instance, but
reflects the operations to create such an instance. Instead of publishing one change a configuration
containing 20 properties would trigger 20 changes.
This will also cause multiple notifications for any "interested" configuration change listener ...
resulting not only in more load, but also possibly in inconsistent configuration sub stages to cope
with: Imagine a service requires two configuration parameters to work correctly. Instead of getting
one configuration notification that contains the two parameters, it will get two notifications and thus will have to handle the case that the second parameter is missing (desperately hoping that sooner or later this second parameter will appear ...).
Phew ' ... so I would suggest that this API should be changed:
- guarantee that on return from an MBean method that changes a configuration, this configuration
can be assumed to be active (so after creating a query call will return the equivalent created config).
- if this appears to be impossible, the bean should at least provide JMX events a client can listen
to in order to be notified when a configuration call is finished.
- in order to play nicely inside the OSGi framework, it should be possible to create configurations
by calling void create(String pid, Map<String, String> config)
- propset / propappend / propdel should be deprecated or (better) directly removed. (They make sense
on console, but not for JMX).