To get this clear. The race, which this PR fixes, is this:

1. The Node A writes parameter 'param' to 1 calling NodeHandle::setParam: 'param' is set in the nodes cache and is transmitted to the parameter server.
2. A reads 'param' with NodeHandle::getParamCached and gets the value 1 out of the cache.
3. A writes 'param' again, but this time the value 2. The value is written to the cache and transmitted to the server.
4. The parameter server notifys all nodes, which have been using 'param' , about the parameter change from step 1. In A's cache the value of 'param' is set to 1.
5. A read 'param' from the cache and gets value 1 (but expected 2).

The example chosen in fff8901 might be contrived, but it is a simple way to reproduce this issue.

Ok, as far as i can tell, the test for the rosmaster fails now, because the tests expects compute_param_updates to have a pretty wide contract. The documentation of Registrations says that its map-members is a map from keys to lists of pairs of the form (caller_id, caller_api) (L162).

But test_rosmaster_paramserver.py tests compute_param_updates passing a map from keys to caller_ids.

I find this rather confusing. Also I do not see any purpose of this rather than making the test easier.

Maybe there should be new tests for compute_param_updates but I'd like to get some feedback about this patch in advance.

@ros-pull-request-builder retest this please

Well, two changes, which are pretty unrelated to the original purpose of this PR, fixed the builds now.
@dirk-thomas I'm willing to create separate PRs for this unrelated changes if it makes it easier for you.

@dirk-thomas As far as I am concerned this can be merged now.

Thanks for the patch.