I share your concern that the placement of the feed is wrong. I feel
like we would be bending Brooklyn's abstractions to fit the case rather
than working out a better abstraction. Of course, deciding what data
reflects a software process is subjective and as you point out we
already break the encapsulation in a variety of places. Was the plan not
always to make locations into entities too? This would be trivial if
that were the case.
I also think we're continually making SoftwareProcess too important.
Svet's suggestion of an enricher is a good one - I want to explicitly
mix the capability in to entities, not enable it with a flag.
Sam
On 16/06/2016 09:13, Andrew Kennedy wrote:
Hi.
For the project I am working on, we would like to use the CPU utilization
as one of the metrics for scaling a cluster. The existing `MachineEntity`
has a sensor feed that produces this data, along with uptime and memory
usage information. The feed works on Linux VMs only, currently, as is uses
SSH commands on the host to generate the values i.e. the `uptime` command,
or the contents of files in `/proc/`.
I would like to propose moving the feed to `SoftwareProcess` so that it is
available to all entities. It would be disabled normally, set by a
`ConfigKey<Boolean>` flag. This would be named "metrics.machine.retrieve"
to correspond to "metrics.usage.retrieve" which enables sensors in feeds
that return application or process specific information. The
`MachineEntity` would obviously have the default value set to "true", to
maintain current behaviour.
The only issue with this change is that the placement of the sensor feed
feels slightly wrong. These are returning data about the _machine_ but the
entity represents a _process_ on that machine, and there may in fact be
multiple entities sharing a single machine, via `SameServerEntity`. The
`MachineEntity` is used to represent a VM without any applications running
on it, and would not normally be part of a blueprint, so these sensors are
not normally accessible. There is some precedent for placing machine data
on an entity, such as the `HOSTNAME` sensor, so I think the break in
encapsulation is quite small.
The PR containing the change is here:
- https://github.com/apache/brooklyn-server/pull/204
I'd appreciate any comments on whether this is a useful change, as well as
a review of the pull request...
Thanks,
Andrew.
