Plugins are queried for reports two times in a second. That's often
enough to get the shortcut reports. The reports are sent together with
the response.
Thanks to that, plugins can react to requests from controls they
exposed.
To make it work, plugins registry modifies each plugin's report by
prepending the plugin ID to the control name the plugin has exposed
before sending it to the app. Then the registry installs the control
request handler for this faked control name, which forwards the
request to the correct plugin.
This adds a new API endpoint to plugins next to "/report" - a
"/control" entry. The body of the request is the JSON-encoded
xfer.Request instance.
It is not a singleton anymore. Instead it is an object with a registry
backend. The default registry backend is provided, which is equivalent
to what used to be before. Custom backend can be provided for testing
purposes.
The registry also supports batch operations to remove and add handlers
as an atomic step.
We will want to put plugin id in a control id, which is sent to an app
and then to GUI. When we get a control request from GUI, we will want
to extract the plugin ID from the control name. To do it unambiguously
we need some separator made of chars that are not allowed in a plugin
name. This is to avoid the situation when there are two plugins:
"Plugin" and "PluginFoo". "Plugin" exposes a control named
"FooControl" and "PluginFoo" exposes a control named "Control". Faking
the control names which will be sent to the app would result in two
"PluginFooControl".
One possible option for plugin ID and control name separator would be
"/", but that won't work, since the request sent from GUI to the app
to <probe>/<node>/<control> would actually be
<probe>/<node>/<plugin>/<control> and as such wouldn't match the URL
template in RegisterControlRoutes().
* Helps reduce garbage (MakeMetric() now takes a slice and there's a shorter version MakeSingletonMetric())
* Fixes bug computing Max (Min) in samples since using MakeMetric()
was causing a default Max/Min of zero.
* Simplifies code a bit
Also:
* Remove Gob encoder/decoder
* Stop using custom encoders/decoders for Timestamps (both ugorji and the Golang JSON codecs use nanosecond precision).
* Use idiomatic way to check for existence in metric.LastSample()
* Rework Scope metrics according to Prometheus conventions.
- counters should end with _total
- elaborated and added units to help strings
- recommended for cache hit/miss metrics: track only the total and the
hits and in separate metrics, since the most common query will be
"hits / total"
- track all times in seconds (base units), which has become the standard
recommendation
- other small changes
There could be more changes that would require more thinking (what
dimensions to use, summaries vs. histograms, etc.), but this is probably
enough controversial material already :)
* Use timeRequestStatus() in sqs_control_router.go.
