[asterisk-dev] ARI StasisEnd event vs. channel variables

[Mark Michelson]

On 10/19/2016 06:45 AM, Matthew Jordan wrote:
> On Tue, Oct 18, 2016 at 4:02 PM, Joshua Colp <jcolp at digium.com> wrote:
>> Matthew Jordan wrote:
>>
>> <snip>
>>
>>
>>> There are a few wrinkles with emitting channel variables with
>>> arbitrary events (of which StasisEnd would qualify).
>>>
>>> When an event is emitted out of the APIs, it's typically being done so
>>> as a result of a message having been published on the Stasis message
>>> bus. The messages on the bus don't contain a reference to a channel,
>>> but rather a snapshot of the channel. Those snapshots, by default,
>>> don't contain any channel variables, primarily for performance
>>> reasons. Additionally, since dialplan functions can be freely
>>> interchanged with channel variables, it's impossible (or very costly)
>>> to evaluate dialplan function expressions on each snapshot creation.
>>>
>>> Unfortunately, you really do need to rely on the messages coming
>>> across the bus. By the time code is processing the sending of a
>>> StasisEnd event, the ast_channel itself may already be gone. As such,
>>> the variables have to be stored on the snapshot in some fashion.
>>>
>>> AMI accomplished this by creating a core concept known as 'manager
>>> variables'. These are set directly on the ast_channel structure, and -
>>> when a snapshot is created - are evaluated and copied onto the
>>> snapshot. Again, this is a rather costly operation, but it does limit
>>> the impact of including variables on messages, as you have to
>>> explicitly tell Asterisk which state you want transmitted along with
>>> the channel.
>>>
>>> If we wanted a similar concept in ARI, the manager variables could be
>>> reused. I would note that you absolutely will take a performance hit
>>> when using them - building and conveying that much state in every
>>> event is not cheap. Because StasisEnd is rather 'special' - often the
>>> channel core isn't aware of it - I'm not sure how you would convey
>>> variables only on that event and not on all the rest.
>>
>> The actual raising of StasisEnd is explicitly done inside of Stasis and
>> lives there. When raised the channel is there and valid. It currently uses a
>> channel blob as the payload but this could be changed to something else that
>> would allow channel variables (only channel variables explicitly set on the
>> channel) to also be added. This would scope the channel variables addition
>> to only StasisEnd. Now... if we wanted channel variables on every message...
>> then yeah, it would be difficult.
>>
> Two thoughts:
> (1) I wonder if anyone needs the state conveyed on events other than
> StasisEnd? If Asterisk is maintaining all of your state for you, and
> you need to know some part of the state in your external application,
> how do you ask for it?
>
> (2) Even if we emit variable state on StasisEnd, how would we know
> which state to emit? Would we want to require users to provide it up
> front in ari.conf, or through some more dynamic mechanism?

I actually have a patch I've written locally in my spare time that adds 
an ari.conf option called "channelvars" that works exactly the same as 
for manager.conf. In other words, it modifies channel snapshots to have 
the channel variables specified in ari.conf. Then, every channel 
snapshot that is part of an event sent on the ARI websocket will have 
those variables.

I came up with this idea because in addition to Sebastien's comment 
about wanting variables on StasisEnd, other people told me they want 
channel variables on other events. Their reasoning being that they don't 
want the delay of having to look up channel variables every time channel 
state changes. By emitting a subset of channel variables (those that you 
care about in ARI), the message size can be limited to whatever you want.

I had thought about making the list of channel variables per-ARI app 
instead of global, but I don't know if that's actually desired.