We're agreed on creating a dedicated redirect server for this process, I'm not sure I agree on the need to ramp up processes on this, when the tradeoff of holding the transaction in shared memory while waiting for an http response exists. To be sure, you're
trading increased shared memory vs increased package memory, but there's frequently other considerations as well. In situations like this, I find that I'm doing a database dip, etc. in addition to the http request . Adding more workers adds more DB connections,
unnecessarily increasing memory consumption.
I'd go so far as to say that the older http_client generally shouldn't be used. If you have very infrequent http requests to execute and want to use it because it's syntactically easier then it might make sense, but beyond that it only offers lesser performance.
CAUTION: This email originated from outside the organization. Do not click links or open attachments unless you recognize the sender and know the content is safe.
One additional thought -- sorry for the numerous messages:
This discussion has got a little too academic. When talking about concurrency, it's important to ask what the workload actually is and how parallelisable it is.
1) All these async tricks are most useful when you have a busy Kamailio server that does lots of different kinds of things. Freeing up worker processes becomes an important objective, because not blocking them with these requests allows them to handle those
requests and those other replies.
2) The other important consideration is locking: how much are shared-memory structures (transaction state, dialog state, htables, etc) being accessed by multiple worker processes?
These issues are typical of Kamailio-based multiservice SIP gateways that perform functions like call routing, registration, presence, etc. all in one system.
Purpose-built redirect servers that query an HTTP API have a trivial and one-track workload. They receive an INVITE, it's allocated to a worker, the worker goes and bangs on an HTTP server, gets a response, and sends it back to the callee. There are no other
requests to juggle, save for the incidental OPTIONS ping maybe, there are no SIP replies to process and relay, and nothing else to do except to absorb ACKs (if the redirects are sent statefully). Along the way, vanishingly little is being accessed apart from
transaction memory (if stateful), and concurrent accesses to that are fast because the locks are pretty granular.
Thus, this type of workload doesn't benefit from general-purpose throughput optimisations much. The proper approach, using whichever HTTP client module, is to spin up as many child processes as can be spun up without diminishing returns, and just dedicating
them to that singular function while studiously avoiding lockable shared data (e.g. htable). This keeps all the cars mostly in their own lanes, and the success or failure of that per se will dictate throughput.
-- Alex
> On Dec 23, 2024, at 11:25 am, Alex Balashov <abalashov@evaristesys.com> wrote:
>
> PS. A key selling point of this is that if the TCP client crashes or becomes unavailable, the transaction just times out after a while, instead of any Kamailio worker processes shitting themselves or blocking because they're unable to connect to something.
There's nothing to connect to.
>
> This alone sold me. Once you go EVAPI, you can't go back.
>
> -- Alex
>
>> On Dec 23, 2024, at 11:23 am, Alex Balashov <abalashov@evaristesys.com> wrote:
>>
>>
>>> On Dec 23, 2024, at 10:56 am, Ben Kaufman via sr-users <sr-users@lists.kamailio.org> wrote:
>>>
>>> This came off a bit more harsh than I intended. I think I understand the advantage of EVAPI in that it can initiate a command to Kamailio (thus Kamailio doesn't need to hold the thread, possibly it doesn't need to store the message even), but it would
still be nice to have a higher level overview of how this would be achieved. Not even example code, but more a base architecture explanation / diagram of the flow.
>>
>> It's probably not complicated enough to warrant a diagram.
>>
>> 1) EVAPI server is initialised:
>>
>> loadmodule "evapi"
>> modparam("evapi", "workers", 3)
>> modparam("evapi", "bind_addr", "xxx:10399")
>>
>> 2) TCP client connects to this socket when it becomes available;
>>
>> 2) Kamailio receives request and serialises it (e.g. JSON), embedding the transaction ID and label in whatever serialisation structure is used, then emits it on the EVAPI bus, e.g.
>>
>> evapi_async_relay("invite_request:$T(id_index):$T(id_label):$(var(data){s.encode.base64})");
>>
>> 3) TCP client reads this package off the wire; the built-in netstring support is ideal. It then processes it, generates a response, serialises it, and puts it back on the wire. Vitally, the $T(id_index) and $T(id_label) are embedded in the response, allowing
Kamailio to resume the transaction:
>>
>> 4) Kamailio receives this message in $evapi(msg), in this event_route, and resumes the transaction into route[RESUME]:
>>
>> event_route[evapi:message-received] {
>> ...
>> jansson_get("tm_trans.tm_index", "$evapi(msg)", "$var(t_index)");
>> jansson_get("tm_trans.tm_label", "$evapi(msg)", "$var(t_label)");
>>
>> t_continue("$var(t_index)", "$var(t_label)", "RESUME");
>>
>> ...
>> }
>>
>> 5) In route[RESUME], INVITE processing continues as normal, enlightened by anything else that was deserialised out of $evapi(msg).
>>
>> EVAPI is my favourite Kamailio module, and the only way I interface with outside services, when it's up to me.
>>
>> -- Alex
>>
>> --
>> Alex Balashov
>> Principal Consultant
>> Evariste Systems LLC
