Dive into SObjectizer-5.5. Sixth part: Synchronous Interaction

Dive into SObjectizer-5.5
SObjectizer Team, Jan 2016
Sixth Part: Synchronous Interaction
(at v.5.5.15)

This is the next part of the series of presentations with deep
introduction into features of SObjectizer-5.5.
This part is dedicated to synchronous interactions between
agents.

Introduction

A classical way of interaction between agents is
asynchronous messages.
Asynchronous interaction is simple, flexible and scalable. It
also could prevent some errors like deadlocks (but could
also lead to other kind of errors).
However, sometimes asynchronous interaction makes code
more complex and requires a lot of programmers work...

For example: sometimes an agent C must send a request to
an agent S and receive back a response with some
important information. The agent C can continue its work
only after arrival of information from agent S.
How we can express that scenario via asynchronous
messages?

We need a request message which must contain a mbox for
the response (since v.5.5.9 any struct/class which is
MoveConstructible can be used as a message):
struct get_config {
so_5::mbox_t reply_to;
};
We need a response which will be sent back:
struct config {
... // Some fields.
config(...); // Some constructor...
};

A skeleton for agent S is quite simple:
class config_manager : public so_5::agent_t {
...
public :
virtual void so_define_agent() override {
so_subscribe_self().event( &config_manager::evt_get_config );
... // Other subscriptions...
}
private :
void evt_get_config( const get_config & request ) {
// Sending a reply back.
so_5::send< config >( request.reply_to, ... /* some args */ );
}
};

But a skeleton of agent C is more complex:
class client : public so_5::agent_t {
// A special state to wait a response.
const so_5::state_t st_wait_config{ this, "wait_config" };
public :
// Agent should start its work in a special state.
this >>= st_wait_config;
// Reply with config data will be handled in that state.
st_wait_config.event( &client::evt_config );
...
}
virtual void so_evt_start() override {
// Request a config at start.
so_5::send< get_config >( config_manager_mbox, so_direct_mbox() );
}
private :
void evt_config( const config & cfg ) { ... /* Handling of configuration. */ }
};

It is a lot of code for such simple operation.
But this code is not robust enough.
For example, there will be no handling of the cases when
get_config request or config response are lost somewhere...

Synchronous interaction can help here.
An agent C can issue a service request to agent S and
receive the result of that request synchronously.

Let’s rewrite our small example with service request instead
of asynchronous messages...

A slight modification for request type. There is no need to
pass reply_to mbox in a request now. So, get_config
becomes a signal:
struct get_config : public so_5::signal_t {};
Response will remain the same:
struct config {
... // Some fields.
config(...); // Some constructors...
};

Service request processor will return config as a result of
event handler:
class config_manager : public so_5::agent_t {
...
public :
so_subscribe_self().event< get_config >( &config_manager::evt_get_config );
... // Other subscriptions...
}
private :
config evt_get_config() {
// Returning a reply back.
return config{ ... /* some args */ };
}
};

Service request issuer will look like:
class client : public so_5::agent_t {
public :
... // No special state, no subscription for config response.
}
// Request a config at start.
// Wait response for 1 second.
auto cfg = so_5::request_value< config, get_config >(
config_manager_mbox, std::chrono::seconds(1) );
}
...
};

That code is not only much simpler ‒ it is also more robust:
● if there is no service request processor behind
config_manager_mbox there will be an exception;
● if there are more than one service request processor behind
config_manager_mbox there will be an exception;
● if config_manager won’t process a request (request is ignored in the
current state of manager) there will be an exception;
● if config_manager can’t process a request, e.g. throw an exception,
that exception will be propagated to service request's issuer;
● if config_manager can’t process a request in the specified time slice
there will be an exception.

It means that in several cases synchronous interaction via
service requests is more appropriate than asynchronous
interaction.
In such cases service requests allow to write more simple,
clean and robust code…
...but everything has its price.

The main disadvantage of service request is a possibility of
deadlocks.
Service request’s issuer and processor must work on
different working threads.
It means that issuer and processor must be bound to
different dispatchers. Or, to the different working threads
inside the same dispatcher.

If an issuer and a processor work on the same working
thread there will be a deadlock.
SObjectizer doesn’t check that. A user is responsible for
binding issuer and processor to the different contexts.

But the case when an issuer and a processor are working on
the same thread is the simplest case of deadlock.
There could be more complex cases:
● agent A calls agent B;
● agent B calls agent C;
● agent C calls agent D;
● agent D calls agent A.
It is another kind of classical deadlock with the same
consequences.

Another disadvantage of service request is a blocking of a
working thread for some time.
If a service request issuer shares the working thread with
different agents (for example, all of them are bound to
one_thread dispatcher instance) then all other agents on that
thread will wait until the service request is completed.
It means that synchronous agents interaction is not very
scalable.

As shown above, the synchronous agents interaction has
significant disadvantages. Based on that, it should be used
with care.

How Does It Work?

There is no any magic behind service requests.
Just an ordinary asynchronous messages and some help
from std::promise and std::future...

When a service request is initiated a special envelope with
service requests params inside is sent as an ordinary
message to service request processor.
This envelope contains not only request params but also a
std::promise object for the response. A value for that
promise is set on processor’s side.
A service request issuer waits on std::future object which is
bound with std::promise from the envelope which was sent.

It means that so_5::request_value call in form:
auto r = so_5::request_value<Result,Request>(mbox, timeout, params);
It is a shorthand for something like that:
// Envelope will contain Request object and std::promise<Result> object.
auto env__ = std::make_unique<msg_service_request_t<Result,Request>>(params);
// Get the future to wait on it.
std::future<Result> f__ = env__.m_promise.get_future();
// Sending the envelope with request params as async message.
mbox->deliver_message(std::move(env__));
// Waiting and handling the result.
auto wait_result__ = f__.wait_for(timeout);
if(std::future_status::ready != wait_result__)
throw exception_t(...);
auto r = f__.get();

Every service request handler in the following form
Result event_handler(const Request & svc_req ) { ... }
is automatically transformed to something like this:
void actual_event_handler(msg_service_request_t<Result,Request> & m) {
try {
m.m_promise.set( event_handler(m.query_param()) );
}
catch(...) {
m.set_exception(std::current_exception());
}
}
This transformation is performed during subscription of
event_handler.

This approach of supporting synchronous interaction means
that service request is handled by SObjectizer just like
dispatching and handling of ordinary message.
As a consequence, a service request handler even doesn’t
know is a message it handles is a part of a service request
or it was sent as a asynchronous message?
Only the service request issuer knows the difference.

It means that a config_manager from the example above will
work in the same way even if get_config signal is sent via
so_5::send() function like any other async message.
But in this case the return value of config_manager::
evt_get_config will be discarded.
It means that there is no special rules for writing service
request handlers: they are just agents with traditional event
handlers. Except one important aspect...

This important aspect is exception handling.
If an agent allows an exception to go out from an event
handler then traditional reaction to unhandled exception will
be initiated:
● SObjectizer Environment will call so_exception_reaction() for that
agent (and may be for coop of that agent and so on);
● appropriate action will be taken (for example, an exception could
be ignored or the whole application will be aborted).

But if an exception is going out from service request handler
then it will be intercepted and returned back to service
request issuer via std::promise/std::future pair.
It means that this exception will be reraised on service
request issuer side during the call to std::future::get()
method. E.g. that exception will be thrown out from
request_value().

request_value and request_future functions

There are several ways of initiating service requests.
The simplest one is the usage of so_5::request_value()
template function. It can be used in the form:
Result r = so_5::request_value<Result, Request>(Target, Timeout [,Args]);
Where Timeout is a value which is defined by std::chrono or
so_5::infinite_wait for non-limited waiting of the result.
Type of Request can be any message or signal type. All
Args (if any) will be passed to the constructor of Request.

Examples of request_value invocations:
// Sending a get_status signal as service request.
auto v = so_5::request_value<engine_status, get_status>(engine,
// Infinite waiting for the response.
so_5::infinite_wait); // No other params because of signal.
// Sending a message convert_value as a service request.
auto v = so_5::request_value<std::string, convert_value>(converter,
// Waiting for 200ms for the response.
std::chrono::milliseconds(200),
"feets", 33000, "metres" ); // Params for the constructor of convert_value.

There is also so_5::request_future() template function. It
returns std::future<Result> object:
std::future<Result> r = so_5::request_future<Result, Request>(Target [,Args]);
There is no Timeout argument. Waiting on the std::future is
responsibility of a programmer.
As in the case of request_value type of Request can be any
of message or signal type. All Args (if any) will be passed to
the constructor of Request.

request_future can be used in more complex scenarios than
request_value. For example, it is possible to issue several
service requests, do some actions and only then request
responses:
auto lights = so_5::request_future<light_status, turn_light_on>(light_controller);
auto heating = so_5::request_future<heating_status, turn_heating_on>(heating_controller);
auto accessories = so_5::request_future<accessories_status, update>(accessories_controller);
... // Some other actions.
if(light_status::on != lights.get())
... // Some reaction...
if(heating_status::on != heating.get())
... // Some reaction...
check_accessories(accessories.get());
...

There could be more complex scenarios:
class accessories_listener : public so_5::agent_t {
public :
...
m_status = so_5::request_future< accessories_status, get_status >(accessories_controller());
// Service request initiated, but the response will be used later.
so_5::send_delayed< check_status >(*this, std::chrono::milliseconds(250));
}
private :
std::future<accessories_status> m_status;
...
void evt_check_status() {
auto status = m_status.get(); // Getting the service request response.
... // Processing it.
// Initiate new request again.
m_status = so_5::request_future< accessories_status, get_status >(accessories_controller());
// Response is expected to be ready on the next timer event
so_5::send_delayed< check_status >(*this, std::chrono::milliseconds(250));
}
};

Service Requests With void As Result Type

Sometimes there is some sense in initiating a service
request which returns void.
A result of such service request means that processing of a
request is completely finished. Sometimes, is it a very useful
information.
For example, if service request processor does flushing of
some important data, finishing transactions, invalidating
caches and so on...

An example of service request with void result:
// Type of agent which implements an intermediate buffer with flushing by demand or by timer.
class data_buffer : public so_5::agent_t {
public :
// Signal for flushing the data.
struct flush : public so_5::signal_t {};
...
private :
// Event which is bound to flush signal.
void evt_flush() {
... // Storing the data to the disk/database/cloud...
}
};
// Initiating flush operation as a service request. Return from request_value
// means the completeness of data flushing.
so_5::request_value<void, data_buffer::flush>(buffer, so_5::infinite_wait);

Additional Information:
Project’s home: http://sourceforge.net/projects/sobjectizer
Documentation: http://sourceforge.net/p/sobjectizer/wiki/
Forum: http://sourceforge.net/p/sobjectizer/discussion/
Google-group: https://groups.google.com/forum/#!forum/sobjectizer
GitHub mirror: https://github.com/masterspline/SObjectizer

Dive into SObjectizer-5.5. Sixth part: Synchronous Interaction

Recommended

Recommended

More Related Content

What's hot

What's hot (20)

Similar to Dive into SObjectizer-5.5. Sixth part: Synchronous Interaction

Similar to Dive into SObjectizer-5.5. Sixth part: Synchronous Interaction (20)

More from Yauheni Akhotnikau

More from Yauheni Akhotnikau (14)

Recently uploaded

Recently uploaded (20)

Dive into SObjectizer-5.5. Sixth part: Synchronous Interaction