[
https://issues.apache.org/jira/browse/THRIFT-3593?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15365420#comment-15365420
]
James E. King, III commented on THRIFT-3593:
--------------------------------------------
The C# implementation in THRIFT-66 used events as you describe to set up
stateful session capabilities.
On the server side if something wants to send out as a client request, it is
presumed that happens from a thread other than a server side thrift thread
processing a connected client. One would not want a connected client method
handler to initiate communication back to the origin - an infinite loop could
occur. A typical sequence would be:
Let's define "A" as the system that initiates the connection. Historically
this is always a client.
Let's define "B" as the system that's listening for connections. Historically
this is always a server.
Remember that every request has a unique ID that is connection specific,
therefore the TConnectedClient acts as a message router on either side of the
connection. The server doesn't need to know or care about separating requests
from different clients.
1. "A" creates a TConnectedEndpoint.
2. "A" registers a service handler for the "PushEventNotification" Service.
3. "A" opens a connection to "B".
4. "B" creates a TConnectedEndpoint for the incoming connection, using the
TProcessor defined by the TServer.
5. "B" calls a ClientConnected callback giving the solution a chance to
establish a stateful session.
6. "B" tells TConnectedEndpoint to serve(), essentially. I don't remember if
that's how the older implementation worked, but it's possible.
7. "A" calls a thrift method in a service. The TConnectedEndpoint sends this
to "B" under mutex to prevent interleaving of message packets, and then blocks
the caller thread on "A" until a reply comes back (unless it is oneway). If
TConnectedEndpoint is thread-safe, as it is in the THRIFT-66 implementation,
then multiple threads can repeat #7. The connection itself issues a unique ID
per outgoing request to distinguish the replies.
8. "B" reads a message fully, then calls the TProcessor to execute it, using a
memory based transport to buffer the output.
9. This method modifies a database, perhaps, or tickles a semaphore.
10. An independent thread on "B" wakes up due to the database update.
11. Meanwhile, with the complete reply to the method call in hand, the memory
buffer reply in "B" is handed to the TConnectedEndpoint for delivery to "A",
unblocking #7 above. The client code that called the thrift method returns
with the response.
12. Meanwhile, the independent thread on "B" uses a PushEventNotification
client to send a oneway message to "A".
13. The TConnectedClient on "B" representing this connection ensures only one
of #11 or #12 is going outbound at a time to prevent interleaving on the socket.
It's been 5 years or more now so I would have to go over the code again.
At my current company we're also looking for push notification to optimize some
polling cycles.
If we ended up implementing it we would need to do so on C++, Java, C#, Perl,
and Python....
> Add new IDL keyword 'signal' to be able that a server can actively send
> messages to all its connected clients to prevent polling
> --------------------------------------------------------------------------------------------------------------------------------
>
> Key: THRIFT-3593
> URL: https://issues.apache.org/jira/browse/THRIFT-3593
> Project: Thrift
> Issue Type: Wish
> Components: AS3 - Compiler, AS3 - Library, C glib - Compiler, C glib
> - Library, C# - Compiler, C# - Library, C++ - Compiler, C++ - Library, Cocoa
> - Compiler, Cocoa - Library, Compiler (General), D - Compiler, D - Library,
> Dart - Compiler, Dart - Library, Delphi - Compiler, Delphi - Library,
> Documentation, Erlang - Compiler, Erlang - Library, Go - Compiler, Go -
> Library, Haskell - Compiler, Haskell - Library, Haxe - Compiler, Haxe -
> Library, Java - Compiler, Java - Library, JavaScript - Compiler, JavaScript -
> Library, Lua - Compiler, Lua - Library, Node.js - Compiler, Node.js -
> Library, Perl - Compiler, Perl - Library, PHP - Compiler, PHP - Library,
> Python - Compiler, Python - Library, Ruby - Compiler, Ruby - Library,
> Smalltalk - Compiler, Smalltalk - Library, Swift - Compiler, Test Suite,
> Tutorial
> Reporter: Sebastian Zenker
> Labels: push
>
> In our applications, we have very often the use case, that we actively want
> to inform all connected Thrift clients about state changes on the server
> side. Let me use a stupid example to explain what I whish. Let's assume we
> have service which represents a fan controller. This service allows to
> configure a target temperature and can be requested for the actual
> temperature and actual RPM.
> {code}
> service FanController
> {
> void setTargetTemperature(int t);
> int getTargetTemperature();
> int getActualTemperature();
> int getActualRPM();
> }
> {code}
> Our client application allows the user to set the target temperature and
> display the actual temperature and RPM.
> To implement such an application, we currently have two options when using
> the Thrift framework:
> 1.) Every client requests the actual temperature and RPM once per second.
> With other words: every client implements polling.
> 2.) We split service FanController into two different Thrift services. One
> which allows to configure the fan controller and a second one which is used
> by the server to notify all its clients about state changes. The first one is
> implemented by the "real" server and the second one is implemented by all
> clients and consists of some oneway methods only. So from a Thrift point of
> view, both sides are server & client. E.g.
> {code}
> service FanController
> {
> void setTargetTemperature(int t);
> int getTargetTemperature();
> void RegisterEvents(string hostname, int port); //use to tell the server,
> that it should establish a connection to hostname+port which implements
> FanControllerEvents
> void UnregisterEvents(string hostname, int port);
> }
> service FanControllerEvents
> {
> oneway void targetTemperatureChanged(int t);
> oneway void actualTemperatureChanged(int t);
> oneway void actualRPMChanged(int rpm);
> }
> {code}
> Both approaches have massive drawbacks. I think it is not worth the effort to
> explain why solution #1 (polling) sucks. But also solution #2 doesn't work
> well, because:
> * It requires every client to register its FanControllerEvents service at
> the server side by using FanController::RegisterEvents(). This doesn't work,
> in case the client resides behind a NAT-router, because so the "real" server
> cannot establish a TCP connection to the client.
> * It always requires at least two TCP connections which makes firewall
> configurations more complex.
> * The "real" server needs to maintain a list with all connected clients in
> the application logic. In case the actual RPM or temperature changes, the
> server needs to iterate over the list of all connected clients and call the
> corresponding function. Maintaining the list in the application logic adds
> extra complexity at the server side, which can be avoided and may be better
> part of the Thrift framework.
> * How to handle the case, if only 1 of the 2 TCP connections gets
> interrupted?
> * The fan controller service - which is logically one thing - gets splitted
> into two Thrift services: FanController + FanControllerEvents which decreases
> readability of the IDL file.
> To solve such a use case, my recommendation is the following: Add a new
> keyword like "signal" to the IDL language. Wouldn't it be cool to be able to
> define something like:
> {code}
> service FanController
> {
> void setTargetTemperature(int t);
> signal void targetTemperatureChanged(int t);
> signal void actualTemperatureChanged(int t);
> signal actualRPMChanged(int t);
> }
> {code}
> E.g. DBus (a IPC framework very often used in Linux environments) allows to
> specify signals in their interfaces. See also:
> http://dbus.freedesktop.org/doc/dbus-tutorial.html#signalprocedure
> It's a very intrusive wish, as it will effect all code generators and runtime
> libraries. What do you think?
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
