Clients: Calling remote objects

This chapter explains how you write code that calls remote objects. Often, a program that calls methods on a Pyro object is called a client program. (The program that provides the object and actually runs the methods, is the server. Both roles can be mixed in a single program.)

Make sure you are familiar with Pyro’s Key concepts before reading on.

Object discovery

To be able to call methods on a Pyro object, you have to tell Pyro where it can find the actual object. This is done by creating an appropriate URI, which contains amongst others the object name and the location where it can be found. You can create it in a number of ways.

  • directly use the object name and location.

    This is the easiest way and you write an URI directly like this: PYRO:someobjectid@servername:9999 It requires that you already know the object id, servername, and port number. You could choose to use fixed object names and fixed port numbers to connect Pyro daemons on. For instance, you could decide that your music server object is always called “musicserver”, and is accessible on port 9999 on your server musicbox.my.lan. You could then simply use:

    uri_string = "PYRO:musicserver@musicbox.my.lan:9999"
    # or use Pyro4.URI("...") for an URI object instead of a string
    

    Most examples that come with Pyro simply ask the user to type this in on the command line, based on what the server printed. This is not very useful for real programs, but it is a simple way to make it work. You could write the information to a file and read that from a file share (only slightly more useful, but it’s just an idea).

  • use a logical name and look it up in the name server.

    A more flexible way of locating your objects is using logical names for them and storing those in the Pyro name server. Remember that the name server is like a phone book, you look up a name and it gives you the exact location. To continue on the previous bullet, this means your clients would only have to know the logical name “musicserver”. They can then use the name server to obtain the proper URI:

    import Pyro4
    nameserver = Pyro4.locateNS()
    uri = nameserver.lookup("musicserver")
    # ... uri now contains the URI with actual location of the musicserver object
    

    You might wonder how Pyro finds the Name server. This is explained in the separate chapter Name Server.

  • use a logical name and let Pyro look it up in the name server for you.

    Very similar to the option above, but even more convenient, is using the meta-protocol identifier PYRONAME in your URI string. It lets Pyro know that it should lookup the name following it, in the name server. Pyro should then use the resulting URI from the name server to contact the actual object. See The PYRONAME protocol type. This means you can write:

    uri_string = "PYRONAME:musicserver"
    # or Pyro4.URI("PYRONAME:musicserver") for an URI object
    

    You can use this URI everywhere you would normally use a normal uri (using PYRO). Everytime Pyro encounters the PYRONAME uri it will use the name server automatically to look up the object for you. [1]

  • use object metadata tagging to look it up (yellow-pages style lookup).

    You can do this directly via the name server for maximum control, or use the PYROMETA protocol type. See The PYROMETA protocol type. This means you can write:

    uri_string = "PYROMETA:metatag1,metatag2"
    # or Pyro4.URI("PYROMETA:metatag1,metatag2") for an URI object
    

    You can use this URI everywhere you would normally use a normal uri. Everytime Pyro encounters the PYROMETA uri it will use the name server automatically to find a random object for you with the given metadata tags. [1]

[1](1, 2) this is not very efficient if it occurs often. Have a look at the Tips & Tricks chapter for some hints about this.

Calling methods

Once you have the location of the Pyro object you want to talk to, you create a Proxy for it. Normally you would perhaps create an instance of a class, and invoke methods on that object. But with Pyro, your remote method calls on Pyro objects go trough a proxy. The proxy can be treated as if it was the actual object, so you write normal python code to call the remote methods and deal with the return values, or even exceptions:

# Continuing our imaginary music server example.
# Assume that uri contains the uri for the music server object.

musicserver = Pyro4.Proxy(uri)
try:
    musicserver.load_playlist("90s rock")
    musicserver.play()
    print("Currently playing:", musicserver.current_song())
except MediaServerException:
    print("Couldn't select playlist or start playing")

For normal usage, there’s not a single line of Pyro specific code once you have a proxy!

Accessing remote attributes

You can access exposed attributes of your remote objects directly via the proxy. If you try to access an undefined or unexposed attribute, the proxy will raise an AttributeError stating the problem. Note that direct remote attribute access only works if the metadata feature is enabled (METADATA config item, enabled by default).

import Pyro4

p = Pyro4.Proxy("...")
velo = p.velocity    # attribute access, no method call
print("velocity = ", velo)

See the attributes example for more information.

Serialization

Pyro will serialize the objects that you pass to the remote methods, so they can be sent across a network connection. Depending on the serializer that is being used, there will be some limitations on what objects you can use.

  • serpent: serializes into Python literal expressions. Accepts quite a lot of different types. Many will be serialized as dicts. You might need to explicitly translate literals back to specific types on the receiving end if so desired, because most custom classes aren’t dealt with automatically. Requires third party library module, but it will be installed automatically as a dependency of Pyro. This serializer is the default choice.
  • json: more restricted as serpent, less types supported. Part of the standard library. Not particularly fast, so you might want to look for a faster 3rd party implementation (such as simplejson). Be sure to benchmark before switching! Use the JSON_MODULE config item to tell Pyro to use the other module instead. Note that it has to support the advanced parameters such as default, not all 3rd party implementations do that.
  • marshal: a very limited but fast serializer. Can deal with a small range of builtin types only, no custom classes can be serialized. Part of the standard library.
  • msgpack: See https://pypi.python.org/pypi/msgpack-python Reasonably fast serializer (and a lot faster if you’re using the C module extension). Can deal with many builtin types, but not all. Not enabled by default because it’s optional, but it’s safe to add to the accepted serializers config item if you have it installed.
  • pickle: the legacy serializer. Fast and supports almost all types. Part of the standard library. Has security problems, so it’s better to avoid using it.
  • cloudpickle: See https://pypi.python.org/pypi/cloudpickle It is similar to pickle serializer, but more capable. Extends python’s ‘pickle’ module for serializing and de-serializing python objects to the majority of the built-in python types. Has security problems though, just as pickle.
  • dill: See https://pypi.python.org/pypi/dill It is similar to pickle serializer, but more capable. Extends python’s ‘pickle’ module for serializing and de-serializing python objects to the majority of the built-in python types. Has security problems though, just as pickle.

You select the serializer to be used by setting the SERIALIZER config item. (See the Configuring Pyro chapter). The valid choices are the names of the serializer from the list mentioned above. If you’re using pickle or dill, and need to control the protocol version that is used, you can do so with the PICKLE_PROTOCOL_VERSION or DILL_PROTOCOL_VERSION config items. If you’re using cloudpickle, you can control the protocol version with PICKLE_PROTOCOL_VERSION as well. By default Pyro will use the highest one available.

It is possible to override the serializer on a particular proxy. This allows you to connect to one server using the default serpent serializer and use another proxy to connect to a different server using the json serializer, for instance. Set the desired serializer name in proxy._pyroSerializer to override.

Note

Since Pyro 4.20 the default serializer is “serpent”. Serpent is secure but cannot serialize all types (by design). Some types are serialized into a different form such as a string or a dict. Strings are serialized/deserialized into unicode at all times – be aware of this if you’re using Python 2.x (strings in Python 3.x are always unicode already).

Note

The serializer(s) that a Pyro server/daemon accepts, is controlled by a different config item (SERIALIZERS_ACCEPTED). This can be a set of one or more serializers. By default it accepts the set of ‘safe’ serializers, so “pickle”, “cloudpickle” and “dill” are excluded. If the server doesn’t accept the serializer that you configured for your client, it will refuse the requests and respond with an exception that tells you about the unsupported serializer choice. If it does accept your requests, the server response will use the same serializer that was used for the request.

Note

Because the name server is just a regular Pyro server as well, you will have to tell it to allow the pickle, cloudpickle or dill serializers if your client code uses them. See Using the name server with pickle, cloudpickle or dill serializers.

Changing the way your custom classes are (de)serialized

Note

The information in this paragraph is not relevant when using the pickle, cloudpickle or dill serialization protocols, they have their own ways of serializing custom classes.

By default, custom classes are serialized into a dict. They are not deserialized back into instances of your custom class. This avoids possible security issues. An exception to this however are certain classes in the Pyro4 package itself (such as the URI and Proxy classes). They are deserialized back into objects of that certain class, because they are critical for Pyro to function correctly.

There are a few hooks however that allow you to extend this default behaviour and register certain custom converter functions. These allow you to change the way your custom classes are treated, and allow you to actually get instances of your custom class back from the deserialization if you so desire.

The hooks are provided via several classmethods:
Pyro4.util.SerializerBase.register_class_to_dict() and Pyro4.util.SerializerBase.register_dict_to_class()
and their unregister-counterparts:
Pyro4.util.SerializerBase.unregister_class_to_dict() and Pyro4.util.SerializerBase.unregister_dict_to_class()

Click on the method link to see its apidoc, or have a look at the ser_custom example and the test_serialize unit tests for more information. It is recommended to avoid using these hooks if possible, there’s a security risk to create arbitrary objects from serialized data that is received from untrusted sources.

Upgrading older code that relies on pickle

What do you have to do with code that relies on pickle, and worked fine in older Pyro versions, but now crashes?

You have three options:

  1. Redesign remote interfaces
  2. Configure Pyro to eable the use of pickle again
  3. Stick to Pyro 4.18 (less preferable)

You can redesign the remote interface to only include types that can be serialized (python’s built-in types and exception classes, and a few Pyro specific classes such as URIs). That way you benefit from the new security that the alternative serializers provide. If you can’t do this, you have to tell Pyro to enable pickle again. This has been made an explicit step because of the security implications of using pickle. Here’s how to do this:

Client code configuration
Tell Pyro to use pickle as serializer for outgoing communication, by setting the SERIALIZER config item to pickle. For instance, in your code: Pyro4.config.SERIALIZER = 'pickle' or set the appropriate environment variable.
Server code configuration
Tell Pyro to accept pickle as incoming serialization format, by including pickle in the SERIALIZERS_ACCEPTED config item list. For instance, in your code: Pyro4.config.SERIALIZERS_ACCEPTED.add('pickle'). Or set the appropriate environment variable, for instance: export PYRO_SERIALIZERS_ACCEPTED=serpent,json,marshal,pickle. If your server also uses Pyro to call other servers, you may also need to configure it as mentioned above at ‘client code’. This is because the incoming and outgoing serializer formats are configured independently.

Proxies, connections, threads and cleaning up

Here are some rules:

  • Every single Proxy object will have its own socket connection to the daemon.

  • You can share Proxy objects among threads, it will re-use the same socket connection.

  • Usually every connection in the daemon has its own processing thread there, but for more details see the Servers: hosting Pyro objects chapter.

  • The connection will remain active for the lifetime of the proxy object. Hence, consider cleaning up a proxy object explicitly if you know you won’t be using it again in a while. That will free up resources and socket connections. You can do this in two ways:

    1. calling _pyroRelease() on the proxy.

    2. using the proxy as a context manager in a with statement. This is the preffered way of creating and using Pyro proxies. This ensures that when you’re done with it, or an error occurs (inside the with-block), the connection is released:

      with Pyro4.Proxy(".....") as obj:
          obj.method()
      

    Note: you can still use the proxy object when it is disconnected: Pyro will reconnect it as soon as it’s needed again.

  • At proxy creation, no actual connection is made. The proxy is only actually connected at first use, or when you manually connect it using the _pyroReconnect() or _pyroBind() methods.

Oneway calls

Normal method calls always block until the response is returned. This can be any normal return value, None, or an error in the form of a raised exception. The client code execution is suspended until the method call has finished and produced its result.

Some methods never return any response or you are simply not interested in it (including errors and exceptions!), or you don’t want to wait until the result is available but rather continue immediately. You can tell Pyro that calls to these methods should be done as one-way calls. For calls to such methods, Pyro will not wait for a response from the remote object. The return value of these calls is always None, which is returned immediately after submitting the method invocation to the server. The server will process the call while your client continues execution. The client can’t tell if the method call was successful, because no return value, no errors and no exceptions will be returned! If you want to find out later what - if anything - happened, you have to call another (non-oneway) method that does return a value.

Note that this is different from Asynchronous (‘future’) remote calls & call chains: they are also executed while your client code continues with its work, but they do return a value (but at a later moment in time). Oneway calls are more efficient because they immediately produce None as result and that’s it.

How to make methods one-way: You mark the methods of your class in the server as one-way by using a special decorator. See Creating a Pyro class and exposing its methods and properties for details on how to do this. See the oneway example for some code that demonstrates the use of oneway methods.

Batched calls

Doing many small remote method calls in sequence has a fair amount of latency and overhead. Pyro provides a means to gather all these small calls and submit it as a single ‘batched call’. When the server processed them all, you get back all results at once. Depending on the size of the arguments, the network speed, and the amount of calls, doing a batched call can be much faster than invoking every call by itself. Note that this feature is only available for calls on the same proxy object.

How it works:

  1. You create a batch proxy object for the proxy object.
  2. Call all the methods you would normally call on the regular proxy, but use the batch proxy object instead.
  3. Call the batch proxy object itself to obtain the generator with the results.

You create a batch proxy using this: batch = Pyro4.batch(proxy) or this (equivalent): batch = proxy._pyroBatch(). The signature of the batch proxy call is as follows:

batchproxy.__call__([oneway=False, async=False])

Invoke the batch and when done, returns a generator that produces the results of every call, in order. If oneway==True, perform the whole batch as one-way calls, and return None immediately. If async==True, perform the batch asynchronously, and return an asynchronous call result object immediately.

Simple example:

batch = Pyro4.batch(proxy)
batch.method1()
batch.method2()
# more calls ...
batch.methodN()
results = batch()   # execute the batch
for result in results:
    print(result)   # process result in order of calls...

Oneway batch:

results = batch(oneway=True)
# results==None

Asynchronous batch

The result value of an asynchronous batch call is a special object. See Asynchronous (‘future’) remote calls & call chains for more details about it. This is some simple code doing an asynchronous batch:

results = batch(async=True)
# do some stuff... until you're ready and require the results of the async batch:
for result in results.value:
    print(result)    # process the results

See the batchedcalls example for more details.

Remote iterators/generators

Since Pyro 4.49 it is possible to simply iterate over a remote iterator or generator function as if it was a perfectly normal Python iterable. Pyro will fetch the items one by one from the server that is running the remote iterator until all elements have been consumed or the client disconnects.

So you can write in your client:

proxy = Pyro4.Proxy("...")
for item in proxy.things():
    print(item)

The implementation of the things method can return a normal list but can also return an iterator or even be a generator function itself. This has the usual benefits of “lazy” generators: no need to create the full collection upfront which can take a lot of memory, possibility of infinite sequences, and spreading computation load more evenly.

By default the remote item streaming is enabled in the server and there is no time limit set for how long iterators and generators can be ‘alive’ in the server. You can configure this however if you want to restrict resource usage or disable this feature altogether, via the ITER_STREAMING and ITER_STREAM_LIFETIME config items.

Lingering when disconnected: the ITER_STREAM_LINGER config item controls the number of seconds a remote generator is kept alive when a disconnect happens. It defaults to 30 seconds. This allows you to reconnect the proxy and continue using the remote generator as if nothing happened (see Pyro4.core.Proxy._pyroReconnect() or even Automatic reconnecting). If you reconnect the proxy and continue iterating again after the lingering timeout period expired, an exception is thrown because the remote generator has been discarded in the meantime. Lingering can be disabled completely by setting the value to 0, then all remote generators from a proxy will immediately be discarded in the server if the proxy gets disconnected or closed.

Notice that you can also use this in your Java or .NET/C# programs that connect to Python via Pyrolite! Version 4.14 or newer of that library supports Pyro item streaming. It returns normal Java and .NET iterables to your code that you can loop over normally with foreach or other things.

There are several examples that use the remote iterator feature. Have a look at the stockquotes tutorial example, or the filetransfer example.

Asynchronous (‘future’) remote calls & call chains

You can execute a remote method call and tell Pyro: “hey, I don’t need the results right now. Go ahead and compute them, I’ll come back later once I need them”. The call will be processed in the background and you can collect the results at a later time. If the results are not yet available (because the call is still being processed) your code blocks but only at the line you are actually retrieving the results. If they have become available in the meantime, the code doesn’t block at all and can process the results immediately. It is possible to define one or more callables (the “call chain”) that should be invoked automatically by Pyro as soon as the result value becomes available.

You set a proxy in async mode using this: Pyro4.async(proxy) or (equivalent): proxy._pyroAsync(). Every remote method call you make on the async proxy, returns a Pyro4.futures.FutureResult object immediately. This object means ‘the result of this will be available at some moment in the future’ and has the following interface:

value

This property contains the result value from the call. If you read this and the value is not yet available, execution is halted until the value becomes available. If it is already available you can read it as usual.

ready

This property contains the readiness of the result value (True meaning that the value is available).

wait([timeout=None])

Waits for the result value to become available, with optional wait timeout (in seconds). Default is None, meaning infinite timeout. If the timeout expires before the result value is available, the call will return False. If the value has become available, it will return True.

then(callable[, *args, **kwargs])

Add a callable to the call chain, to be invoked when the results become available. The result of the current call will be used as the first argument for the next call. Optional extra arguments can be provided via args and kwargs.

iferror(errorhandler)

Specify the exception handler to be invoked (with the exception object as only argument) when asking for the result raises an exception. If no exception handler is set, any exception result will be silently ignored (unless you explicitly ask for the value). Returns self so you can easily chain other calls.

A simple piece of code showing an asynchronous method call:

proxy._pyroAsync()
asyncresult = proxy.remotemethod()
print("value available?", asyncresult.ready)
# ...do some other stuff...
print("resultvalue=", asyncresult.value)

Note

Batched calls can also be executed asynchronously. Asynchronous calls are implemented using a background thread that waits for the results. Callables from the call chain are invoked sequentially in this background thread.

See the async example for more details and example code for call chains.

Async calls for normal callables (not only for Pyro proxies)

The async proxy discussed above is only available when you are dealing with Pyro proxies. It provides a convenient syntax to call the methods on the proxy asynchronously. For normal Python code it is sometimes useful to have a similar mechanism as well. Pyro provides this too, see Asynchronous (‘future’) normal function calls for more information.

Pyro Callbacks

Usually there is a nice separation between a server and a client. But with some Pyro programs it is not that simple. It isn’t weird for a Pyro object in a server somewhere to invoke a method call on another Pyro object, that could even be running in the client program doing the initial call. In this case the client program is a server itself as well.

These kinds of ‘reverse’ calls are labeled callbacks. You have to do a bit of work to make them possible, because normally, a client program is not running the required code to also act as a Pyro server to accept incoming callback calls.

In fact, you have to start a Pyro daemon and register the callback Pyro objects in it, just as if you were writing a server program. Keep in mind though that you probably have to run the daemon’s request loop in its own background thread. Or make heavy use of oneway method calls. If you don’t, your client program won’t be able to process the callback requests because it is by itself still waiting for results from the server.

Exceptions in callback objects: If your callback object raises an exception, Pyro will return that to the server doing the callback. Depending on what the server does with it, you might never see the actual exception, let alone the stack trace. This is why Pyro provides a decorator that you can use on the methods in your callback object in the client program: @Pyro4.callback. This way, an exception in that method is not only returned to the caller, but also logged locally in your client program, so you can see it happen including the stack trace (if you have logging enabled):

import Pyro4

class Callback(object):

    @Pyro4.expose
    @Pyro4.callback
    def call(self):
        print("callback received from server!")
        return 1//0    # crash!

Also notice that the callback method (or the whole class) has to be decorated with @Pyro4.expose as well to allow it to be called remotely at all. See the callback example for more details and code.

Miscellaneous features

Pyro provides a few miscellaneous features when dealing with remote method calls. They are described in this section.

Error handling

You can just do exception handling as you would do when writing normal Python code. However, Pyro provides a few extra features when dealing with errors that occurred in remote objects. This subject is explained in detail its own chapter: Exceptions and remote tracebacks.

See the exceptions example for more details.

Timeouts

Because calls on Pyro objects go over the network, you might encounter network related problems that you don’t have when using normal objects. One possible problems is some sort of network hiccup that makes your call unresponsive because the data never arrived at the server or the response never arrived back to the caller.

By default, Pyro waits an indefinite amount of time for the call to return. You can choose to configure a timeout however. This can be done globally (for all Pyro network related operations) by setting the timeout config item:

Pyro4.config.COMMTIMEOUT = 1.5      # 1.5 seconds

You can also do this on a per-proxy basis by setting the timeout property on the proxy:

proxy._pyroTimeout = 1.5    # 1.5 seconds

There is also a server setting related to oneway calls, that says if oneway method calls should be executed in a separate thread or not. If this is set to False, they will execute in the same thread as the other method calls. This means that if the oneway call is taking a long time to complete, the other method calls from the client may actually stall, because they’re waiting on the server to complete the oneway call that came before them. To avoid this problem you can set this config item to True (which is the default). This runs the oneway call in its own thread (regardless of the server type that is used) and other calls can be processed immediately:

Pyro4.config.ONEWAY_THREADED = True     # this is the default

See the timeout example for more details.

Also, there is a automatic retry mechanism for timeout or connection closed (by server side), in order to use this automatically retry:

Pyro4.config.MAX_RETRIES = 3      # attempt to retry 3 times before raise the exception

You can also do this on a pre-proxy basis by setting the max retries property on the proxy:

proxy._pyroMaxRetries = 3      # attempt to retry 3 times before raise the exception

Be careful to use when remote functions have a side effect (e.g.: calling twice results in error)! See the autoretry example for more details.

Automatic reconnecting

If your client program becomes disconnected to the server (because the server crashed for instance), Pyro will raise a Pyro4.errors.ConnectionClosedError. You can use the automatic retry mechanism to handle this exception, see the autoretry example for more details. Alternatively, it is also possible to catch this and tell Pyro to attempt to reconnect to the server by calling _pyroReconnect() on the proxy (it takes an optional argument: the number of attempts to reconnect to the daemon. By default this is almost infinite). Once successful, you can resume operations on the proxy:

try:
    proxy.method()
except Pyro4.errors.ConnectionClosedError:
    # connection lost, try reconnecting
    obj._pyroReconnect()

This will only work if you take a few precautions in the server. Most importantly, if it crashed and comes up again, it needs to publish its Pyro objects with the exact same URI as before (object id, hostname, daemon port number).

See the autoreconnect example for more details and some suggestions on how to do this.

The _pyroReconnect() method can also be used to force a newly created proxy to connect immediately, rather than on first use.

Proxy sharing

Due to internal locking you can freely share proxies among threads. The lock makes sure that only a single thread is actually using the proxy’s communication channel at all times. This can be convenient but it may not be the best way to approach things. The lock essentially prevents parallelism. If you want calls to go in parallel, give each thread its own proxy.

Here are a couple of suggestions on how to make copies of a proxy:

  1. use the copy module, proxy2 = copy.copy(proxy)
  2. create a new proxy from the uri of the old one: proxy2 = Pyro4.Proxy(proxy._pyroUri)
  3. simply create a proxy in the thread itself (pass the uri to the thread instead of a proxy)

See the proxysharing example for more details.

Metadata from the daemon

A proxy contains some meta-data about the object it connects to. It obtains the data via the (public) Pyro4.core.DaemonObject.get_metadata() method on the daemon that it connects to. This method returns the following information about the object (or rather, its class): what methods and attributes are defined, and which of the methods are to be called as one-way. This information is used to properly execute one-way calls, and to do client-side validation of calls on the proxy (for instance to see if a method or attribute is actually available, without having to do a round-trip to the server). Also this enables a properly working hasattr on the proxy, and efficient and specific error messages if you try to access a method or attribute that is not defined or not exposed on the Pyro object. Lastly the direct access to attributes on the remote object is also made possible, because the proxy knows about what attributes are available.

For backward compatibility with old Pyro4 versions (4.26 and older) you can disable this mechanism by setting the METADATA config item to False (it’s True by default). You can tell if you need to do this if you’re getting errors in your proxy saying that ‘DaemonObject’ has no attribute ‘get_metadata’. Either upgrade the Pyro version of the server, or set the METADATA config item to False in your client code.