Java collections are automatically mapped to Python collections so that standard Python operations such as slicing work on Java collections. Here is the mapping of the collection:
Java Collection |
Python Collection |
Py4J Implementation |
|---|---|---|
Array |
Sequence [1] |
|
java.util.List |
MutableSequence |
|
java.util.Set |
MutableSet |
|
java.util.Map |
MutableMapping |
|
java.util.Iterator |
Iterator Protocol |
Java methods are still accessible when using the Python version of a Java collection. Here are some usage examples for each collection class. These examples do not cover the entire API.
It is possible to change this default mapping by registering an output
converter with register_output_converter
>>> gateway = JavaGateway()
>>> int_class = gateway.jvm.int
>>> int_array = gateway.new_array(int_class,2)
>>> int_array[0] = 1
>>> int_array[1] = 2
>>> int_array[0]
1
>>> int_array[2]
Traceback (most recent call last):
...
IndexError: list index out of range
>>> for i in int_array:
... print(i)
...
1
2
>>> sarray = gateway.new_array(gateway.jvm.java.lang.String,2,3)
>>> len(sarray)
2
>>> len(sarray[0])
3
>>> sarray[0][1] = 'hello'
>>> sarray[0][1]
u'hello'
>>> sarray[0][0] == None
True
>>> l = gateway.jvm.java.util.ArrayList()
>>> l.append(1) # calling Python interface
>>> l.add('hello') # calling Java interface
>>> for elem in l:
... print elem
...
1
hello
>>> l[0] = 2
>>> l.append(3)
>>> str(l)
"[2, u'hello', 3]"
>>> l2 = l[0:-1]
>>> l2[0] = 999
>>> l
[2, u'hello', 3]
>>> l2 # l2 is a copy of l and not a view so a change in l2 does not affect l
[999, u'hello']
>>> del(l[0])
>>> l
[u'hello', 3]
>>> s = gateway.jvm.java.util.HashSet()
>>> s.add(1)
>>> s.add('hello')
>>> s
set([1, u'hello'])
>>> 1 in s
True
>>> s.remove(u'hello')
>>> s
set([1])
>>> m = gateway.jvm.java.util.HashMap()
>>> m["a"] = 0
>>> m.put("b",1)
>>> m
{u'a': 0, u'b': 1}
>>> u"b" in m
True
>>> del(m["a"])
>>> m
{u'b': 1}
>>> m["c"] = 2
>>> for key in m:
... print("%s:%i" % (key,m[key]))
...
b:1
c:2
Since version 0.7, Py4J automatically passes Java byte array (i.e., byte[]) by value and convert them to Python bytearray (2.x) or bytes (3.x) and vice versa. The rationale is that byte array are often used for binary processing and are often immutable: a program reads a series of byte from a data source and interpret it (or transform it into another byte array).
Prior to 0.7, a Python program could interact with a Java byte[] but each access to a byte required a call between the Python and the Java interpreter (i.e., the Python program only had a reference to the array).
In summary:
If from Java, you return a byte[], Py4J will convert the byte[] to a bytearray (Python 2.x) or bytes (Python 3.x) variable in Python.
If from Python, you pass a bytearray or bytes variable to the Java side, Py4J will convert it to a byte[].
If you pass a byte[] from Java to Python, both arrays are disconnected: a change to the array on one side will not be reflected on the other side.
If you want to pass an array by reference, use a Byte[] instead of a byte[] or you can use Bytes.asList from Guava to obtain a list backed by a byte array.
Since version 0.3, Py4J allows Python classes to implement Java interfaces so that the JVM can call back Python objects. In the following example, a Python class implements a Java listener interface.
Here is the code of listener interface:
// py4j/examples/ExampleListener.java
package py4j.examples;
public interface ExampleListener {
Object notify(Object source);
}
Here is the code of the main Java application. The program has a main method starting a GatewayServer. The entry point, a ListenerApplication instance, has two methods, one to register listeners, the other to notify all listeners. Start this program before running the Python code below.
// py4j/examples/ListenerApplication.java
package py4j.examples;
import py4j.GatewayServer;
import java.util.ArrayList;
import java.util.List;
public class ListenerApplication {
List<ExampleListener> listeners = new ArrayList<ExampleListener>();
public void registerListener(ExampleListener listener) {
listeners.add(listener);
}
public void notifyAllListeners() {
for (ExampleListener listener: listeners) {
Object returnValue = listener.notify(this);
System.out.println(returnValue);
}
}
@Override
public String toString() {
return "<ListenerApplication> instance";
}
public static void main(String[] args) {
ListenerApplication application = new ListenerApplication();
GatewayServer server = new GatewayServer(application);
server.start(true);
}
}
Here is the Python program that implements the Java interface and calls the ListenerApplication to notify all listeners. Then multiple exchanges between Java and Python occurs (e.g., Python calls System.out.println on the Java side).
from py4j.java_gateway import JavaGateway, CallbackServerParameters
class PythonListener(object):
def __init__(self, gateway):
self.gateway = gateway
def notify(self, obj):
print("Notified by Java")
print(obj)
gateway.jvm.System.out.println("Hello from python!")
return "A Return Value"
class Java:
implements = ["py4j.examples.ExampleListener"]
if __name__ == "__main__":
gateway = JavaGateway(
callback_server_parameters=CallbackServerParameters())
listener = PythonListener(gateway)
gateway.entry_point.registerListener(listener)
gateway.entry_point.notifyAllListeners()
gateway.shutdown()
The PythonListener class is a standard Python class that has one method, notify. The signature of the method contains one parameter. When interfaces contain overloaded methods, the python method must accept all possible combinations of parameters (with *args and **kwargs or with default parameters).
Py4J recognizes that the PythonListener class implements a Java interface because it declares an internal class called Java, which has a member named implements. This member is a list of string representing the fully qualified name of implemented Java interfaces.
Finally, the Python program contains a main method that starts a gateway, initializes a PythonListener instance and registers it to the ListenerApplication instance on the Java side. Then, it calls the notifyAllListeners() method, which will notify all listeners. Py4J takes care of creating the necessary proxies: the notify method of the PythonListener class is called in the Java VM, but the method is executed in the Python interpreter.
Warning
Python classes can only implement Java interfaces. Abstract or concrete classes are not supported because Java does not natively support dynamic proxies for classes. Extending classes may be supported in future releases of Py4J.
As a workaround, a subclass of the abstract class could be created on the Java side. The methods of the subclass would call the methods of a custom interface that a Python class could implement.
Warning
If you want to implement an interface declared in a class (i.e., an internal class), you need to prefix the name of the interface with a dollar sign. For example, if the interface Operator is declared in the class package1.MyClass, you will have to write:
implements = [‘package1.MyClass$Operator’]
When you use a JavaGateway and a CallbackServer, Py4J will create a new thread every time a command is received and there is no thread available. This can lead to the creation of many threads if there are recursive calls between Python and Java. For example, let’s consider this code:
class PythonPlayer(object):
def start(self, player):
return player.firstPing(self)
def firstPong(self, player):
return player.secondPing(self)
def secondPong(self, player):
return "Success"
class Java:
implements = ["py4j.examples.PongPlayer"]
# Start the JVM with "java -cp py4j.jar py4j.examples.ExampleApplication"
from py4j.java_gateway import JavaGateway, CallbackServerParameters
gateway = JavaGateway(
callback_server_parameters=CallbackServerParameters())
ping_player = java_gateway.jvm.py4j.examples.PingPlayer()
pong_player = PongPlayer()
print(pong_player.start(ping_player))
// PongPlayer.java
package py4j.examples;
public interface PongPlayer {
String firstPong(PingPlayer player);
String secondPong(PingPlayer player);
}
// PingPlayer.java
package py4j.examples;
public class PingPlayer {
public String firstPing(PongPlayer player) {
return player.firstPong(this);
}
public String secondPing(PongPlayer player) {
return player.secondPong(this);
}
}
When this code is executed, the Python side creates 2 threads and the Java side creates two threads (note that the Java and Python sides both start with an implicit thread so there are a minimum of three threads on either side).
The Python client initiates the conversation by calling firstPing() and waits for a response.
The Java server receives the call and creates Java Thread 1 to execute the code. Java then calls firstPong and waits for a response.
The Python server receives the call and creates Python Thread 1 to execute the code. Python then calls secondPing and waits for a response.
The Java server receives the call and creates Java Thread 2 to execute the code. Java then calls secondPong and waits for a response.
The Python server receives the call and creates Python Thread 2 to execute
the code. Python then returns the response "Success". Python Thread 2
can now be used to execute other calls.
Java Thread 2 (secondPing) receives a response and sends back the response to the Python side. Java Thread 2 can now be used to execute other calls.
Python Thread 1 (firstPong) receives a response and sends back the response to the Java side. Python Thread 1 can now be used to execute other calls.
Java Thread 1 (firstPing) receives a response and sends back the response to the Python side. Java Thread 1 can now be used to execute other calls.
The Python side receives the response and prints it.
Although the multi threading model is sound and has been used by many distributed systems (e.g., Java RMI), it has two main issues: (1) it creates a lot of threads when the Java and Python sides exchange many calls such as when there is indirect recursion between the two sides, and (2) it is not possible to control in which thread the code will be executed. The first issue can severely impact performance and the second issue limits the usefulness of Py4J in certain applications (e.g., GUI applications).
To work around these limitations, the py4j.clientserver module
implements a single threading model also called pinned thread model that
ensures that callbacks received from the other side will be executed in the
initiating thread. For example, if we modify the previous code with this one:
# Start the JVM with "java -cp py4j.jar py4j.examples.SingleThreadApplication"
from py4j.clientserver import ClientServer, JavaParameters, PythonParameters
gateway = ClientServer(
java_parameters=JavaParameters(),
python_parameters=PythonParameters())
ping_player = gateway.jvm.py4j.examples.PingPlayer()
pong_player = PongPlayer()
print(pong_player.start(ping_player))
Only one Java Thread will be created and no Python thread will be created:
The Python client initiates the conversation from Python Thread 1 by calling firstPing() and waits for either a response or a call to execute.
The Java server receives the call and creates Java Thread 1 to execute the code. Java then calls firstPong and waits for a response or a call to execute.
Python Thread 1 receives the call and calls secondPing. It waits for a response or a call.
Java Thread 1 receives the call and then call secondPong. It waits for a response or a call.
Python Thread 1 receives the call responds with "Success".
Java Thread 1 (secondPing) receives the response and sends the response to the Python side.
Python Thread 1 (firstPong) receives the response and sends the response to the Java side.
Java Thread 1 (firstPing) receives the response and sends the response to the Python side.
Python Thread 1 (start) receives the response and prints it.
In total, only one thread was created (on the Java side). If for whatever reason multiple threads are created on the Python side and they each call the Java side, a corresponding Java thread will be created. For example, if three threads from the Python side are calling the Java side, three Java threads will be created.
All the previous examples assume that you want to initiate the communication between Python and Java from the Python side. What if you want to start calling Python from the Java side?
If you want Java to call Python first, you need to:
Start the Python process and create either a JavaGateway or a ClientServer.
Pass a Python instance implementing a Java interface as a
python_server_entry_point parameter.
Start a GatewayServer or ClientServer on the Java side.
Call getPythonServerEntryPoint by providing the list of interfaces the
Python entry point is expected to implement.
// IHello.java
package py4j.examples;
public interface IHello {
public String sayHello();
public String sayHello(int i, String s);
}
// ExampleClientApplication.java
package py4j.examples;
import py4j.GatewayServer;
public class ExampleClientApplication {
public static void main(String[] args) {
GatewayServer.turnLoggingOff();
GatewayServer server = new GatewayServer();
server.start();
IHello hello = (IHello) server.getPythonServerEntryPoint(new Class[] { IHello.class });
try {
hello.sayHello();
hello.sayHello(2, "Hello World");
} catch (Exception e) {
e.printStackTrace();
}
server.shutdown();
}
}
class SimpleHello(object):
def sayHello(self, int_value=None, string_value=None):
print(int_value, string_value)
return "Said hello to {0}".format(string_value)
class Java:
implements = ["py4j.examples.IHello"]
# Make sure that the python code is started first.
# Then execute: java -cp py4j.jar py4j.examples.SingleThreadClientApplication
from py4j.java_gateway import JavaGateway, CallbackServerParameters
simple_hello = SimpleHello()
gateway = JavaGateway(
callback_server_parameters=CallbackServerParameters(),
python_server_entry_point=simple_hello)
In the following example, the Java side is initiating the conversation by making the first call to the Python side. If the call is initiated from a UI thread, then all subsequent call from Python to Java will be executed in the Java UI thread.
// IHello.java
package py4j.examples;
public interface IHello {
public String sayHello();
public String sayHello(int i, String s);
}
// SingleThreadClientApplication.java
package py4j.examples;
import py4j.ClientServer;
import py4j.GatewayServer;
public class SingleThreadClientApplication {
public static void main(String[] args) {
ClientServer clientServer = new ClientServer(null);
// We get an entry point from the Python side
IHello hello = (IHello) clientServer.getPythonServerEntryPoint(new Class[] { IHello.class });
// Java calls Python without ever having been called from Python
System.out.println(hello.sayHello());
System.out.println(hello.sayHello(2, "Hello World"));
clientServer.shutdown();
}
}
class SimpleHello(object):
def sayHello(self, int_value=None, string_value=None):
print(int_value, string_value)
return "Said hello to {0}".format(string_value)
class Java:
implements = ["py4j.examples.IHello"]
# Make sure that the python code is started first.
# Then execute: java -cp py4j.jar py4j.examples.SingleThreadClientApplication
from py4j.clientserver import ClientServer, JavaParameters, PythonParameters
simple_hello = SimpleHello()
gateway = ClientServer(
java_parameters=JavaParameters(),
python_parameters=PythonParameters(),
python_server_entry_point=simple_hello)
If you do not want to use Py4J’s default port (25333 on the Java side and 25334 on the Python side to receive callback), you can use the port 0 and Py4J will use the next available port. Once a port is assigned, the other side needs to be aware of this port. Here is one way to do it:
from py4j.java_gateway import (
JavaGateway, CallbackServerParameters, GatewayParameters,
launch_gateway)
# launch Java side with dynamic port and get back the port on which the
# server was bound to.
port = launch_gateway()
# connect python side to Java side with Java dynamic port and start python
# callback server with a dynamic port
gateway = JavaGateway(
gateway_parameters=GatewayParameters(port=port),
callback_server_parameters=CallbackServerParameters(port=0))
# retrieve the port on which the python callback server was bound to.
python_port = gateway.get_callback_server().get_listening_port()
# tell the Java side to connect to the python callback server with the new
# python port. Note that we use the java_gateway_server attribute that
# retrieves the GatewayServer instance.
gateway.java_gateway_server.resetCallbackClient(
gateway.java_gateway_server.getCallbackClient().getAddress(),
python_port)
# Test that callbacks work
from py4j.tests.java_callback_test import IHelloImpl
hello = IHelloImpl()
example = gateway.jvm.py4j.examples.ExampleClass()
example.callHello(hello)
If you try to pass a Python collection to a method that expects a Java collection, an error will be thrown:
>>> my_list = [3,2,1]
>>> gateway.jvm.java.util.Collections.sort(my_list)
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
File "py4j/java_gateway.py", line 347, in __call__
args_command = ''.join([get_command_part(arg, self.pool) for arg in new_args])
File "py4j/protocol.py", line 195, in get_command_part
command_part = REFERENCE_TYPE + parameter._get_object_id()
AttributeError: 'list' object has no attribute '_get_object_id'
You can explicitly convert Python collections using one of the following converter located in the py4j.java_collections module: SetConverter, MapConverter, ListConverter.
>>> from py4j.java_collections import SetConverter, MapConverter, ListConverter
>>> java_list = ListConverter().convert(my_list, gateway._gateway_client)
>>> gateway.jvm.java.util.Collections.sort(java_list)
>>> java_list
[1, 2, 3]
>>> my_list
[3, 2, 1]
Note that the Python list is totally disconnected from the Java list. The Java
List is actually a copy. You can also ask Py4J to automatically convert Python
collections to Java Collections when calling a Java method: just set
auto_convert=True when creating a JavaGateway:
>>> gateway = JavaGateway(gateway_parameters=GatewayParameters(auto_convert=True))
>>> my_list
[3, 2, 1]
>>> gateway.jvm.java.util.Collections.sort(my_list)
>>> my_list # The python list is not sorted!
[3, 2, 1]
>>> gateway.jvm.java.util.Collections.frequency(my_list,2)
1
Again, note that my_list is not sorted because when calling
Collections.sort(), Py4J only makes a copy of the Python list. Still, a copy
can be useful if you do not expect the list to be modified by the Java method
like in the call to frequency().
Order of Automatic Conversion
When auto_convert=True, Py4J will attempt to automatically convert Python
objects that are not an instance of basestring or JavaObject. By
default, Py4J performs the following checks and conversions:
If the Python object is an instance of collections.Set, it is converted to a HashSet.
If the object has the methods keys() and __getitem__, it is converted to a HashMap
If the object is iterable, it is converted to an ArrayList.
Otherwise, standard Py4J primitive type conversion is attempted (e.g., bool to boolean).
It is possible to add custom converters by calling
register_input_converter(). Look at the source code of the
default converters for an example. Note that automatic conversion makes calling
Java methods slightly less efficient because in the worst case, Py4J needs to
go through all registered converters for all parameters. This is why automatic
conversion is disabled by default.
Py4J can raise three exceptions on the Python side:
Py4JJavaError. This exception is
raised when an exception occurs in the Java client code. For example, if you
try to pop an element from an empty stack. The instance of the Java exception
thrown is stored in the java_exception member.
Py4JNetworkError. This exception is
raised when a problem occurs during network transfer (e.g., connection lost).
Py4JError. This exception is raised when
any other error occurs such as when the client program tries to access an
object that no longer exists on the Java side.
Both Py4JJavaError and Py4JNetworkError inherits from Py4JError so it is possible to catch all related Py4J errors with one except clause:
try:
java_object.doSomething()
except Py4JError:
traceback.print_exc()
If a Py4J Exception wraps another exception, the original exception will be
available in the clause field.
try:
java_object.doSomething()
except Py4JError as e:
if e.cause:
print(e.cause)
Py4J allows you to import packages so that you don’t have to type the fully qualified name of the classes you want to instantiate. The java.lang package is always automatically imported.
>>> from py4j.java_gateway import JavaGateway
>>> gateway = JavaGateway()
>>> from py4j.java_gateway import java_import
>>> java_import(gateway.jvm,'java.util.*')
>>> jList = gateway.jvm.ArrayList()
>>> jMap = gateway.jvm.HashMap()
>>> gateway.jvm.java.lang.String("a")
u'a'
>>> gateway.jvm.String("a")
u'a'
As opposed to Java where import statements do not cross compilation units (java source files), the jvm instance can be shared across multiple Python modules: in other words, import statements are global.
The recommended way to use import statements is to use one JVMView instance per Python module. Here is an example on
how to create and use a JVMView:
>>> module1_view = gateway.new_jvm_view()
>>> module2_view = gateway.new_jvm_view()
>>> jList2 = module1_view.ArrayList()
Py4JError: Trying to call a package.
...
>>> java_import(module1_view,'java.util.ArrayList')
>>> jList2 = module1_view.ArrayList()
>>> jList3 = module2_view.ArrayList()
Py4JError: Trying to call a package.
...
As you can see from the previous example, the import of java.util.ArrayList only affects module1_view.
Note
In fact, the gateway.jvm member is also an instance of JVMView. It is automatically created when a gateway is
initialized.
Py4J can be used with Eclipse like any normal Java program. A plug-in needs to instantiate and start a GatewayServer. By default, the GatewayServer will only be able to access the classes declared in the plug-in or one of its dependencies.
Unless they have specific needs, users are encouraged to use the Eclipse plug-ins provided by Py4J available on the following update site:
http://eclipse.py4j.org/
The plug-in org.py4j.py4j-java contains the same code as the production jar published on maven or pypi and it also contains a OSGi MANIFEST.MF. See the build instructions to build the OSGi bundle locally.
You can also use legacy plug-ins that will eventually be moved to the new build process on the following update site:
http://eclipse-legacy.py4j.org/
The first plug-in, net.sf.py4j, provides all the Py4J Java classes such as GatewayServer. The plug-in comes with the source and the javadoc. The plug-in also declares a global buddy policy which allows the GatewayServer to access any class declared in any plug-in loaded with Eclipse.
The second plug-in, net.sf.py4j.defaultserver, instantiates a GatewayServer and starts it as soon as Eclipse is started (no lazy loading). The ports used by the default server can be changed in the Py4J Preferences page. The server is also accessible at runtime:
import net.sf.py4j.defaultserver.DefaultServerActivator;
...
GatewayServer server = DefaultServerActivator.getDefault().getServer();
Here is a short example of what you could do with Py4J and Eclipse:
>>> from py4j.java_gateway import JavaGateway, java_import
>>> gateway = JavaGateway()
>>> jvm = gateway.jvm
>>> java_import(jvm, 'org.eclipse.core.resources.*')
>>> workspace_root = jvm.ResourcesPlugin.getWorkspace().getRoot()
>>> gateway.help(workspace_root,'*Projects*')
Help on class WorkspaceRoot in package org.eclipse.core.internal.resources:
WorkspaceRoot extends org.eclipse.core.internal.resources.Container implements org.eclipse.core.resources.IWorkspaceRoot {
|
| Methods defined here:
|
| getProjects() : IProject[]
|
| getProjects(int) : IProject[]
|
| ------------------------------------------------------------
| Fields defined here:
|
| ------------------------------------------------------------
| Internal classes defined here:
|
}
>>> project_names = [project.getName() for project in workspace_root.getProjects()]
>>> print(project_names)
[u'test2', u'testplugin', u'testplugin2']
Support for Eclipse was introduced in Py4J 0.5 and more features may be added in the future.
On the Java side, it is possible to listen to GatewayServer events by
registering an instance of py4j.GatewayServerListener with
GatewayServer.addListener.
On the Python side, it is possible to listen to CallbackServer events using signals. Recognizing that signals may be a foreign concept to Java developers, here is a full example:
from py4j.java_gateway import (
server_connection_started, server_connection_stopped,
server_started, server_stopped, pre_server_shutdown, post_server_shutdown,
JavaGateway, GatewayParameters, CallbackServerParameters)
def started(sender, **kwargs):
server = kwargs["server"]
# do something
def connection_started(sender, **kwargs):
connection = kwargs["connection"]
# do something
def connection_stopped(sender, **kwargs):
connection = kwargs["connection"]
# do something
def stopped(sender, **kwargs):
server = kwargs["server"]
# do something
def pre_shutdown(sender, **kwargs):
server = kwargs["server"]
# do something
def post_shutdown(sender, **kwargs):
server = kwargs["server"]
# do something
# Because the server is created and started at the same time,
# you must listen to all started signals. You will receive the server
# instance that was started as a parameter and sender.
server_started.connect(started)
gateway = JavaGateway(
gateway_parameters=GatewayParameters(),
callback_server_parameters=CallbackServerParameters())
server_stopped.connect(
stopped, sender=gateway.get_callback_server())
server_connection_started.connect(
connection_started,
sender=gateway.get_callback_server())
server_connection_stopped.connect(
connection_stopped,
sender=gateway.get_callback_server())
pre_server_shutdown.connect(
pre_shutdown, sender=gateway.get_callback_server())
post_server_shutdown.connect(
post_shutdown, sender=gateway.get_callback_server())
You can read more information about the signals module or
the Py4J signals.
Java objects sent to the Python side
Every time a Java object is sent to the Python side, a reference to the object is kept on the Java side (in the Gateway class). Once the object is garbage collected on the Python VM (reference count == 0), the reference is removed on the Java VM: if this was the last reference, the object will likely be garbage collected too. When a gateway is shut down, the remaining references are also removed on the Java VM.
Because Java objects on the Python side are involved in a circular reference
(JavaObject and JavaMember reference each other), these objects are not
immediately garbage collected once the last reference to the object is removed
(but they are guaranteed to be eventually collected if the Python garbage
collector runs before the Python program exits).
In doubt, users can always call the detach function on the Python gateway to
explicitly delete a reference on the Java side. A call to gc.collect() also
usually works.
Python objects sent to the Java side (callback)
Every time a Python object is sent to the Java side, a reference to this object
is kept on the Python side (by a PythonProxyPool). Once a python object is garbage
collected on the Java side, a message is sent to the Python side to remove the
reference to the Python object. When a gateway is shut down, the remaining
references are removed from the Python VM.
Unfortunately, there is no guarantee that the garbage collection message will ever be sent to the Python side (it usually works on Sun/Oracle VM). It might thus be necessary to manually remove the reference to the Python objects. Some helper functions will be developed in the future, but it is unlikely that garbage collection will be guarenteed because of the specifications of Java finalizers (which are surprisingly worse than Python finalizer strategies).
In its default mode, Py4J allocates one thread per connection. The design of Py4j is symmetrical on the Python and Java sides. A Python GatewayClient communicates with the Java GatewayServer and is then associated with a GatewayConnection. A Java CallbackClient (for callbacks) communicates with the Python CallbackServer and is then associated with a CallbackConnection. A connection runs in the calling thread.
And now, for the details:
On the Python side: calling Java
Py4J on the Python side does not explicitly create a thread to call Java methods. When a method is called, a connection to the Java GatewayServer is established in the calling thread. If multiple threads are calling Java methods concurrently, Py4J will ensure that each thread has its own connection by requesting more connections.
To be extra clear: if you only call Java, Py4J on the Python side will never create a thread.
To be extra extra clear: Py4J is thread-safe so if you use a JavaGateway from multiple threads that you created, Py4J will make manage the network resources appropriately.
On the Python side: receiving callbacks from Java
Py4J explicitly creates a thread to run the
CallbackServer, which accepts
callback connection requests, and a thread for each callback connection
request. As long as there is no concurrent callback from the Java side, the
same callback connection/thread will be used.
These threads are necessary to prevent deadlocks. For example, if we only had a single thread to handle callbacks from Java, Py4J would deadlock as soon as it would encounter an indirect recursion between Java and Python functions. Early versions of Py4J made this mistake :-)
On the Java side: receiving calls from Python
Py4J explicitly creates a thread to run the GatewayServer, which accepts connection requests (from a GatewayClient), and a thread for each connection request. As long as there is no concurrent call from the Python side, the same connection/thread will be used.
On the Java side: calling back Python
Py4J on the Java side does not explicitly create a thread to make a callback to a Python object. When a callback is called, a connection to the CallbackServer is established in the calling thread. If you created multiple threads in Java to call back Python concurrently, Py4J will ensure that each thread has its own CallbackConnection.
Alternative threading model
If both sides are making multiple calls to each other (e.g., indirect recursion), or if you want to control which thread will execute the calls, you should use the single threading model, which provide a more efficient threading model.
Py4J supports TLS for both connections from Python to the JVM and the callback connections to Python. This requires configuring both Java and Python:
On the Python side
Use Python’s ssl module to create an ssl.SSLContext. For the connection to the JVM you’ll need something like:
client_ssl_context = ssl.SSLContext(ssl.PROTOCOL_TLSv1)
client_ssl_context.verify_mode = ssl.CERT_REQUIRED
client_ssl_context.check_hostname = True
client_ssl_context.load_verify_locations(cafile='/path/to/pem/ca/certs')
This ssl.SSLContext should be set as the ssl_context parameter when constructing an instanceof of GatewayParameters. If you’re setting check_hostname to True like the code above then the address parameter of the GatewayParameters must match (one of) the hostnames in the certificate the Java GatewayServer presents.
For the socket listening for callbacks from the JVM, you’ll need something like:
server_ssl_context = ssl.SSLContext(ssl.PROTOCOL_TLSv1)
server_ssl_context.load_cert_chain(
'/path/to/pem/private/key',
password='key-password')
This ssl.SSLContext should be set as the ssl_context parameter when constructing an instance of CallbackServerParameters. You can see a full example in the test file py4j-python/src/tests/java_tls_test.py.
On the Java side
Use the GatewayServer constructor that takes a ServerSocketFactory, and pass an instance of type SSLServerSocketFactory. The easiest way of constructing one of these is using SSLServerSocketFactory.getDefault(), however for more control construct a SSLContext.
To add TLS to callbacks, create a CallbackClient with the constructor that takes a SSLSocketFactory. You can get an instance of this via getDefault() or a SSLContext instance similarly.
A runnable example is in py4j.examples.ExampleSSLApplication.
By default, Py4J does not authenticate client connections. This allows any user who can open a connection to the Py4J daemons to potentially execute code as the user running those daemons.
Authentication can be enabled by providing an authentication token to the different servers. See the API documentation for how to provide the authentication token through the API.
Py4J leaves it to the application to generate and distribute the authentication tokens to all the needed processes.
The authentication protocol transfers the token over the socket connection, so it relies on transport-level encryption for secrecy. This means that for proper security, connections should use loopback addresses, or use TLS.
Here is a quick example of how to use an authentication token.
GatewayServer server = new GatewayServer.GatewayServerBuilder().entryPoint(
new ExampleEntryPoint()).authToken("HelloWorld").build();
server.start();
from py4j.java_gateway import JavaGateway, GatewayParameters
params = GatewayParameters(auth_token="HelloWorld")
gateway = JavaGateway(gateway_parameters=params)
Boxed Java primitives, such as java.lang.Integer are unboxed in the JVM and transmitted as primitives, so they will appear as Python int and not JavaObjects.
We track performance with each new release by running a benchmark suite that tests various features of Py4J.
The benchmark data and performance evolution charts are made available as google charts.
The following chart illustrates the progress Py4J made so far for transferring 1 MB byte arrays.