Quickstart¶
This assumes you already have the capnp library installed. If you don’t, please follow the instructions at Installation first.
In general, this library is a very light wrapping of the Cap’n Proto C++ library. You can refer to its docs for more advanced concepts, or just to get a basic idea of how the python library is structured.
Load a Cap’n Proto Schema¶
First you need to import the library:
import capnp
Then you can load the Cap’n Proto schema with:
import addressbook_capnp
This will look all through all the directories in your sys.path/PYTHONPATH, and try to find a file of the form ‘addressbook.capnp’. If you want to disable the import hook magic that import capnp adds, and load manually, here’s how:
capnp.remove_import_hook()
addressbook_capnp = capnp.load('addressbook.capnp')
For future reference, here is the Cap’n Proto schema. Also available in the github repository under examples/addressbook.capnp:
# addressbook.capnp
@0x934efea7f017fff0;
const qux :UInt32 = 123;
struct Person {
id @0 :UInt32;
name @1 :Text;
email @2 :Text;
phones @3 :List(PhoneNumber);
struct PhoneNumber {
number @0 :Text;
type @1 :Type;
enum Type {
mobile @0;
home @1;
work @2;
}
}
employment :union {
unemployed @4 :Void;
employer @5 :Text;
school @6 :Text;
selfEmployed @7 :Void;
# We assume that a person is only one of these.
}
}
struct AddressBook {
people @0 :List(Person);
}
Const values¶
Const values show up just as you’d expect under the loaded schema. For example:
print addressbook_capnp.qux
# 123
Build a message¶
Initialize a New Cap’n Proto Object¶
Now that you’ve imported your schema, you need to allocate an actual struct from that schema. In this case, we will allocate an AddressBook:
addresses = addressbook_capnp.AddressBook.new_message()
Notice that we used addressbook_capnp from the previous section: Load a Cap’n Proto Schema.
Also as a shortcut, you can pass keyword arguments to the new_message function, and those fields will be set in the new message:
person = addressbook_capnp.Person.new_message(name='alice')
# is equivalent to:
person = addressbook_capnp.Person.new_message()
person.name = 'alice'
List¶
Allocating a list inside of an object requires use of the init function:
people = addresses.init('people', 2)
For now, let’s grab the first element out of this list and assign it to a variable named alice:
alice = people[0]
Note
It is a very bad idea to call init more than once on a single field. Every call to init allocates new memory inside your Cap’n Proto message, and if you call it more than once, the previous memory is left as dead space in the message. See Tips and Best Practices for more details.
Primitive Types¶
For all primitive types, from the Cap’n Proto docs:
Boolean: Bool
Integers: Int8, Int16, Int32, Int64
Unsigned integers: UInt8, UInt16, UInt32, UInt64
Floating-point: Float32, Float64
Blobs: Text, Data
You can assign straight to the variable with the corresponding Python type. For Blobs, you use strings. Assignment happens just by using the . syntax on the object you contstructed above:
alice.id = 123
alice.name = 'Alice'
alice.email = 'alice@example.com'
Note
Text fields will behave differently depending on your version of Python. In Python 2.x, Text fields will expect and return a bytes string, while in Python 3.x, they will expect and return a unicode string. Data fields will always a return bytes string.
Enums¶
First we’ll allocate a length one list of phonenumbers for alice:
alicePhone = alice.init('phones', 1)[0]
Note that even though it was a length 1 list, it was still a list that was returned, and we extracted the first (and only) element with [0].
Enums are treated like strings, and you assign to them like they were a Text field:
alicePhone.type = 'mobile'
If you assign an invalid value to one, you will get a ValueError:
alicePhone.type = 'foo'
---------------------------------------------------------------------------
ValueError Traceback (most recent call last)
...
ValueError: src/capnp/schema.c++:326: requirement not met: enum has no such enumerant; name = foo
Unions¶
For the most part, you just treat them like structs:
alice.employment.school = "MIT"
Now the school field is the active part of the union, and we’ve assigned ‘MIT’ to it. You can query which field is set in a union with which(), shown in Reading Unions
Also, one weird case is for Void types in Unions (and in general, but Void is really only used in Unions). For these, you will have to assign None to them:
bob.employment.unemployed = None
Note
One caveat for unions is having structs as union members. Let us assume employment.school was actually a struct with a field of type Text called name:
alice.employment.school.name = "MIT"
# Raises a ValueError
The problem is that a struct within a union isn’t initialized automatically. You have to do the following:
school = alice.employment.init('school')
school.name = "MIT"
Note that this is similar to init for lists, but you don’t pass a size. Requiring the init makes it more clear that a memory allocation is occurring, and will hopefully make you mindful that you shouldn’t set more than 1 field inside of a union, else you risk a memory leak
Writing to a File¶
Once you’re done assigning to all the fields in a message, you can write it to a file like so:
f = open('example.bin', 'w+b')
addresses.write(f)
There is also a write_packed function, that writes out the message more space-efficientally. If you use write_packed, make sure to use read_packed when reading the message.
Read a message¶
Reading from a file¶
Much like before, you will have to de-serialize the message from a file descriptor:
f = open('example.bin', 'rb')
addresses = addressbook_capnp.AddressBook.read(f)
Note that this very much needs to match the type you wrote out. In general, you will always be sending the same message types out over a given channel or you should wrap all your types in an unnamed union. Unnamed unions are defined in the .capnp file like so:
struct Message {
union {
person @0 :Person;
addressbook @1 :AddressBook;
}
}
Reading Fields¶
Fields are very easy to read. You just use the . syntax as before. Lists behave just like normal Python lists:
for person in addresses.people:
print(person.name, ':', person.email)
for phone in person.phones:
print(phone.type, ':', phone.number)
Reading Unions¶
The only tricky one is unions, where you need to call .which() to determine the union type. The .which() call returns an enum, ie. a string, corresponding to the field name:
which = person.employment.which()
print(which)
if which == 'unemployed':
print('unemployed')
elif which == 'employer':
print('employer:', person.employment.employer)
elif which == 'school':
print('student at:', person.employment.school)
elif which == 'selfEmployed':
print('self employed')
print()
Serializing/Deserializing¶
Files¶
As shown in the examples above, there is file serialization with write():
addresses = addressbook_capnp.AddressBook.new_message()
...
f = open('example.bin', 'w+b')
addresses.write(f)
And similarly for reading:
f = open('example.bin', 'rb')
addresses = addressbook_capnp.AddressBook.read(f)
There are packed versions as well:
addresses.write_packed(f)
f.seek(0)
addresses = addressbook_capnp.AddressBook.read_packed(f)
Multi-message files¶
The above methods only guaranteed to work if your file contains a single message. If you have more than 1 message serialized sequentially in your file, then you need to use these convenience functions:
addresses = addressbook_capnp.AddressBook.new_message()
...
f = open('example.bin', 'w+b')
addresses.write(f)
addresses.write(f)
addresses.write(f) # write 3 messages
f.seek(0)
for addresses in addressbook_capnp.AddressBook.read_multiple(f):
print addresses
There is also a packed version:
for addresses in addressbook_capnp.AddressBook.read_multiple_packed(f):
print addresses
Dictionaries¶
There is a convenience method for converting Cap’n Proto messages to a dictionary. This works for both Builder and Reader type messages:
alice.to_dict()
For the reverse, all you have to do is pass keyword arguments to the new_message constructor:
my_dict = {'name' : 'alice'}
alice = addressbook_capnp.Person.new_message(**my_dict)
# equivalent to: alice = addressbook_capnp.Person.new_message(name='alice')
It’s also worth noting, you can use python lists/dictionaries interchangably with their Cap’n Proto equivalent types:
book = addressbook_capnp.AddressBook.new_message(people=[{'name': 'Alice'}])
...
book = addressbook_capnp.AddressBook.new_message()
book.init('people', 1)
book.people[0] = {'name': 'Bob'}
Byte Strings/Buffers¶
There is serialization to a byte string available:
encoded_message = alice.to_bytes()
And a corresponding from_bytes function:
alice = addressbook_capnp.Person.from_bytes(encoded_message)
There are also packed versions:
alice2 = addressbook_capnp.Person.from_bytes_packed(alice.to_bytes_packed())
Byte Segments¶
Note
This feature is not supported in PyPy at the moment, pending investigation.
Cap’n Proto supports a serialization mode which minimizes object copies. In the C++ interface, capnp::MessageBuilder::getSegmentsForOutput() returns an array of pointers to segments of the message’s content without copying. capnp::SegmentArrayMessageReader performs the reverse operation, i.e., takes an array of pointers to segments and uses the underlying data, again without copying. This produces a different wire serialization format from to_bytes() serialization, which uses capnp::messageToFlatArray() and capnp::FlatArrayMessageReader (both of which use segments internally, but write them in an incompatible way).
For compatibility on the Python side, use the to_segments() and from_segments() functions:
segments = alice.to_segments()
This returns a list of segments, each a byte buffer. Each segment can be, e.g., turned into a ZeroMQ message frame. The list of segments can also be turned back into an object:
alice = addressbook_capnp.Person.from_segments(segments)
For more information, please refer to the following links:
Advice on minimizing copies from Cap’n Proto (from the author of Cap’n Proto)
Advice on using Cap’n Proto over ZeroMQ (from the author of Cap’n Proto)
Discussion about sending and reassembling Cap’n Proto message segments in C++ (from the Cap’n Proto mailing list; includes sample code)
RPC¶
Cap’n Proto has a rich RPC protocol. You should read the RPC specification as well as the C++ RPC documentation before using pycapnp’s RPC features. As with the serialization part of this library, the RPC component tries to be a very thin wrapper on top of the C++ API.
The examples below will be using calculator.capnp. Please refer to it to understand the interfaces that will be used.
Asyncio support was added to pycapnp in v1.0.0 utilizing the TwoWayPipe interface to libcapnp (instead of having libcapnp control the socket communication). The main advantage here is that standard Python socket libraries can be used with pycapnp (more importantly, TLS/SSL). Asyncio requires a bit more boiler plate to get started but it does allow for a lot more control than using the pycapnp socket wrapper.
Client¶
There are two ways to start a client: libcapnp socket wrapper and asyncio. The wrapper is easier to implement but is very limited (doesn’t support SSL/TLS with Python). asyncio requires more setup and can be harder to debug; however, it does support SSL/TLS and has more control over the socket error conditions. asyncio also helps get around the threading limitations around the current pycapnp implementation has with libcapnp (pycapnp objects and functions must all be in the same thread).
Starting a Client¶
Starting a client is very easy:
import capnp
import calculator_capnp
client = capnp.TwoPartyClient('localhost:60000')
Note
You can also pass a raw socket with a fileno() method to TwoPartyClient
Note
This will not work with SSL/TLS, please see Starting a Client (asyncio)
Starting a Client (asyncio)¶
Asyncio takes a bit more boilerplate than using the socket wrapper, but it gives you a lot more control. The example here is very simplistic. Here’s an example of full error handling (with reconnection on server failure): hidio client.
At a basic level, asyncio splits the input and output streams of the tcp socket and sends it to the libcapnp TwoWayPipe interface. An async reader Python function/method is used to consume the incoming byte stream and an async writer Python function/method is used to write outgoing bytes to the socket.
Note
You’ll need to be using the async keyword on some of the Python function/methods. If you’re unsure, look at the full example code. Also, read up on recent Python asyncio tutorials if you’re new to the concept. Make sure the tutorial is 3.7+, asyncio changed a lot from when it was first introduced in 3.4.
First you’ll need two basic async functions:
async def myreader(client, reader):
while True:
data = await reader.read(4096)
client.write(data)
async def mywriter(client, writer):
while True:
data = await client.read(4096)
writer.write(data.tobytes())
await writer.drain()
Note
There’s no socket error handling here, so this won’t be sufficient for anything beyond a simple example.
Next you’ll need to define an async function that sets up the socket connection. This is equivalent to client = capnp.TwoPartyClient(‘localhost:60000’) in the earlier example:
async def main(host):
addr = 'localhost'
port = '6000'
# Handle both IPv4 and IPv6 cases
try:
print("Try IPv4")
reader, writer = await asyncio.open_connection(
addr, port,
family=socket.AF_INET
)
except Exception:
print("Try IPv6")
reader, writer = await asyncio.open_connection(
addr, port,
family=socket.AF_INET6
)
# Start TwoPartyClient using TwoWayPipe (takes no arguments in this mode)
client = capnp.TwoPartyClient()
# Assemble reader and writer tasks, run in the background
coroutines = [myreader(client, reader), mywriter(client, writer)]
asyncio.gather(*coroutines, return_exceptions=True)
## Bootstrap Here ##
Note
On systems that have both IPv4 and IPv6 addresses, IPv6 is often resolved first and needs to be handled separately. If you’re certain IPv6 won’t be used, you can remove it (you should also avoid localhost, and stick to something like 127.0.0.1).
Finally, you’ll need to start the asyncio function:
if __name__ == '__main__':
asyncio.run(main(parse_args().host))
Note
This is the simplest way to start asyncio and usually not sufficient for most applications.
SSL/TLS Client¶
SSL/TLS setup effectively wraps the socket transport. You’ll need an SSL certificate, for this example we’ll be using a self-signed certificate. Most of the asyncio setup is the same as above:
async def main(host):
addr = 'localhost'
port = '6000'
# Setup SSL context
ctx = ssl.create_default_context(ssl.Purpose.SERVER_AUTH, cafile=os.path.join(this_dir, 'selfsigned.cert'))
# Handle both IPv4 and IPv6 cases
try:
print("Try IPv4")
reader, writer = await asyncio.open_connection(
addr, port,
ssl=ctx,
family=socket.AF_INET
)
except Exception:
print("Try IPv6")
reader, writer = await asyncio.open_connection(
addr, port,
ssl=ctx,
family=socket.AF_INET6
)
# Start TwoPartyClient using TwoWayPipe (takes no arguments in this mode)
client = capnp.TwoPartyClient()
# Assemble reader and writer tasks, run in the background
coroutines = [myreader(client, reader), mywriter(client, writer)]
asyncio.gather(*coroutines, return_exceptions=True)
## Bootstrap Here ##
Due to a bug in Python 3.7 and 3.8 asyncio client needs to be initialized in a slightly different way:
if __name__ == '__main__':
loop = asyncio.get_event_loop()
loop.run_until_complete(main(parse_args().host))
Bootstrap¶
Before calling any methods you’ll need to bootstrap the Calculator interface:
calculator = client.bootstrap().cast_as(calculator_capnp.Calculator)
There’s two things worth noting here. First, we are asking for the server capability. Secondly, you see that we are casting the capability that we receive. This is because capabilities are intrinsically dynamic, and they hold no run time type information, so we need to pick what interface to interpret them as.
Calling methods¶
There are 2 ways to call RPC methods. First the more verbose request syntax:
request = calculator.evaluate_request()
request.expression.literal = 123
eval_promise = request.send()
This creates a request for the method named ‘evaluate’, sets expression.literal in that call’s parameters to 123, and then sends the request and returns a promise (all non-blocking).
The shorter syntax for calling methods is:
eval_promise = calculator.evaluate({"literal": 123})
The major shortcoming with this method is that expressing complex fields with many nested sub-structs can become very tedious.
Once you have a promise, there are 2 ways of getting to the result. The first is to wait for it:
result = eval_promise.wait()
The second is to build a promise chain by calling then:
def do_stuff(val):
...
eval_promise.then(do_stuff).wait()
Pipelining¶
If a method returns values that are themselves capabilites, then you can access these fields before having to call wait. Doing this is called pipelining, and it allows Cap’n Proto to chain the calls without a round-trip occurring to the server:
# evaluate returns `value` which is itself an interface.
# You can call a new method on `value` without having to call wait first
read_promise = eval_promise.value.read()
read_result = read_promise.wait() # only 1 wait call
You can also chain promises with then and the same pipelining will occur:
read_result = eval_promise.then(lambda ret: ret.value.read()).wait()
Server¶
There are two ways to start a server: libcapnp socket wrapper and asyncio. The wrapper is easier to implement but is very limited (doesn’t support SSL/TLS with Python). asyncio requires more setup and can be harder to debug; however, it does support SSL/TLS and has more control over the socket error conditions. asyncio also helps get around the threading limitations around the current pycapnp implementation has with libcapnp (pycapnp objects and functions must all be in the same thread). The asyncio Server is a bit more work to implement than an asyncio client as more error handling is required to deal with client connection/disconnection/timeout events.
Starting a Server¶
To start a server:
server = capnp.TwoPartyServer('*:60000', bootstrap=CalculatorImpl())
server.run_forever()
Note
You can also pass a raw socket with a fileno() method to TwoPartyServer. In that case, run_forever will not work, and you will have to use on_disconnect.wait().
Note
This will not work with SSL/TLS, please see Starting a Server (asyncio)
Starting a Server (asyncio)¶
Like the asyncio client, an asyncio server takes a bunch of boilerplate as opposed to using the socket wrapper. Servers generally have to handle a lot more error conditions than clients so they are generally more complicated to implement with asyncio.
Just like the asyncio client, both the input and output socket streams are handled by reader/writer callback functions/methods.
Note
You’ll need to be using the async keyword on some of the Python function/methods. If you’re unsure, look at the full example code. Also, read up on recent Python asyncio tutorials if you’re new to the concept. Make sure the tutorial is 3.7+, asyncio changed a lot from when it was first introduced in 3.4.
To simplify the callbacks use a server class to define the reader/writer callbacks.:
class Server:
async def myreader(self):
while self.retry:
try:
# Must be a wait_for so we don't block on read()
data = await asyncio.wait_for(
self.reader.read(4096),
timeout=0.1
)
except asyncio.TimeoutError:
print("myreader timeout.")
continue
except Exception as err:
print("Unknown myreader err: %s", err)
return False
await self.server.write(data)
print("myreader done.")
return True
async def mywriter(self):
while self.retry:
try:
# Must be a wait_for so we don't block on read()
data = await asyncio.wait_for(
self.server.read(4096),
timeout=0.1
)
self.writer.write(data.tobytes())
except asyncio.TimeoutError:
print("mywriter timeout.")
continue
except Exception as err:
print("Unknown mywriter err: %s", err)
return False
print("mywriter done.")
return True
We need an additional myserver() method in the Server class to handle each of the incoming socket connections:
async def myserver(self, reader, writer):
# Start TwoPartyServer using TwoWayPipe (only requires bootstrap)
self.server = capnp.TwoPartyServer(bootstrap=CalculatorImpl())
self.reader = reader
self.writer = writer
self.retry = True
# Assemble reader and writer tasks, run in the background
coroutines = [self.myreader(), self.mywriter()]
tasks = asyncio.gather(*coroutines, return_exceptions=True)
while True:
self.server.poll_once()
# Check to see if reader has been sent an eof (disconnect)
if self.reader.at_eof():
self.retry = False
break
await asyncio.sleep(0.01)
# Make wait for reader/writer to finish (prevent possible resource leaks)
await tasks
Finally, we’ll need to start an asyncio server to spawn a new async myserver() with it’s own Server() object for each new connection:
async def new_connection(reader, writer):
server = Server()
await server.myserver(reader, writer)
async def main():
addr = 'localhost'
port = '60000'
# Handle both IPv4 and IPv6 cases
try:
print("Try IPv4")
server = await asyncio.start_server(
new_connection,
addr, port,
family=socket.AF_INET
)
except Exception:
print("Try IPv6")
server = await asyncio.start_server(
new_connection,
addr, port,
family=socket.AF_INET6
)
async with server:
await server.serve_forever()
if __name__ == '__main__':
asyncio.run(main())
Note
On systems that have both IPv4 and IPv6 addresses, IPv6 is often resolved first and needs to be handled separately. If you’re certain IPv6 won’t be used, you can remove it (you should also avoid localhost, and stick to something like 127.0.0.1). If you’re broadcasting in general, you’ll probably want to use 0.0.0.0 (IPv4) or ::/0 (IPv6).
SSL/TLS Server¶
Adding SSL/TLS support for a pycapnp asyncio server is fairly straight-forward. Just create an SSL context before starting the asyncio server:
async def main():
addr = 'localhost'
port = '60000'
# Setup SSL context
ctx = ssl.create_default_context(ssl.Purpose.CLIENT_AUTH)
ctx.load_cert_chain(os.path.join(this_dir, 'selfsigned.cert'), os.path.join(this_dir, 'selfsigned.key'))
# Handle both IPv4 and IPv6 cases
try:
print("Try IPv4")
server = await asyncio.start_server(
new_connection,
addr, port,
ssl=ctx,
family=socket.AF_INET
)
except Exception:
print("Try IPv6")
server = await asyncio.start_server(
new_connection,
addr, port,
ssl=ctx,
family=socket.AF_INET6
)
async with server:
await server.serve_forever()
Implementing a Server¶
Here’s a part of how you would implement a Calculator server:
class CalculatorImpl(calculator_capnp.Calculator.Server):
"Implementation of the Calculator Cap'n Proto interface."
def evaluate(self, expression, _context, **kwargs):
return evaluate_impl(expression).then(lambda value: setattr(_context.results, 'value', ValueImpl(value)))
def defFunction_context(self, context):
params = context.params
context.results.func = FunctionImpl(params.paramCount, params.body)
def getOperator(self, op, **kwargs):
return OperatorImpl(op)
Some major things worth noting.
You must inherit from your_module_capnp.YourInterface.Server, but don’t worry about calling __super__ in your __init__
Method names of your class must either match the interface exactly, or have ‘_context’ appended to it
If your method name is exactly the same as the interface, then you will be passed all the arguments from the interface as keyword arguments, so your argument names must match the interface spec exactly. You will also receive a _context parameter which is equivalent to the C++ API’s Context. I highly recommend having **kwargs as well, so that even if your interface spec is upgraded and arguments were added, your server will still operate fine.
Returns work with a bit of magic as well. If you return a promise, then it will be handled the same as if you returned a promise from a server method in the C++ API. Otherwise, your return statement will be filled into the results struct following the ordering in your spec, for example:
# capability.capnp file interface TestInterface { foo @0 (i :UInt32, j :Bool) -> (x: Text, i:UInt32); } # python code class TestInterface(capability_capnp.TestInterface.Server): def foo(self, i, j, **kwargs): return str(j), i
If your method ends in _context, then you will only be passed a context parameter. You will have to access params and set results yourself manually. Returning promises still works as above, but you can’t return anything else from a method.
Full Examples¶
Full examples are available on github. There is also an example of a very simplistic RPC available in test_rpc.py.