Release Notes¶
0.9.2¶
- Consistent error message in py2 and py3 for error in Timestamp decode
0.9.1¶
- Fix issues in KATCPReply __repr__ in py3
0.9.0¶
- Add asyncio compatible ioloop to ioloop manager.
0.8.0¶
- Added bulk sensor sampling feature.
0.7.2¶
- Support for handling generator expressions in
Discrete
type. - Fix handling of strings and bytes in
get_sensor
in testutils. - Allow strings or bytes for
assert_request_fails
andtest_assert_request_succeeds
function arguments. - Handle
str
type correctly (‘easier’) intestutils.get_sensor
for python 2 and python 3. - Allow bytes and strings in
test_sensor_list
comparison of sensors. - Correct handling of floats
test_sensor_list
. - black formatting on certain test files.
0.7.1¶
- All params in
future_get_sensor
are now cast to byte strings. - Added tests to
test_fake_clients.py
andtest_inspecting_client.py
. - Ensure
testutils
method casts expected requests to byte strings.
0.7.0¶
- Added Python 3 compatibility.
See also CHANGELOG.md
for more details on changes.
Important changes for Python 3 compatibility¶
General notes¶
The package is now compatible with both Python 2 and 3. The goals of the migration were:
- Do not change the public API.
- Do not break existing functionality for Python 2.
- Ease migration of packages using katcp to Python 3.
Despite these goals, some of the stricter type checking that has been added may force minor updates in existing code. E.g., using integer for the options of a discrete sensor is no longer allowed.
Asynchronous code is still using tornado in the same Python 2 way. The new
Python 3.5 async
and await
keywords are not used. The tornado version
is also pinned to older versions that support both Python 2 and 3. The 5.x
versions also support Python 2, but they are avoided as some significant
changes result in test failures.
The Python future
package was used for the compatibility layer. The use of
the newstr
and newbytes
compatibility types was avoided, to reduce
confusion. I.e., from builtins import str, bytes
is not done.
Docstrings¶
In docstrings the interpretation of parameter and return types described
as “str” has changed slightly. In Python 2 the str
type is a byte
string, while in Python 3, str
is a unicode string. The str
type
is referred to as the “native” string type. In code, native literal strings
would have no prefix, for example: "native string"
, as opposed to
explicit byte strings, b"byte string"
, and explicit unicode strings,
u"unicode string"
. In the docstrings “bytes” means a byte string is
expected (or returned), “str” means a native string, and “str or bytes”
means either type.
Changes to types¶
As part of the Python 3 compatibility update, note the following:
katcp.Message
. -arguments
andmid
attributes will be forced to byte strings in all Python versions. This is to match what is sent on the wire (serialised byte stream). -name
: is expected to be a native string. -repr()
: the result will differ slightly in Python 3 - the arguments will be shown as quoted byte strings. E.g., Python 2:"<Message reply ok (123, zzz)>"
, vs. Python 3:"<Message reply ok (b'123', b'zzz')>"
. In all versions, arguments longer than 1000 characters are now truncated.katcp.Sensor
. -name
,description
,units
,params
(for discrete sensors):__init__
can take byte strings or native strings, but attributes will be coerced to native strings. -set_formatted
,parse_value
: only accept byte strings (stricter checking). - Thefloat
andstrict_timestamp
sensor values are now encoded usingrepr()
instead of"%.15g"
. This means that more significant digits are transmitted on the wire (16 to 17, instead of 15), and the client will be able to reconstruct the exact some floating point value.
Non-ASCII and UTF-8¶
Prior to these changes, all strings were byte strings, so there was no encoding
required. Arbitrary bytes could be used for message parameters and string
sensor values. After these changes, strings sensors and Str
types are
considered “text”. In Python 3, UTF-8 encoding will be used when changing
between byte strings and unicode strings for “text”. This has the following
effects:
katcp.Message
- thearguments
are always using byte strings, so arbitrary bytes can still be sent and received using this class directly.katcp.Sensor
- Values forstring
anddiscrete
sensor types cannot be arbitrary byte strings in Python 3 - they need to be UTF-8 compatible.kattypes.Str
,kattypes.Discrete
,kattypes.DiscreteMulti
- These types is still used inrequest
andreply
decorators. - For sending messages, they accept any type of object, but UTF-8 encoding is used if values are not already byte strings. - When decoding received messages, “native” strings are returned.
Keep in mind that a Python 2 server may be communicating with a Python 3 client, so sticking to ASCII is safest. If you are sure both client and server are on Python 3 (or understand the encoding the same), then UTF-8 could be used. That is also the encoding option used by the aiokatcp package.
Performance degradation¶
Adding the compatibility results in more checks and conversions. From some basic benchmarking, there appears to be up to 20% performance degradation when instantiating message objects.
Benchmark, in ipython:
import random, katcp
args_groups = []
for i in range(1000):
args_groups.append((random.randint(0, 1) == 1,
random.randint(0, 1000),
random.random(),
str(random.random())))
def benchmark():
for args in args_groups:
tx_msg = katcp.Message.reply('foo', *args)
serialised = bytes(tx_msg)
parser = katcp.MessageParser()
rx_msg = parser.parse(serialised)
assert tx_msg == rx_msg
%timeit benchmark()
- Old Py2: 10 loops, best of 3: 23.4 ms per loop
- New Py2: 10 loops, best of 3: 29.9 ms per loop
- New Py3: 25.1 ms ± 86.8 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)
0.6.4¶
- Fix some client memory leaks, and add until_stopped methods.
- Increase server MAX_QUEUE_SIZE to handle more clients.
- Use correct ioloop for client AsyncEvent objects.
See also CHANGELOG.md
for more details on changes.
Important API changes¶
Stopping KATCP clients¶
When stopping KATCP client classes that use a managed ioloop (i.e., create their
own in a new thread), the traditional semantics are to call stop()
followed by
join()
from another thread. This is unchanged. In the case of an unmanaged
ioloop (i.e., an existing ioloop instance is provided to the client), we typically
stop from the same thread, and calling join()
does nothing. For the case of
unmanaged ioloops, a new method, until_stopped()
, has been added. It returns a
future that resolves when the client has stopped. The caller can yield
on this
future to be sure that the client has completed all its coroutines. Using this new
method is not required. If the ioloop will keep running, the stopped client’s
coroutines will eventually exit. However, it is useful in some cases, e.g., to
verify correct clean up in unit tests.
The new method is available on katcp.DeviceClient
and derived classes, on
katcp.inspecting_client.InspectingClientAsync
, and on the high-level
clients katcp.KATCPClientResource
and
katcp.KATCPClientResourceContainer
.
An additional change is that the inspecting client now sends a state update
(indicating that it is disconnected and not synced) when stopping. This means
high-level clients that were waiting on until_not_synced
when the client was
stopped will now be notified. Previously, this was not the case.
0.6.3¶
- Put docs on readthedocs.
- Better error handling for messages with non-ASCII characters (invalid).
- Increase container sync time to better support large containers.
- Limit tornado version to <5.
- Allow sampling strategy to be removed from cache.
- Improve error messages for DeviceMetaClass assertions.
- Increase server’s message queue length handle more simultaneous client connections.
- Improve Jenkins pipeline configuration.
- Add information on how to contribute to the project.
See also CHANGELOG.md
for more details on changes.
0.6.2¶
- Various bug fixes
- Docstring and code style improvements
- Bumped the tornado dependency to at least 4.3
- Added the ability to let ClientGroup wait for a quorum of clients
- Added default request-timeout-hint implementation to server.py
- Moved IOLoopThreadWrapper to ioloop_manager.py, a more sensible location
- Added a random-exponential retry backoff process
See also CHANGELOG.md
for more details on changes.
0.6.1¶
- Various bug fixes
- Improvements to testing utilities
- Improvements to various docstrings
- Use katversion to determine version string on install
- Better dependency management using setup.py with setuptools
- Fixed a memory leak when using KATCPResourceContainer
See also CHANGELOG.md
for more details on changes.
0.6.0¶
- Major change: Use the tornado event loop and async socket routines.
See also CHANGELOG.md
for more details on changes.
Important API changes¶
Tornado based event loop(s)¶
While the networking stack and event loops have been re-implemented using
Tornado, this change should be largely invisible to existing users of the
library. All client and server classes now expose an ioloop attribute that is
the tornado.ioloop.IOLoop
instance being used. Unless new server or
client classes are used or default settings are changed, the thread-safety and
concurrency semantics of 0.5.x versions should be retained. User code that made
use of non-public interfaces may run into trouble.
High level auto-inspecting KATCP client APIs added¶
The high level client API inspects a KATCP device server and present requests as method calls and sensors as objects. See Using the high-level client API.
Sensor observer API¶
The katcp.Sensor
sensor observer API has been changed to pass the
sensor reading in the observer.update() callback, preventing potential lost
updates due to race conditions. This is a backwards incompatible change.
Whereas before observers were called as observer.update(sensor), they are now
called as observer.update(sensor, reading), where reading is an instance of
katcp.core.Reading
.
Sample Strategy callback API¶
Sensor strategies now call back with the sensor object and raw Python datatype
values rather than the sensor name and KATCP formatted values. The sensor
classes have also grown a katcp.Sensor.format_reading()
method that
can be used to do KATCP-version specific formatting of the sensor reading.
0.5.5¶
- Various cleanups (logging, docstrings, base request set, minor refactoring)
- Improvements to testing utilities
- Convenience utility functions in katcp.version, katcp.client, katcp.testutils.
0.5.4¶
- Change event-rate strategy to always send an update if the sensor has changed and shortest-period has passed.
- Add differential-rate strategy.
0.5.3¶
Add convert_seconds()
method to katcp client classes that converts seconds
into the device timestamp format.
0.5.2¶
Fix memory leak in sample reactor, other minor fixes.
0.5.1¶
Minor bugfixes and stability improvements
0.5.0¶
First stable release supporting (a subset of) KATCP v5. No updates apart from documentation since 0.5.0a0; please refer to the 0.5.0a release notes below.
0.5.0a0¶
First alpha release supporting (a subset of) KATCP v5. The KATCP v5 spec brings a number of backward incompatible changes, and hence requires care. This library implements support for both KATCP v5 and for the older dialect. Some API changes have also been made, mainly in aid of fool-proof support of the Message ID feature of KATCP v5. The changes do, however, also eliminate a category of potential bugs for older versions of the spec.
Important API changes¶
CallbackClient.request()¶
Renamed request()
to callback_request()
to be more consistent with superclass
API.
Sending replies and informs in server request handlers¶
The function signature used for request handler methods in previous versions of
this library were request_requestname(self, sock, msg), where sock is a
raw python socket object and msg is a katcp Message
object. The sock
object was never used directly by the request handler, but was passed to methods
on the server to send inform or reply messages.
Before:
class MyServer(DeviceServer):
def request_echo(self, sock, msg):
self.inform(sock, Message.inform('echo', len(msg.arguments)))
return Message.reply('echo', 'ok', *msg.arguments)
The old method requires the name of the request to be repeated several times,
inviting error and cluttering code. The user is also required to instantiate
katcp Message
object each time a reply is made. The new method passes a
request-bound connection object that knows to what request it is replying, and
that automatically constructs Message
objects.
Now:
class MyServer(DeviceServer):
def request_echo(self, req, msg):
req.inform(len(msg.arguments)))
return req.make_reply('ok', *msg.arguments)
A req.reply()
method with the same signature as req.make_reply()
is also available for asyncronous reply handlers, and
req.reply_with_message()
which takes a Message
instance rather
than message arguments. These methods replace the use of
DeviceServer.reply()
.
The request object also contains the katcp request Message
object
(req.msg), and the equivalent of a socket object
(req.client_connection). See the next section for a description of
client_connection.
Using the server methods with a req object in place of sock will still work as before, but will log deprecation warnings.
Connection abstraction¶
Previously, the server classes internally used each connection’s low-level
sock object as an identifier for the connection. In the interest of
abstracting out the transport backend, the sock object has been replaced by a
ClientConnectionTCP
object. This object is passed to all server handler
functions (apart from request handlers) instead of the sock object. The
connection object be used in the same places where sock was previously
used. It also defines inform()
, reply_inform()
and reply()
methods for sending Message
objects to a client.
Backwards incompatible KATCP V5 changes¶
Timestamps¶
Excerpted from NRF-KAT7-6.0-IFCE-002-Rev5.pdf
:
All core messages involving time (i.e. timestamp or period specifications) have changed from using milliseconds to seconds. This provides consistency with SI units. Note also that from version five timestamps should always be specified in UTC time.
Message Identifiers (mid)¶
Excerpted from NRF-KAT7-6.0-IFCE-002-Rev5.pdf
:
Message identifiers were introduced in version 5 of the protocol to allow replies to be uniquely associated with a particular request. If a client sends a request with a message identifier the server must include the same identifier in the reply. Message identifiers are limited to integers in the range 1 to 231 − 1 inclusive. It is the client’s job to construct suitable identifiers – a server should not assume that these are unique. Clients that need to determine whether a server supports message identifiers should examine the #version-connect message returned by the server when the client connects (see Section 4). If no #version-connect message is received the client may assume message identifiers are not supported.
also:
If the request contained a message id each inform that forms part of the response should be marked with the original message id.
Support for message IDs is optional. A properly implemented server should never use mids in replies unless the client request has an mid. Similarly, a client should be able to detect whether a server supports MIDs by checking the #version-connect informs sent by the server, or by doing a !version-list request. Furthermore, a KATCP v5 server should never send #build-state or #version informs.
Server KATCP Version Auto-detection¶
The DeviceClient
client uses the presence of
#build-state or #version informs as a heuristic to detect pre-v5 servers,
and the presence of #version-connect informs to detect v5+ servers. If mixed
messages are received the client gives up auto-detection and disconnects. In
this case preset_protocol_flags()
can be used to
configure the client before calling start()
.
Level of KATCP support in this release¶
This release implements the majority of the KATCP v5 spec; excluded parts are:
- Support for optional warning/error range meta-information on sensors.
- Differential-rate sensor strategy.