Open GoogleCodeExporter opened 8 years ago
Hi David,
Thanks for your feedback.
Perhaps the documentation isn't clear enough. The intent was that 'valid frame'
and 'unrecognized response' were two separate concepts. A 'valid frame' is a
binary API frame structure that passes check-summing, with no concern for the
data contained inside it.
An 'unrecognized response' occurs at a higher level. The API frame structure
was already declared valid, but its contents did not match any defined
response. Thus, the internal structure of the frame's contents could not be
parsed.
Since invalid frames could never be parsed correctly, they are silently
discarded in _wait_for_frame (xbee/base.py, near line 132). However, receiving
a valid, yet unrecognizable frame was deemed an error since it implies that
python-xbee was not working properly. In all probability, it should have been
able to parse the given response, but was unable to do so. The alternative of
silently ignoring expected, legal responses would prove very frustrating to
those who aren't aware of the behavior.
My apologies for the confusion; I hope this clarifies a few things. I would be
happy to accept your definitions of the missing response types; I'll try to
merge in your change later this evening.
Original comment by pmalms...@gmail.com
on 21 Jun 2011 at 5:15
Thanks for the clarification.
As things are now, code written according to the documented API is not
forwards-compatible and will crash when new response codes are introduced on
the network. Not all "response" frames are actually responses; some will be
unsolicited, like the routing frames, so it's entirely likely that a client
application will exist with no expectation that it's going to be receiving
these new frame types. Some alternatives to silently discarding unknown frames:
1. Log them with logging.error (and then discard them so that the application
won't crash)
2. Raise an exception only if the user requests so (e.g. a "be_tolerant" flag
in the initialization of the object)
3. Clearly document the exception as part of the API, so that users know to
catch it.
I suspect the general case for #3 will be that users will ignore it, which is
essentially what I optimized for in my suggested patch, but I understand that
some people might like to know that the module saw a frame that it didn't
understand.
I can try and put together a documentation fix if you think that's the best
path forward.
Original comment by david@fastolfe.net
on 21 Jun 2011 at 5:48
Note that this exception can also occur in the background/callback thread,
which causes the background thread to crash with no indication to the caller
and no way to restart it. The user code will just fail to ever see any new
events.
Original comment by david@fastolfe.net
on 22 Jun 2011 at 5:01
Those are good points; this definitely needs to be fixed.
I like all three of your suggestions. I'm not sure what the convention is for
using the logging module within library code; do you know?
My first thought would be to add two optional arguments to the XBeeBase
constructor: a logger object, and the 'raises exceptions' flag you suggested.
If a logger is provided, we use that, otherwise we create/use a default one.
The exceptions flag would be false by default.
For the background thread, I think we can just log and squash any/all
exceptions raised.
What do you think?
Original comment by pmalms...@gmail.com
on 22 Jun 2011 at 5:59
It would be great if at least David's ignore patch were integrated. The
current behavior is worse than silently ignoring unknown packet ids. It would
be nice to be able to use this otherwise fine library without having to patch
it first.
Original comment by ergo...@gmail.com
on 22 Jan 2013 at 7:48
Original issue reported on code.google.com by
david@fastolfe.net
on 21 Jun 2011 at 4:16