Closed rpdrewes closed 4 years ago
Thinking this might be a duplicate of issue #59. Testing.
Yes indeed forcing tornado 4.3 in the docker image fixed it. Thanks @craklyn and @phiresky, and of course @jcsilva.
I do wonder what the root of the problem is ... why is forcing tornado 4.3 in the docker image necessary for things to work properly? I had a vague feeling it was something about EOS message being interpreted as unicode but that is just a wild guess ...
Thanks for your work, jcsilva, and your excellent instructions in the README.md. I have the recognizer working, sort of: the recognition is performed correctly but a result is not returned to the client, and if I kill the client and retry, there seems to be something wrong with the server and subsequent attempts fail until the server is restarted.
My test client is the client.py referenced in the README.md. I have the suggested ws4py version (0.3.2). I have also tried Konele client with similar results.
I am running both worker and master in the same docker, as shown in the Practical Example portion of the README.md. I just pulled the docker image yesterday. I am running on Ubuntu 19.10. I get similar behavior when I run worker and master in separate docker instances. I get similar behavior when my client is on the same physical machine as the worker/master and also when the client is on another machine on the local network.
In detail, when I invoke client.py with an audio test file, the end of the worker.log looks like this:
The audio is correctly recognized ("here is a more complicated audio sample to try to recognize"). But something seems to go wrong at the end of things. No result seems to make it back to the client. The client hangs like this:
During this period the master log looks like this:
If I kill the client (control-C does nothing; I have to control-Z and then kill the process manually), then the worker sees the connection drop:
But the next time I try the client, the worker only does this:
And the client hangs as above. The master log looks like this during this period:
The same thing happens when I use the test audio files (bill_gates-TED.mp3, 1272-128104-0000.wav) at the end of the README.md: the recognition is performed once, but no result is returned to the client (the client simply hangs) and subsequent submissions from client fail until the server is stopped and restarted.
Any suggestion for how to most quickly find the problem?