reinforcement_learning
                                
                                 reinforcement_learning copied to clipboard
                                
                                    reinforcement_learning copied to clipboard
                            
                            
                            
                        Fix wheels produced by CI
Currently wheel files produced by CI is not usable.
Tested on WSL2 Ubuntu 20.04 python 3.8.10
Running https://github.com/VowpalWabbit/reinforcement_learning/blob/master/examples/python/basic_usage.py hangs with the following outputs:
ataymano@ataymano-sb1:~/src/github/VowpalWabbit/reinforcement_learning/examples/python$ python3 basic_usage.py
event_id: 032575ce-fc7a-4837-a8bb-517012c980f8
model_id: N/A
chosen action id: 1
all action probabilities [(1, 0.5), (0, 0.5)]
event_id: c8cf584c-ec46-4639-8ca9-545a6836d3cc
model_id: N/A
chosen action id: 1
all action probabilities [(1, 0.5), (0, 0.5)]
event_id: 10223fb5-8987-4908-8378-181d657efc46
model_id: N/A
chosen action id: 1
all action probabilities [(1, 0.5), (0, 0.5)]
And there is no data on server side - probably something is wrong with cpprest.
I can try to take a look. Instead of CI, manual build should be working.
Some observations:
- Verified that building wheel locally works on WSL2 ubuntu 20.04
- Using the CI-produced wheel, the program hangs due to stuck background threads never joining (observed in gdb)
- auditwheeltries to package dependencies from a centos docker image, including- libssl, but it seems to have some binary incompatibility issues on ubuntu, causing the issue. I tried to replace some dependencies without much success.
I wonder if it is possible to not bundle libssl. We should most likely be depending on the system copy right?
yes agreed, so they can for example enjoy security updates.