`loader.py`: use `urllib3`, support custom timeout about kivy

[az75014]

Sure! To prove to myself that adding pooling would make the difference for me I created a MyAsyncImage/myasyncimage.py so that I could make it load my hacky MyLoader/myloader.py rather than the Kivy code one. It was basically copy/paste, apart from the changes indicated below.

This was just a test, I'm sure it needs more thought than this.

...
import urllib3
pool_mgr = urllib3.PoolManager(maxsize=10)
"""
Note: I set maxsize to 10 because I was getting warnings due to too many concurrent requests:
[WARNING] [Connection pool is full, discarding connection]
I tested this with a value of 1000 also with seemingly no ill effect. 
Maybe it should match the Loader.num_workers value? 
It's set to 10 in my case as the images I am loading are rather small.
I left num_pools unchanged as not relevant in my case: there was only ever going to be 1 pool,
namely the connection to my image server. (default is 10)
"""
Cache.register('kv.loader', limit=500, timeout=500)
...
    def _load_urllib(self, filename, kwargs):
        '''(internal) Loading a network file. First download it, save it to a
        temporary file, and pass it to _load_local().'''	
        import tempfile
        global pool_mgr
        if not pool_mgr:
            pool_mgr = urllib3.PoolManager(maxsize=10)
...
        try:
            _out_filename = ''
            """
            Note: in my hacked version I removed a whole section here about smb, 
            headers, custom context as not relevant to my test
            """
            fd = pool_mgr.request("GET", filename)

            if '#.' in filename:
                # allow extension override from URL fragment
                suffix = '.' + filename.split('#.')[-1]
            else:
                ctype = fd.headers['Content-Type']
...
            idata = fd.data
            # removed: fd.close()
            # removed: fd = None
...

[Julian-O]

I am working on a refactor of loader.py in the background, so listening with interest. Do you want to share the changes you tried?

[NomadDemon]

And how it behave when you change default network backend to "requests"? Its single line in config file

Can someone benchmark it?

[az75014]

The loader.py currently doesn't use the config file settings (apart from user agent) -- urlib.request is hardcoded.

There is no session mechanism for the "requests" implementation, so that shouldn't make a difference?

On my app I make several requests to my API on start up, I had a huge efficiency boost after I implemented sessions for that.

Currently, I create a session in my App class, save it as a class attribute and then pass the session object as an argument to e.g. my API request module. It works fine but seems a bit hacky to do it this way.

There might be scope to handle this better?

[NomadDemon]

The loader.py currently doesn't use the config file settings (apart from user agent) -- urlib.request is hardcoded.

There is no session mechanism for the "requests" implementation, so that shouldn't make a difference?

On my app I make several requests to my API on start up, I had a huge efficiency boost after I implemented sessions for that.

Currently, I create a session in my App class, save it as a class attribute and then pass the session object as an argument to e.g. my API request module. It works fine but seems a bit hacky to do it this way.

There might be scope to handle this better?

it uses

Try

Config.set("network", "implementation", "requests")

it should work from kivy 2.2.0

kivy/kivy/network/urlrequest.py

Line 781 in 340ea0d

implementation_map = {