picows websocket client

[Tapanhaz]

I tried implementing picows for connecting to a broker websocket. https://github.com/Tapanhaz/Shoonya-Websocket

[tarasko]

Thank you for trying out picows!

I quickly ran through the code in on_ws_frame and noticed a couple of thing that you maybe would find interesting to change:

• I would advice to add

if frame.msg_type == WSMsgType.PING:
     transport.send_pong(frame.get_payload_as_bytes())

to make sure you client always replies to ping messages.

• Also, it is likely that Shoonya never fragments messages, I believe it is very common not to fragment messages for brokers. It does not have much practical benefits. It only make sense to have this logic

if frame.fin:
    ...
else:
    ...

if you have actually observed non fin messages. Imho, it may actually suffice to just have this instead:

def on_ws_frame(self, transport, frame):
    assert frame.fin, "unexpected fragmented websocket message from Shoonya"
    if frame.msg_type == WSMsgType.PING:
         transport.send_pong(frame.get_payload_as_bytes())
    elif frame.msg_type == WSMsgType.CLOSE:
         # It is good to maybe log close reason and close message here and disconnect
         # logger.info("Shoonya disconnected, code=%s, reason=%s, frame.get_close_code(), frame.get_close_message().decode())
         transport.disconnect()
    else:
        assert frame.msg_type == WSMsgType.TEXT, f"Shoonya is expected to send text messages, instead received {frame.msg_type}"
        msg = frame.get_payload_as_utf8_text()
        self.parent.on_data_callback(msg)

Notice that I also used get_payload_as_utf8_text instead of get_payload_as_bytes, as it is faster than first getting bytes and then decoding them with decode('utf8')

I wish all the best with your project!

[Tapanhaz]

I was wrong. The server sends pong .., modified my code. Maybe I made some huge mistakes there. With low number of subscriptions it performs quite well but when token numbers increased, it is lagging. i checked with websocket-client (running in asyncio.to_thread) side by side, put a counter in on_ticks to check the number of ticks it gets per sec / total tokens subscribed to check average ticks per sec (both with same sets of tokens) :


I am trying to debug at my end but not able to get success till now....

[tarasko]

Without profiling it is difficult to say where the bottleneck is, have you tried running it with cProfile?

Regarding running code in asyncio.to_thread, it most likely wouldn't improve things because of GIL. I would expect it to actually slow down everything unless you are doing some heavy slow IO there.

Note Due to the GIL, asyncio.to_thread() can typically only be used to make IO-bound functions non-blocking. However, for extension modules that release the GIL or alternative Python implementations that don’t have one, asyncio.to_thread() can also be used for CPU-bound functions.

One thing that I noticed, that may help a little bit in ShoonyaClient
Instead of

msg = json.loads(msg)

Create json decoder once and reuse it

# Somewhere
json_decoder = json.JSONDecoder()
...
# In on_data_callback
        try:
            msg = json_decoder.loads(msg)

But again without profiling it is not possible to say where the bottleneck is

[Tapanhaz]

I profile it with cProfile and scalene and didn't find anything other than your mentioned line . scalene detects ssl_context.verify_mode = ssl.CERT_REQUIRED and another line in on_ws_frame which I forget to comment out . But as the market is closed as of now, I will paste here the live profiling report on monday. Also I am thinking of replacing json with orjson.

[Tapanhaz]

This is the cprofile result ( running with 252 tokens subscribed for 60 seconds): Please remove the .txt extension

shoonya_ticker.prof.txt

[tarasko]

Hm, I don't see anything suspicious in the profiler result.

Does ticker messages arrive at constant rate or are they only coming when trade tick happens? Could it be that there was simply not much messages from shoonya during those 60 seconds?

I saw that on_ws_frame was called 1582 times during 60 seconds, which suggest 26 messages (ticks) per seconds, which is not 0.09 but it is a very small number anyway.
Just to give some numbers, I'm running a single threaded app where I connect to a few websockets via picows and sometimes it is more than 10000-20000 messages per second.

During those 60 seconds was CPU heavily consumed by the client?

Do you use threads/locks in the client? You mentioned earlier asyncio.to_thread, would it be possible to see what do main and __subscribe_callback provided to ShoonyaTicker do?

I'm asking because I saw

call_soon_threadsafe

which suggesting that threading is somehow involved.

I'd guess it is either not much messages from the broker or a deadlock somewhere, probably related to multithreading and synchronization primitives

[Tapanhaz]

Extremely sorry, I made a huge mistake in subscription :

Sorry again for the trouble I created because of own silly mistake..

[Tapanhaz]

This is scalene report now :

and this is cprofile result :

shoonya_ticker_new.prof.txt

[Tapanhaz]

with orjson :

[tarasko]

Cheers! A good luck with your library :)

[Tapanhaz]

On network disconnection, it behaves differently in windows and linux. In windows (python 3.11 and 3.12), It disconnects immediately on network disconnects but in linux (Ubuntu 24.04 LTS, python 3.12.3 and 3.12.6) it shows a indefinite behavior. So I make a loop checking for pong after each ping with a timeout of 20 secs:

If not received, It will disconnects itself and the retry mechanism will try to reconnect it with a gap of 1 sec:

But transport.disconnect() takes a significant amount of time in between:

[Tapanhaz]

Thanks it is working...

I really loved current low level minimalistic interface, User gets more fine grained control over each thing.

[tarasko]

Awesome! Thanks for the testing!

[tarasko]

Hi @Tapanhaz ,
I recently added an automatic ping-pong mechanism for detecting stale connections. This might help to simplify your code a bit.

https://picows.readthedocs.io/en/stable/guides.html#auto-ping-pong

[Tapanhaz]

Thanks for the update. That sounds great !! This will definitely help simplify things on my end.

[Tapanhaz]

I recently tried implementing auto ping in another repo : https://github.com/Tapanhaz/Kite-Ticker . This is my first hands on cython too..