Call a Python program that keeps running with PYPELINE()

Total.js 5 · total5@0.0.19 pythonpypelineintegrationsai Executed against Total.js 5

PYPELINE(name, [options]) starts pypelines/name.py with python3 and keeps a connection to it, so the application can use Python libraries (data, models, image processing) without starting a new process for every request. It returns an object with send(data) to send a message, on('data', fn) to receive the answers and close() to stop the script. A name that starts with ~ is the path of a file instead (~/opt/job.py), and a text that is code (it has line breaks or looks like Python) is run as an inline program with python3 -c.

The protocol. The script is started as python3 -u script.py <socket>: the last argument is the path of a Unix socket that the application listens on, and the script connects to it. Every message, in both directions, is a frame: 4 bytes with the length (big endian) followed by the body, which is JSON by default.
Options: type ('json' default, 'text', or anything else for raw Buffers), autorestart (false to not restart the script when it ends; by default it is started again), threads and socket: false (no socket, only start the program).
Messages sent before the script has connected wait and are delivered when it does.
Events: open (the script connected), data (a message), restart and close.
The path of the socket has a limit. The socket is a file in the tmp folder of the application and the system limits the whole path (about 104 characters on macOS and 108 on Linux). In a project that is deep in the disk the server fails with listen EINVAL, the script never connects and the messages wait forever. The path is the property pypeline.socket: set it right after PYPELINE() (the file is checked and the server started afterwards) to a short one, as the example does.
The Python program is a normal program that runs with the rights of the application: it is only as safe as the code of the file, and Python 3 has to be installed (python3 in the PATH).

pypelines/shout.py

import json
import socket
import struct
import sys

# The last argument is the socket that the application listens on
sock = socket.socket(socket.AF_UNIX, socket.SOCK_STREAM)
sock.connect(sys.argv[-1])

def read(count):
    data = b''
    while len(data) < count:
        chunk = sock.recv(count - len(data))
        if not chunk:
            raise SystemExit(0)
        data += chunk
    return data

while True:
    size = struct.unpack('>I', read(4))[0]
    message = json.loads(read(size).decode('utf-8'))
    answer = json.dumps({ 'id': message['id'], 'upper': message['text'].upper(), 'words': len(message['text'].split()) }).encode('utf-8')
    sock.sendall(struct.pack('>I', len(answer)) + answer)

controllers/py.js

exports.install = function() {
	ROUTE('POST /shout/ @json', shout);
};

const pypeline = PYPELINE('shout');

// A short path for the socket: the folder of a project can be too deep for a Unix socket
pypeline.socket = require('os').tmpdir() + '/forge-shout-' + process.pid + '.socket';
const pending = {};
let counter = 0;

// Every answer says which message it is for
pypeline.on('data', function(answer) {
	const callback = pending[answer.id];
	if (callback) {
		delete pending[answer.id];
		callback(answer);
	}
});

function shout($) {
	const id = ++counter;
	const timer = setTimeout(() => $.invalid('The script did not answer'), 5000);
	pending[id] = function(answer) {
		clearTimeout(timer);
		$.json(answer);
	};
	pypeline.send({ id: id, text: $.body.text });
}