''' Server Central management object for both file serving systems (static server, live reloading) and job execution (routing and listening). Routers and Listeners can be started and managed independently, but a single Server instance can house, start, and shutdown listeners in one place. TODO: as it stands, the Server requires address and port details, effectively needing one of the HTTP items (static file serving or livereloading) to be initialized appropriately. But there is a clear use case for just managing disparate Routers and their associated Listeners. Should perhaps separate this "grouped listener" into another object, or just make the Server definition more flexible. ''' import re import asyncio import logging import threading from functools import partial from contextlib import asynccontextmanager import uvicorn from inotify_simple import flags from fastapi import FastAPI, WebSocket from fastapi.staticfiles import StaticFiles from execlog.routers.path import PathRouter from execlog.handler import Handler as LREndpoint logger = logging.getLogger(__name__) class Server: ''' Server class. Wraps up a development static file server and live reloader. ''' def __init__( self, host, port, root, static : bool = False, livereload : bool = False, managed_listeners : list | None = None, ): ''' Parameters: host: host server address (either 0.0.0.0, 127.0.0.1, localhost) port: port at which to start the server root: base path for static files _and_ where router bases are attached (i.e., when files at this path change, a reload event will be propagated to a corresponding client page) static: whether or not to start a static file server livereload: whether or not to start a livereload server managed_listeners: auxiliary listeners to "attach" to the server process, and to propagate the shutdown signal to when the server receives an interrupt. ''' self.host = host self.port = port self.root = root self.static = static self.livereload = livereload if managed_listeners is None: managed_listeners = [] self.managed_listeners = managed_listeners self.listener = None self.userver = None self.server = None self.server_text = '' self.server_args = {} self.loop = None self._server_setup() def _wrap_static(self): self.server.mount("/", StaticFiles(directory=self.root), name="static") def _wrap_livereload(self): self.server.websocket_route('/livereload')(LREndpoint) #self.server.add_api_websocket_route('/livereload', LREndpoint) def _server_setup(self): ''' Set up the FastAPI server. Only a single server instance is used here, optionally mounting the static route (if static serving enabled) and providing a websocket endpoint (if livereload enabled). Note that, when present, the livereload endpoint is registered first, as the order in which routes are defined matters for FastAPI apps. This allows `/livereload` to behave appropriately, even when remounting the root if serving static files (which, if done in the opposite order, would "eat up" the `/livereload` endpoint). ''' # enable propagation and clear handlers for uvicorn internal loggers; # allows logging messages to propagate to my root logger log_config = uvicorn.config.LOGGING_CONFIG log_config['loggers']['uvicorn']['propagate'] = True log_config['loggers']['uvicorn']['handlers'] = [] log_config['loggers']['uvicorn.access']['propagate'] = True log_config['loggers']['uvicorn.access']['handlers'] = [] log_config['loggers']['uvicorn.error']['propagate'] = False log_config['loggers']['uvicorn.error']['handlers'] = [] self.server_args['log_config'] = log_config self.server_args['host'] = self.host self.server_args['port'] = self.port @asynccontextmanager async def lifespan(app: FastAPI): yield self.shutdown() if self.static or self.livereload: self.server = FastAPI(lifespan=lifespan) #self.server.on_event('shutdown')(self.shutdown) if self.livereload: self._wrap_livereload() self._listener_setup() self.server_text += '+reload' if self.static: self._wrap_static() self.server_text += '+static' def _listener_setup(self): ''' flags.MODIFY okay since we don't need to reload non-existent pages ''' if self.loop is None: self.loop = asyncio.new_event_loop() asyncio.set_event_loop(self.loop) #self.listener = listener.WatchFS(loop=self.loop) self.router = PathRouter(loop=self.loop) self.router.register( path=str(self.root), func=LREndpoint.reload_clients, delay=100, flags=flags.MODIFY, ) self.listener = self.router.get_listener() self.managed_listeners.append(self.listener) def start(self): ''' Start the server. Note: Design This method takes on some extra complexity in order to ensure the blocking Watcher and FastAPI's event loop play nicely together. The Watcher's `start()` method runs a blocking call to INotify's `read()`, which obviously cannot be started directly here in the main thread. Here we have a few viable options: 1. Simply wrap the Watcher's `start` call in a separate thread, e.g., ```py watcher_start = partial(self.watcher.start, loop=loop) threading.Thread(target=self.watcher.start, kwargs={'loop': loop}).start() ``` This works just fine, and the watcher's registered async callbacks can still use the passed event loop to get messages sent back to open WebSocket clients. 2. Run the Watcher's `start` inside a thread managed by event loop via `loop.run_in_executor`: ```py loop.run_in_executor(None, partial(self.watcher.start, loop=loop)) ``` Given that this just runs the target method in a separate thread, it's very similar to option #1. It doesn't even make the outer loop context available to the Watcher, meaning we still have to pass this loop explicitly to the `start` method. The only benefit here (I think? there may actually be no difference) is that it keeps things under one loop, which can be beneficial for shutdown. See related discussions: - https://stackoverflow.com/questions/55027940/is-run-in-executor-optimized-for-running-in-a-loop-with-coroutines - https://stackoverflow.com/questions/70459437/how-gil-affects-python-asyncio-run-in-executor-with-i-o-bound-tasks Once the watcher is started, we can kick off the FastAPI server (which may be serving static files, handling livereload WS connections, or both). We provide `uvicorn` access to the manually created `asyncio` loop used to the run the Watcher (in a thread, that is), since that loop is made available to the `Watcher._event_loop` method. This ultimately allows async methods to be registered as callbacks to the Watcher and be ran in a managed loop. In this case, that loop is managed by FastAPI, which keeps things consistent: the Watcher can call `loop.call_soon_threadsafe` to queue up a FastAPI-based response _in the same FastAPI event loop_, despite the trigger for that response having originated from a separate thread (i.e., where the watcher is started). This works smoothly, and keeps the primary server's event loop from being blocked. Note that, due to the delicate Watcher behavior, we must perform a shutdown explicitly in order for things to be handled gracefully. This is done in the server setup step, where we ensure FastAPI calls `watcher.stop()` during its shutdown process. Note: on event loop management The uvicorn server is ran with `run_until_complete`, intended as a long-running process to eventually be interrupted or manually disrupted with a call to `shutdown()`. The `shutdown` call attempts to gracefully shutdown the uvicorn process by setting a `should_exit` flag. Upon successful shutdown, the server task will be considered complete, and we can then manually close the loop following the interruption. So a shutdown call (which is also attached as a lifespan shutdown callback for the FastAPI object) will disable listeners and shut down their thread pools, gracefully close up the Uvicorn server and allow the serve coroutine to complete, and finally close down the event loop. ''' if self.loop is None: self.loop = asyncio.new_event_loop() asyncio.set_event_loop(self.loop) for listener in self.managed_listeners: #loop.run_in_executor(None, partial(self.listener.start, loop=loop)) if not listener.started: listener.start() if self.server: logger.info(f'Server{self.server_text} @ http://{self.host}:{self.port}') uconfig = uvicorn.Config(app=self.server, loop=self.loop, **self.server_args) self.userver = uvicorn.Server(config=uconfig) self.loop.run_until_complete(self.userver.serve()) self.loop.close() def shutdown(self): ''' Additional shutdown handling after the FastAPI event loop receives an interrupt. This is attached as a "shutdown" callback when creating the FastAPI instance, which generally appears to hear interrupts and propagate them through. This method can also be invoked programmatically, such as from a thread not handling the main event loop. Note that either of the following shutdown approaches of the Uvicorn server do not appear to work well in this case; they both stall the calling thread indefinitely (in the second case, when waiting on the shutdown result), or simply don't shutdown the server (in the first). Only setting `should_exit` and allowing for a graceful internal shutdown appears to both 1) handle this gracefully, and 2) shut down the server at all. ``` self.loop.call_soon_threadsafe(self.userver.shutdown) # OR # future = asyncio.run_coroutine_threadsafe(self.userver.shutdown(), self.loop) # and wait for shutdown future.result() ``` The shutdown process goes as follows: 1. Stop any managed listeners: close out listener loops and/or thread pools by calling `stop()` on each of the managed listeners. This prioritizes their closure so that no events can make their way into the queue. 2. Gracefully shut down the wrapper Uvicorn server. This is the process that starts the FastAPI server instance; set the `should_exit` flag. If this completes successfully, in the thread where Uvicorn was started the server task should be considered "completed," at which point the event loop can be closed successfully. ''' logger.info("Shutting down server...") # stop attached auxiliary listeners, both internal & external if self.managed_listeners: logger.info(f"Stopping {len(self.managed_listeners)} listeners...") for listener in self.managed_listeners: listener.stop() # stop FastAPI server if started if self.userver is not None: def set_should_exit(): self.userver.should_exit = True self.loop.call_soon_threadsafe(set_should_exit)