From c7e835656abeabac81fe3b8cddd70343df0c4eba Mon Sep 17 00:00:00 2001
From: Andrew Svetlov <andrew.svetlov@gmail.com>
Date: Mon, 11 Sep 2017 11:11:38 +0300
Subject: [PATCH] Document web handlers cancellation (#2257)

* Sketch for webhandler cancellation docs

* Temp fix

* Finish a chapter about tasks cancellation
---
 docs/web.rst | 96 +++++++++++++++++++++++++++++++++++++++++++++++++++-
 1 file changed, 95 insertions(+), 1 deletion(-)

diff --git a/docs/web.rst b/docs/web.rst
index 5a4761d7dd9..a1710420877 100644
--- a/docs/web.rst
+++ b/docs/web.rst
@@ -320,7 +320,6 @@ viewed using the :meth:`UrlDispatcher.named_resources` method::
    :meth:`UrlDispatcher.resources` instead of
    :meth:`UrlDispatcher.named_routes` / :meth:`UrlDispatcher.routes`.
 
-
 Alternative ways for registering routes
 ---------------------------------------
 
@@ -383,6 +382,101 @@ own.
 
 .. versionadded:: 2.3
 
+Web Handler Cancellation
+------------------------
+
+.. warning::
+
+   :term:`web-handler` execution could be canceled on every ``await``
+   if client drops connection without reading entire response's BODY.
+
+   The behavior is very different from classic WSGI frameworks like
+   Flask and Django.
+
+Sometimes it is a desirable behavior: on processing ``GET`` request the
+code might fetch data from database or other web resource, the
+fetching is potentially slow.
+
+Canceling this fetch is very good: the peer dropped connection
+already, there is no reason to waste time and resources (memory etc) by
+getting data from DB without any chance to send it back to peer.
+
+But sometimes the cancellation is bad: on ``POST`` request very often
+is needed to save data to DB regardless to peer closing.
+
+Cancellation prevention could be implemented in several ways:
+* Applying :func:`asyncio.shield` to coroutine that saves data into DB.
+* Spawning a new task for DB saving
+* Using aiojobs_ or other third party library.
+
+:func:`asyncio.shield` works pretty good. The only disadvantage is you
+need to split web handler into exactly two async functions: one
+for handler itself and other for protected code.
+
+For example the following snippet is not safe::
+
+   async def handler(request):
+       await asyncio.shield(write_to_redis(request))
+       await asyncio.shield(write_to_postgres(request))
+       return web.Response('OK')
+
+Cancellation might be occurred just after saving data in REDIS,
+``write_to_postgres`` will be not called.
+
+Spawning a new task is much worse: there is no place to ``await``
+spawned tasks::
+
+   async def handler(request):
+       request.loop.create_task(write_to_redis(request))
+       return web.Response('OK')
+
+In this case errors from ``write_to_redis`` are not awaited, it leads
+to many asyncio log messages *Future exception was never retrieved*
+and *Task was destroyed but it is pending!*.
+
+Moreover on :ref:`aiohttp-web-graceful-shutdown` phase *aiohttp* don't
+wait for these tasks, you have a great chance to loose very important
+data.
+
+On other hand aiojobs_ provides an API for spawning new jobs and
+awaiting their results etc. It stores all scheduled activity in
+internal data structures and could terminate them gracefully::
+
+   from aiojobs.aiohttp import setup, spawn
+
+   async def coro(timeout):
+       await asyncio.sleep(timeout)  # do something in background
+
+   async def handler(request):
+       await spawn(request, coro())
+       return web.Response()
+
+   app = web.Application()
+   setup(app)
+   app.router.add_get('/', handler)
+
+All not finished jobs will be terminated on
+:attr:`aiohttp.web.Application.on_cleanup` signal.
+
+To prevent cancellation of the whole :term:`web-handler` use
+``@atomic`` decorator::
+
+   from aiojobs.aiohttp import atomic
+
+   @atomic
+   async def handler(request):
+       await write_to_db()
+       return web.Response()
+
+   app = web.Application()
+   setup(app)
+   app.router.add_post('/', handler)
+
+It prevents all ``handler`` async function from cancellation,
+``write_to_db`` will be never interrupted.
+
+.. _aiojobs: http://aiojobs.readthedocs.io/en/latest/
+
 Custom Routing Criteria
 -----------------------