linkcheck builder: begin using requests.Session functionality during linkchecking

[jayaddison]

Is your feature request related to a problem? Please describe.
At the moment, the linkcheck builder performs individual request.get (or similar HTTP request method) operations during linkchecking, without any explicit connection or session pooling.

This may be inefficient, because it seems likely that for many use cases, linkchecking will make multiple requests to the same host (because documentation references are likely to have host-locality).

Describe the solution you'd like
Confirmation that connection pooling is not currently in use would be a good starting point; in other words: we should confirm that linkchecking of multiple URLs on a single host results in multiple TCP connections. Ideally this should be written as a test case.

If we can confirm that the problem exists, then we may be able to use some of the Session object functionality from the requests library that's already in use here to enable connection pooling.

Describe alternatives you've considered
None so far, although open to suggestions (and improvements on the definition of this feature request).

Additional context

• See note / suggestion about use of context managers: linkcheck builder: streamed HTTP requests do not close their local socket #11317 (comment)

[jayaddison]

Confirmation that connection pooling is not currently in use would be a good starting point; in other words: we should confirm that linkchecking of multiple URLs on a single host results in multiple TCP connections. Ideally this should be written as a test case.

I think that the mocket socket-mocking library could be a very nice match for this testing use case.

I'm experimenting with it locally at the moment; initially it seems to be good at mocking the client-side sockets, but when mocking the communication between the linkcheck (client) and test HTTP server, something isn't working as expected.

cc @mindflayer (I'll likely open an issue report or two for mocket when I can figure out what is going on here)

[mindflayer]

Remember that using mocket means mocking socket based modules, and it has to be done in the same thread. What you can do, which is best for many testing use-cases, is using its recording mode. You'll end up with a bunch of fixtures that you can add to your repository.
See https://github.com/mindflayer/python-mocket#example-of-how-to-record-real-socket-traffic.

[jayaddison]

Thanks! The test HTTP server does run on a separate thread, so that's likely the problem. The recording feature looks perfect to determine whether the use of session-based requests has the intended effect.

[mindflayer]

The test HTTP server does run on a separate thread, so that's likely the problem.

That's exactly why I mentioned that. Let me know if you need more support, I am always happy to help.

[jayaddison]

Thanks @mindflayer - I'd be grateful for your help.

I'd been planning to try switching to aiohttp as the test HTTP server, with the aim of handling HTTP server-and-client interaction within a single thread (and simplifying socket mocking).

Does that sound like a sensible approach, and/or do you have other suggestions?

[mindflayer]

You won't need to reach the actual server when introducing mocket, but aiohttp should work effectively as a client.

[jayaddison]

(a shorter summary of a longer draft message) Are you suggesting that instantiating a server is not necessary during these tests?

[mindflayer]

Correct, mocket will be responsible to fool the client and provide the response you registered.

[jayaddison]

Ok, thanks. If possible I would like to retain the existing client/server communication during the tests, so I'm weighing up the options available.

I ran into problems using the recording mode of mocket with the tests as they exist today: it seems that, the server's use of sockets is interfered with after mocket becomes enabled (I can see the listen, bind, accept calls arriving on MocketSocket, but clearly they are not going to behave correctly for a server).

During that same attempt, there was also a problem with a call to getsockname before the self._address attribute had been assigned; I've temporarily updated this line locally to match this one (and I've read some discussion about this on your issue tracker, although I'm not yet sure whether to report this, particularly if the server-side use case is not supported).

[mindflayer]

I am not sure why would you want to use mocket while maintaining a client/server setup.
The main reason for using it is decoupling tests from having to talk to a "real" server (aka mocking). Some mocking libraries do that providing a fake client, mocket allows you to use the original client you'd use in production and it achieves it through monkey-patching the socket module.

[jayaddison]

re: socket mocks

Ok, it seems I didn't match the use-case for mocket correctly here, apologies for the distraction.

@mindflayer: I have some ideas about whether it might be able to support mocket recording mode when both client-and-server are within the same process (as here). Is that worth reporting as a feature request, or is it too obscure a use case? I could do some more development testing locally before creating that, to check whether the ideas are valid.

re: connection pool tracing

I have been able to (hackily) confirm the expected result locally by collecting a set of distinct connections retrieved from urllib3.poolmanager.connection_from_url with-and-without using requests.Session -- as expected (and assertable), the connection count is lower with sessions.

This did require use of (upgrade to) an HTTP/1.1 protocol server during testing -- HTTP/1.0 doesn't keep connections open by default, so the connection count does not change under HTTP/1.0.

[mindflayer]

Is that worth reporting as a feature request, or is it too obscure a use case?

Feel free to open an issue on mocket side after you've done your evaluation. I'll do my best to understand the context and give you my pov on the topic.

[picnixz]

For the linkcheck builder, since each check is a task performed by a HyperlinkAvailabilityCheckWorker and there is a fixed and known number of workers at the beginning of the check, one should:

• order the links by domain before submitting them to the workers. That way, we would process links by "chunks" instead of checking them one by one.
• each worker has a dedicated session object instance. That way, the number of sessions existing simultaneously is upper-bounded by the number of maximum workers (5 by default)
• instead of submitting a single link to a worker, we should split the whole links to check as a list of chunks, each chunk representing a domain with some links to check within that specific domain. Depending on the number of links in the domain, a single worker can either process all the domain, or multiple workers share the burden of processing that domain (possibly all workers). When the tasks for the domain have been submitted, we move to the next domain and do the same with the remaining available workers.

Ideally, if there are a lot of domains but only checking a single link everytime, the chunk approach is not good. In this case, we should do it normally. If there are a lot of links for a single domain, this would reduce the number of sessions being created.

[jayaddison]

I have a branch in progress at https://github.com/jayaddison/sphinx/tree/issue-11324/linkcheck-sessioned-requests that:

• Updates the test HTTP servers to HTTP/1.1 (enabling connection keepalive, and -- I expect -- more closely resembling the webservers that most Sphinx application instances linkcheck)
• Begins using requests.Session functionality
• Removes the linkcheck_timeout configuration (see notes below)

My changes seem naive and don't function particularly well when the tests run.

In particular: without adjusting the timeout, many of the tests fail due to timeout errors that appear in the report results:

(           links: line    3) broken    http://localhost:7777/ - HTTPConnectionPool(host='localhost', port=7777): Read timed out. (read timeout=0.05)

With the timeouts removed, then most of the tests succeed, but run extremely slowly (20 seconds or so per unit test).

Running strace on the pytest process shows repeat calls to futex during the waiting time:

futex(0xa5b8d0, FUTEX_WAKE_PRIVATE, 1)  = 1
futex(0xa5b8d8, FUTEX_WAKE_PRIVATE, 1)  = 1
newfstatat(AT_FDCWD, "/tmp/pytest-of-jka/pytest-5/linkcheck/links.rst", {st_mode=S_IFREG|0644, st_size=648, ...}, 0) = 0
getpid()                                = 5395
getpid()                                = 5395
futex(0xa5b8d4, FUTEX_WAKE_PRIVATE, 1)  = 1
futex(0xa5b8d8, FUTEX_WAKE_PRIVATE, 1)  = 1
newfstatat(AT_FDCWD, "/tmp/pytest-of-jka/pytest-5/linkcheck/links.rst", {st_mode=S_IFREG|0644, st_size=648, ...}, 0) = 0

Socket-level networking and syscall tracing aren't things I have much expertise with, but strace often helps me to get a sense for where problems could exist when a userspace program is running into problems.

It's interesting to me that the tests do eventually pass -- that seems to indicate that both consumer (client) and producer (server) threads are co-operating; my guesses are:

• that there are some extremely inefficient tight loops that occur along the way that near-block the threads until one or both of them receive enough CPU time to process relevant events
  and/or
• that the test server's use of a limited number of sockets (todo: how many?) results in resource contention

Perhaps there's something unusual going on in the socket block/timeout configuration, and/or polling.

It feels like it should be possible to get these two threads to co-operate without these performance penalties, and then to restore the timeout configuration to ensure that the tests run quickly (in terms of elapsed time).