Client._transport may not exist during Client.__del__() or Client.close()

[j178]

>>> import httpx
>>> httpx.get("http://httpbin.org", proxies="socks5://127.0.0.1:1080")  # this will raise as we don't support socks proxy right now
>>> exit()
# then an `AttributeError` will occur here

httpx/httpx/_client.py

Lines 609 to 620 in c75ddc2

	allow_env_proxies = trust_env and app is None and transport is None
	proxy_map = self._get_proxy_map(proxies, allow_env_proxies)
	self._transport = self._init_transport(
	verify=verify,
	cert=cert,
	http2=http2,
	limits=limits,
	transport=transport,
	app=app,
	trust_env=trust_env,
	)

Since the error occured in _get_proxy_map(), before _transport was created, so Client._transport does not exist actuallly.

Then in __del__ and close(), access to _transport will cause AttributeError:

httpx/httpx/_client.py

Lines 1108 to 1118 in c75ddc2

	def close(self) -> None:
	"""
	Close transport and proxies.
	"""
	if self._state != ClientState.CLOSED:
	self._state = ClientState.CLOSED
	self._transport.close()
	for transport in self._mounts.values():
	if transport is not None:
	transport.close()

[lovelydinosaur]

Gotcha. Surprised me slightly that there was actually a client instance in scope there, that got garbage collected, because we never actually get to the point of returning one from __init__().

Presumably the local exception context is hanging onto a reference to a partially instantiated client.

Pull request #1541 neatens up this behaviour, and have confirmed in the console that it resolves this issue. I don't particularly want to futz around with a test case for it because it's a bit of a weird fiddly edge case, and the change in behaviour feels self-evidently correct enough that I'm personally okay without one.