Resuse the same Client for connecting to Google DNS

[freddyaboulton]

safehttpx/safehttpx/__init__.py

Line 44 in c3d25e1

async with httpx.AsyncClient() as client:

We should use a single client instance in this case as it will save us from opening a new connection every time.

[abidlabs]

I don't think that's possible as the main AsyncClient here:

https://github.com/gradio-app/safefetch/blob/e1343f9ac9fabf65f5f3450dde933fea68bd6cd1/safehttpx/__init__.py#L136

requires passing in the transport, which requires us to have already made the call to the Google DNS to get the IP address

[freddyaboulton]

I was referring to the client in resolve_hostname_google - it's creating it's own client but I think we can just use the same one esp since we're always connecting to the same host in that function.

[Zaczero]

I was working on a similar issue and this design decision led me to create a new package that is hopefully more compatible (not just GET, supporting redirects) and lets the user reuse the AsyncClient for maximum efficiency (otherwise the requests are unacceptably slow). https://github.com/Zaczero/httpx-secure I hope it can help somebody too. It's still very fresh and was designed from scratch, so small implementation details will differ.

[abidlabs]

Thanks @Zaczero, if you'd like to upstream your changes to fix safehttpx, that'd be welcome and appreciated!

[Zaczero]

I think the biggest issue with safehttpx is that it doesn't evict cache after some time. All DNS change and having permanent cache (which lru_cache generally is) will lead to bad experience for the users.

The 2nd priority issue would be to validate that the IPs are not only public but also connectable: many DNS servers list several servers for both load-balancing and fallback in case of downtime. safehttpx doesn't support fallback which again will lead to worse experiences (and this issue together with never evicting cache further compounds the problem).

The 3rd priority issue would be to fix missing client reuse that leads to notably degraded performance of all requests, both in async_resolve_hostname_google and get methods.

I released my package under a 0BSD license so you can reuse parts of the code as you see fit. It hopefully addresses all of these issues and some more.