ConnectionPool and AsyncConnectionPool remove the underlying __cause__ of the exception

[jhominal]

I have found that both synchronous and asynchronous connection pools will catch an error in handle_(async_)request and re-raise it while clearing out the __cause__ (with raise exc from None)

Why do we clear the __cause__ in this case? This is particularly bothersome for me because anyio does not usually set messages when raising BrokenResourceError (and other exceptions), and that results in me having only a httpcore.ReadError with an empty message, which is not very useful.

In other words: I would llike to modify that line in connection_pool so that __cause__ is not replaced with None, and so that the underlying error can appear when I collect tracebacks.

[TayfurYldz]

This is still present on current main in both pool implementations. The cleanup block ends with:

raise exc from None

after removing the request and reassigning the queue.

There are two separate traceback mechanisms here. from None sets __suppress_context__, so Python hides an exception that was merely active as __context__. It does not erase an explicit __cause__ that was already attached to exc. The practical problem is that backend errors are commonly translated while handling another exception, and suppressing the context at this outer pool boundary can remove the only useful low-level diagnostic from the displayed traceback.

I don't see a pool-specific reason for doing that suppression. The pool is doing bookkeeping and then re-raising the same exception object; it is not translating the exception into a new abstraction at this point. A bare raise after cleanup preserves the original exception propagation semantics and traceback chain while still allowing the queue cleanup to happen.

I would test this at the pool boundary in both sync and async variants: make the network backend fail with a distinctive underlying exception, assert the public httpcore exception remains the same type, and assert that the original diagnostic remains reachable/displayable through its exception chain. That avoids tying the regression test specifically to AnyIO's current error text.