If a document references an external entity or a schema references and include file and retrieving the other HTTP resource takes for too long, the HTTP connection for the document/schema times out.

Instead of leaving the original socket hanging around without reading from it, the stream from the socket should somehow be exhausted anyway even if the parser wants to read something else first.

Solving this problem neatly is hard. Exhausting the streams sequentially would require changes to the parsers. Using worker threads to exhaust socket threads would add threads and thread synchronization which would probably be inefficient in the common case.