1 files changed, 61 insertions, 0 deletions
diff --git a/doc/src/guide/broken_clients.ezdoc b/doc/src/guide/broken_clients.ezdoc
new file mode 100644
index 0000000..26568a3
--- /dev/null
+++ b/doc/src/guide/broken_clients.ezdoc
@@ -0,0 +1,61 @@
+::: Dealing with broken clients
+
+There exists a very large number of implementations for the
+HTTP protocol. Most widely used clients, like browsers,
+follow the standard quite well, but others may not. In
+particular custom enterprise clients tend to be very badly
+written.
+
+Cowboy tries to follow the standard as much as possible,
+but is not trying to handle every possible special cases.
+Instead Cowboy focuses on the cases reported in the wild,
+on the public Web.
+
+That means clients that ignore the HTTP standard completely
+may fail to understand Cowboy's responses. There are of
+course workarounds. This chapter aims to cover them.
+
+:: Lowercase headers
+
+Cowboy converts all headers it receives to lowercase, and
+similarly sends back headers all in lowercase. Some broken
+HTTP clients have issues with that.
+
+A simple way to solve this is to create an `onresponse` hook
+that will format the header names with the expected case.
+
+``` erlang
+capitalize_hook(Status, Headers, Body, Req) ->
+    Headers2 = [{cowboy_bstr:capitalize_token(N), V}
+        || {N, V} <- Headers],
+    {ok, Req2} = cowboy_req:reply(Status, Headers2, Body, Req),
+    Req2.
+```
+
+Note that SPDY clients do not have that particular issue
+because the specification explicitly says all headers are
+lowercase, unlike HTTP which allows any case but treats
+them as case insensitive.
+
+:: Camel-case headers
+
+Sometimes it is desirable to keep the actual case used by
+clients, for example when acting as a proxy between two broken
+implementations. There is no easy solution for this other than
+forking the project and editing the `cowboy_protocol` file
+directly.
+
+:: Chunked transfer-encoding
+
+Sometimes an HTTP client advertises itself as HTTP/1.1 but
+does not support chunked transfer-encoding. This is invalid
+behavior, as HTTP/1.1 clients are required to support it.
+
+A simple workaround exists in these cases. By changing the
+Req object response state to `waiting_stream`, Cowboy will
+understand that it must use the identity transfer-encoding
+when replying, just like if it was an HTTP/1.0 client.
+
+``` erlang
+Req2 = cowboy_req:set(resp_state, waiting_stream).
+```