> Hi Evgeny. Daniel Shahaf just noticed that r1803143 extended the
> svn_delta_editor_t interface. So I took a look and see in the log
> message that also:
>> This requires a minor tweak to the Subversion's HTTP protocol, and
>> it's the reason why streaming would only work against new servers.
> It would have been nice to announce these API and protocol changes
> before or after making them, to give devs a better chance to review. I
> don't watch commits closely and hadn't seen this at all.
Thank you for this reminder, I have now taken this to a separate thread.
Adding support for streaming svndiff deltas over ra_serf required a minor
extension of the svn_delta_editor_t interface, and of the Subversion's HTTP
protocol. Both of these changes were implemented in a backward-compatible
manner, and they don't alter the existing behavior of the API or 3rd-party
(1) The new svn_delta_editor_t.apply_textdelta_stream() callback was added.
It works in an inverted way, compared to apply_textdelta(), by allowing
the editor driver to set a callback (svn_txdelta_stream_open_func_t)
that will be called when the implementation requires the txdelta stream.
This is what makes streaming deltas possible in the implementation of
the ra_serf's commit editor.
Perhaps, a potential alternative would be to implement such optimization
in the Ev2, without altering the Ev1 interface. But, as long as the
work on the Ev2 isn't finalized, this relatively small and compatible
extension to the existing Ev1 interface allows us to solve the long-
standing issue with the commits working slower than expected and
requiring disk space to store the temporary files.
(2) The HTTP protocol was extended to report the result checksum via the
"X-SVN-Result-Fulltext-MD5" header in the response to a successful
PUT request. This capability is advertised in the response to the
initial OPTIONS request.
An extended explanation of why this is required is the following:
Currently, all PUT requests include a special header that contains the
result checksum, which is used by the server to validate the integrity
of the result after it applies the delta received over the wire. While
this approach works fine if the client first creates a temporary file
with the delta and only then starts sending it to the server (the result
checksum is calculated while preparing the temporary file), it can't
be used in the stream approach, as with it we'd need to know the result
checksum _before_ we start sending data.
So we turn the existing scheme inside out, and teach mod_dav_svn to send
the result checksum in the responses to PUT requests. Then, the client
would check if the received checksum matches what it calculated, and,
possibly, return a checksum mismatch error (thus aborting the edit and
I hope that I didn't miss something subtle about this, and, as well as that,
I am thinking that these changes are worth the result.