Re: The case for Content-Unencoded-Length

On 16/07/2026 20:45, Guoye Zhang wrote:
> Hi,
> 
> We recently enabled zstd support in Safari and received complaints that downloads from a popular cloud storage website are no longer showing progress and no longer resumable. Turns out that they are doing streaming zstd compression so they don’t know the Content-Length.
> 
> Are there interests in standardizing a new header field Content-Unencoded-Length (similar to Unencoded-Digest)? I'm also interested in discussing a new range unit unencoded-bytes for download resumption.
> 

I think you have a use-case where Transfer-Encoding is what should be 
used instead of Content-Encoding.


Download resumption requires that when transaction #2 is received the 
server still has the response representation that was truncated from 
transaction #1.  When that condition is true the client can just send a 
range request for bytes "615-" to resume an unknown-length response 
truncated at byte 614).
If that requirement is false, there is no hope of resuming, ever.


For example; the a file may be stream compressed _before_ storage at the 
server, then served in its pre-compressed form to clients - download 
resume will "just work".

BUT, if the file is stream compressed for each unique transaction, then 
the server providing the resumed response likely no longer have the 
compression context to know what binary output should be delivered 
"on-wire". It may not even be the same machine.


A server which is initializing a new compression stream for each client 
request should use Transfer-Encoding instead of Content-Encoding. Any 
Range offsets for download resume will apply to the un-compressed 
response representation, and a new compression stream can trivially be 
initiated for each range response delivered.

HTH
AYJ

Received on Friday, 17 July 2026 03:24:22 UTC