- From: Henrik Nordstrom <henrik@henriknordstrom.net>
- Date: Fri, 31 Jul 2009 11:36:42 +0200
- To: HTTP Working Group <ietf-http-wg@w3.org>
The recommendation that proxies should only return the requested range to the client when receiving a 200 response from the server has a some undesired network effects. The text currently reads: p5-range 5.4.2. Range Retrieval Requests If a proxy that supports ranges receives a Range request, forwards the request to an inbound server, and receives an entire entity in reply, it SHOULD only return the requested range to its client. It SHOULD store the entire received response in its cache if that is consistent with its cache allocation policies. Proposed change: Reduce "SHOULD only return" to a MAY level requirement, and remove the caching part (covered elsewhere), making the text read If a proxy that supports ranges receives a Range request, forwards the request to an inbound server, and receives an entire entity in reply, it MAY return only the requested range to its client. And probably this should be restricted to 200 responses, and reminding that "supports ranges" also involve processing conditionals as suitable. The intentions of the original text is to optimize the last hop to the client, but unfortunately it has some quite noticeable bad effectsand often do not make sense to implement as specified (SHOULD). Some examples: * A client making a range request for the last 200 bytes of a 8TB file. As the response "never" comes the client usually times out. * "download accelerators" accelerating the problem by making many Range requests for different parts and as the proxy is masking the problem these "download accelerators" have no chance of realizing things have gone "bad". * Guaranteed extra network load if the resulting object is not cachable (or when there is no cache in the proxy). Many times the client do really intend to request the rest a little later, and if the object is not cached this results in yet another full download of the object by the proxy. * Bandwidth allocation policy. It's relatively easy to implement a reasonable bandwidth policy by downloading objects at about the same rate those can be delivered to the requesting client, but far from trivial to select a suitable download rate when only spooling the data into cache with no client waiting for the data currently received. We (Squid) originally had Range implemented as specified, but due to this frequently causing more bandwidth issues and confusion than it helped we changed the implementation many years ago to by default NOT implement Range ourselves when getting a 200 response to a forwarded Range request, only implementing Range on cached responses. Regarding the "SHOULD store the entity" part this is highly redundant with other parts of the specification and do not really add anything to be specified here, just confusion making one think that there is something special with how 200 responses to a Range request should be cached differently from 200 responses to non-Range requests. Regards Henrik
Received on Friday, 31 July 2009 09:37:18 UTC