Feature: Support caching of partial responses

Details

(from the bug report): When range_offset_limit is set to -1, Squid tries to fetch the entire object in response to an HTTP range request.

  • Bug: The entire object is fetched even when it is not cacheable (e.g. because it is larger than maximum_object_size or some other criteria).

    • Squid should revert to fetching just the range if the entire object cannot be cached.
    • Squid should be able to fetch just the section(s) of object necessary and fill the gaps in the background. Otherwise, a patch fetching mechanism such as Windows Update, which fetches a patch file in N chunks, will cause the file to be fetched in its entirety N times. This can cause huge inefficiencies. Squid should always check for cacheability before fetching the entire file.

The proper fix for this is to add caching of partial responses, eleminating the need for range_offset_limit entirely.

Proposal 1: chunked sequencing

  • by Faysal Banna (with some small updates)

Squid should store each object as a series of N-byte chunks. When a range is requested the whole of the chunk containing those bytes should be fetched into the relevat chunk position.

For example; in my cache i would save it in ranges of 1KB which means 40KB file will be 40 chunks of 1KB each.

Suppose my file is called faysal.data

chunk0 = faysal.data,0-1KB
chunk1 = faysal,data,1-2KB
...

Now suppose the file is not cached yet nor requested and I request a range: bytes=2100-3000.

What Squid should do is:

  1. skip the 1024-2048 chunk
  2. fetch the 2049-3072 chunk
  3. optionally skip the 3073+ chunks, or contiue fetching.

Caching the chunks received and marking the response as incomplete.

  • rock cache is already operating as a series of ~32KB chunks for each object. It just needs extending to store non-sequential chunks, and detecting when needed chunks are missing to start background request(s) to fill gaps if a client asks more than already present.

Proposal 2: sparse cache file

Squid should open a disk file for the full file size and fill only the bytes received. The rest can be loaded in background requests later.

  • NP: requires some method of mapping which file bytes have been received and accountig for the variable-length metadata headers in each cache object file.

Proposal 3: caching by Range

Squid caches Vary responses using additional cache key components. Range responses should be cached using a similar mechanism altering the cache key for the range.

  • NP: needs logic to represent a list of all stored ranges in the variant meta entry. Vary caching uses X-Vary-Headers. Range would need a payload.
  • Update: a Store-ID helper can possibly be used to track which URLs are stored incompletely and handle the key alterations.


CategoryFeature

Features/PartialResponsesCaching (last edited 2014-05-10 13:38:16 by AmosJeffries)