Re: Removing dropped entry from ref set

Yes, that's a reasonable interpretation.  While the spec could be clearer, I think that's the correct reading.  That still leaves open the question about correct behavior, though.  You said "You'll need to take the appropriate action when an entry in the reference set gets dropped, which is most likely explicitly emitting it before it gets dropped."  Is this appropriate action on the part of the compressor, or the decompressor?

If the compressor, what action is that -- does it need to notice that an entry it cares about is going to be dropped and immediately encode two indexed references to it, forcing it to be emitted before it's dropped?  Or is it an appropriate action on the decompressor's part, which means the compressor needs to ensure that nothing in the reference set it wishes to remove falls off the list before it gets removed?

Sent from Windows Mail

From: Fred Akalin<mailto:akalin@google.com>
Sent: ?Monday?, ?September? ?23?, ?2013 ?1?:?51? ?PM
To: Mike Bishop<mailto:Michael.Bishop@microsoft.com>
Cc: ietf-http-wg@w3.org<mailto:ietf-http-wg@w3.org>

On Mon, Sep 23, 2013 at 1:37 PM, Mike Bishop <Michael.Bishop@microsoft.com<mailto:Michael.Bishop@microsoft.com>> wrote:
Found a corner case that I'm not sure HPACK handles, or if it does I'm not finding it and we might want to clarify....

Sequence of events:

  *
Header X is present in reference set and header table.
  *
Some other addition to the header table causes header X to be dropped.  (This could be a header that causes X to be dropped, or a substitution operation that replaces X.)

  *
Now the compressor wishes to remove X from the reference set.  How does it do so?

Per 3.2.1, the only operation that has the effect of removing an entry from the reference set is an indexed representation of a header already in the reference set.  But you can't send the indexed representation of a header no longer in the header table.

>From 3.1.3: "A reference set is defined as an unordered set of references to entries of the header table." I interpret this to mean that when an entry is removed from the header table, it's removed from the reference set also.

To address it, we could say:

  *
Replacing X with another entry in the header table also replaces X in the reference set.  This makes perfect sense, but needs to be stated.
  *
Dropping X from the header table also drops it from the reference set.  This is more problematic, because the compressor can't just pass over headers it wants to communicate which are already in the reference set.  Instead, it needs to leave them until the end, then check if they're all still in the reference set.  If not, it needs to take another action to emit one or more, then check that all the remaining are still in the reference set...  This loops until the entire remaining set is still present.

You'll need to take the appropriate action when an entry in the reference set gets dropped, which is most likely explicitly emitting it before it gets dropped. I ran into this same problem, which Tatsuhiro Tsujikawa pointed out: see https://github.com/akalin-chromium/httpbis-header-compression/issues/1 .