[xiph-rtp] My impressions about the discussion about codebook delivery.

[Ramón García]

With regard to sending by HTTP a package with all the codebooks.
RealNetworks is interested in changing dynamically the compression
according to network conditions. This implyies dynamic codebook
changes, and probably unpredicatable. It does make sense to tell
clients in advance which codebooks are going to be used, but that
could be done with an attribute in the SDP. Compared to a TAR file
with all the codebooks, this avoids a client downloading the full set
of codebooks instead of only the necessary ones.

I do not see any practical inconvenient in using the main RTP stream
for data-only. This is a mental path that one naturally develops after
reading the spec, but design must be based on balance of pros and
cons. Thus, for sending unexpected codebook changes, I vote for using
the RTP data. After all, there is a marker bit in the header that can
be used for marking packets with codebooks.

Finally, the probability of hash collision is

n*(n - 1)/2/2^b

n = number of codebooks
b = number of bits.

with a good 32 bit hash function, one is safe if the number of
codebooks is very small compared to 2^16, and I think that this is the
case.