> The additional benefit of the input script/outpoint filter is to watch for
> unexpected spends (coins getting stolen or spent from another wallet) or
> transactions without a unique change or output address. I think this is a
> reasonable implementation, and it would be nice to be able to download
that
> filter without any input elements.

As someone who's implemented a complete integration of the filtering
technique into an existing wallet, and a higher application I disagree.
There's not much gain to be had in splitting up the filters: it'll result in
additional round trips (to fetch these distinct filter) during normal
operation, complicate routine seed rescanning logic, and also is detrimental
to privacy if one is fetching blocks from the same peer as they've
downloaded the filters from.

However, I'm now convinced that the savings had by including the prev output
script (addr re-use and outputs spent in the same block as they're created)
outweigh the additional booking keeping required in an implementation (when
extracting the precise tx that matched) compared to using regular outpoint
as we do currently. Combined with the recently proposed re-parametrization
of the gcs parameters[1], the filter size should shrink by quite a bit!

I'm very happy with the review the BIPs has been receiving as of late. It
would've been nice to have this 1+ year ago when the draft was initially
proposed, but better late that never!

Based on this thread, [1], and discussions on various IRC channels, I plan
to make the following modifications to the BIP:

  1. use P=2^19 and M=784931 as gcs parameters, and also bind these to the
     filter instance, so future filter types may use distinct parameters
  2. use the prev output script rather than the prev input script in the
     regular filter
  3. remove the txid from the regular filter(as with some extra book-keeping
     the output script is enough)
  4. do away with the extended filter all together, as our original use case
     for it has been nerfed as the filter size grew too large when doing
     recursive parsing. instead we watch for the outpoint being spent and
     extract the pre-image from it if it matches now

The resulting changes should slash the size of the filters, yet still ensure
that they're useful enough for our target use case.

[1]:
https://lists.linuxfoundation.org/pipermail/bitcoin-dev/2018-May/016029.html

-- Laolu