Karl wrote:
> I am also curious if you have considered digests containing multiple
> blocks. Retaining a permanent binsearchable record of the entire chain is
> obviously too space costly, but keeping the last X blocks as binsearchable
> could speed up syncing for clients tremendously, I feel.
Originally we hadn't considered such an idea. Grasping the concept a bit
better, I can see how that may result in considerable bandwidth savings
(for purely negative queries) for clients doing a historical sync, or
catching up to the chain after being inactive for months/weeks.
If we were to purse tacking this approach onto the current BIP proposal,
we could do it in the following way:
* The `getcfilter` message gains an additional "Level" field. Using
this field, the range of blocks to be included in the returned filter
would be Level^2. So a level of 0 is just the single filter, 3 is 8
blocks past the block hash etc.
* Similarly, the `getcfheaders` message would also gain a similar field
with identical semantics. In this case each "level" would have a
distinct header chain for clients to verify.
> How fast are these to create? Would it make sense to provide digests on
> demand in some cases, rather than keeping them around indefinitely?
For larger blocks (like the one referenced at the end of this mail) full
construction of the regular filter takes ~10-20ms (most of this spent
extracting the data pushes). With smaller blocks, it quickly dips down to
the nano to micro second range.
Whether to keep _all_ the filters on disk, or to dynamically re-generate a
particular range (possibly most of the historical data) is an
implementation detail. Nodes that already do block pruning could discard
very old filters once the header chain is constructed allowing them to
save additional space, as it's unlikely most clients would care about the
first 300k or so blocks.
> Ahh, so you actually make a separate digest chain with prev hashes and
> everything. Once/if committed digests are soft forked in, it seems a bit
> overkill but maybe it's worth it.
Yep, this is only a hold-over until when/if a commitment to the filter is
soft-forked in. In that case, there could be some extension message to
fetch the filter hash for a particular block, along with a merkle proof of
the coinbase transaction to the merkle root in the header.
> I created digests for all blocks up until block #469805 and actually ended
> up with 5.8 GB, which is 1.1 GB lower than what you have, but may be worse
> perf-wise on false positive rates and such.
Interesting, are you creating the equivalent of both our "regular" and
"extended" filters? Each of the filter types consume about ~3.5GB in
isolation, with the extended filter type on average consuming more bytes
due to the fact that it includes sigScript/witness data as well.
It's worth noting that those numbers includes the fixed 4-byte value for
"N" that's prepended to each filter once it's serialized (though that
doesn't add a considerable amount of overhead). Alex and I were
considering instead using Bitcoin's var-int encoding for that number
instead. This would result in using a single byte for empty filters, 1
byte for most filters (< 2^16 items), and 3 bytes for the remainder of the
cases.
> For comparison, creating the digests above (469805 of them) took
> roughly 30 mins on my end, but using the kstats format so probably
> higher on an actual node (should get around to profiling that...).
Does that include the time required to read the blocks from disk? Or just
the CPU computation of constructing the filters? I haven't yet kicked off
a full re-index of the filters, but for reference this block[1] on testnet
takes ~18ms for the _full_ indexing routine with our current code+spec.
[1]: 000000000000052184fbe86eff349e31703e4f109b52c7e6fa105cd1588ab6aa
-- Laolu
On Sat, Jun 3, 2017 at 2:55 AM, Alex Akselrod via bitcoin-dev
<bitcoin-dev@lists.linuxfoundation.org> wrote:
> Without a soft fork, this is the only way for light clients to verify that
> peers aren't lying to them. Clients can request headers (just hashes of the
> filters and the previous headers, creating a chain) and look for conflicts
> between peers. If a conflict is found at a certain block, the client can
> download the block, generate a filter, calculate the header by hashing
> together the previous header and the generated filter, and banning any peers
> that don't match. A full node could prune old filters if you wanted and
> recalculate them as necessary if you just keep the filter header chain info
> as really old filters are unlikely to be requested by correctly written
> software but you can't guarantee every client will follow best practices
> either.
Ahh, so you actually make a separate digest chain with prev hashes and
everything. Once/if committed digests are soft forked in, it seems a
bit overkill but maybe it's worth it. (I was always assuming committed
digests in coinbase would come after people started using this, and
that people could just ask a couple of random peers for the digest
hash and ensure everyone gave the same answer as the hash of the
downloaded digest..).
> The simulations are based on completely random data within given parameters.
I noticed an increase in FP hits when using real data sampled from
real scriptPubKeys and such. Address reuse and other weird stuff. See
"lies.h" in github repo for experiments and chainsim.c initial part of
main where wallets get random stuff from the chain.
> I will definitely try to reproduce my experiments with Golomb-Coded
> sets and see what I come up with. It seems like you've got a little
> less than half the size of my digests for 1-block digests but I
> haven't tried making digests for all blocks (and lots of early blocks
> are empty).
>
>
> Filters for empty blocks only take a few bytes and sometimes zero when the
> coinbase output is a burn that doesn't push any data (example will be in the
> test vectors that I'll have ready shortly).
I created digests for all blocks up until block #469805 and actually
ended up with 5.8 GB, which is 1.1 GB lower than what you have, but
may be worse perf-wise on false positive rates and such.
> How fast are these to create? Would it make sense to provide digests
> on demand in some cases, rather than keeping them around indefinitely?
>
>
> They're pretty fast and can be pruned if desired, as mentioned above, as
> long as the header chain is kept.
For comparison, creating the digests above (469805 of them) took
roughly 30 mins on my end, but using the kstats format so probably
higher on an actual node (should get around to profiling that...).
_______________________________________________
bitcoin-dev mailing list
bitcoin-dev@lists.linuxfoundation.org
https://lists.linuxfoundation.org/mailman/listinfo/bitcoin-dev