32 vs 64-bit timestamp fields
13 messages, 8 to 9 May 2013. Started by Addy Yeow.
Jeff Garzik (3), Peter Todd (3), John Dillon (3), Pieter Wuille (2), Addy Yeow, Mike Hearn
Hi list,
Can someone explain why do we have 32-bit and 64-bit timestamp fields instead of all being 64-bit?
https://en.bitcoin.it/wiki/Protocol_specification
Cheers, Addy
On Wed, May 8, 2013 at 7:39 PM, Addy Yeow <ayeowch@gmail•com> wrote:
Quoted text (6 lines)
> Hi list, > > Can someone explain why do we have 32-bit and 64-bit timestamp fields > instead of all being 64-bit? > > https://en.bitcoin.it/wiki/Protocol_specification
Hysterical raisins.
On Thu, May 09, 2013 at 09:39:10AM +1000, Addy Yeow wrote:
Quoted text (6 lines)
> Hi list, > > Can someone explain why do we have 32-bit and 64-bit timestamp fields > instead of all being 64-bit? > > https://en.bitcoin.it/wiki/Protocol_specification
Who knows?
Satoshi used 32-bits and those fields can't be changed now without every single Bitcoin user changing all at once. (a "hard-fork" change)
We'll probably need to do one of those eventually for other reasons, so we might as well leave fixing the timestamps until then.
On Wed, May 8, 2013 at 11:44 PM, Peter Todd <pete@petertodd•org> wrote:
Quoted text (7 lines)
> Who knows? > > Satoshi used 32-bits and those fields can't be changed now without every > single Bitcoin user changing all at once. (a "hard-fork" change) > > We'll probably need to do one of those eventually for other reasons, so > we might as well leave fixing the timestamps until then.
Perhaps Satoshi did this delibrately, knowing that at some point a hard-fork would be a good idea, so that we all would have a good excuse to do one?
On Wed, May 8, 2013 at 9:00 PM, John Dillon <john.dillon892@googlemail•com> wrote:
Quoted text (2 lines)
> Perhaps Satoshi did this delibrately, knowing that at some point a hard-fork > would be a good idea, so that we all would have a good excuse to do one?
Guffaw :) The year 2038 is so far in the future that it is not really relevant, from that angle.
We need a hard fork to break the 1MB limit, and Satoshi explicitly presumed that would happen sometime in the future.
On Wed, May 08, 2013 at 09:08:34PM -0400, Jeff Garzik wrote:
Quoted text (7 lines)
> On Wed, May 8, 2013 at 9:00 PM, John Dillon > <john.dillon892@googlemail•com> wrote: > > Perhaps Satoshi did this delibrately, knowing that at some point a hard-fork > > would be a good idea, so that we all would have a good excuse to do one? > > Guffaw :) The year 2038 is so far in the future that it is not really > relevant, from that angle.
"Meh". I think it's highly unlikely we'll break the block header format, as it pretty much means invalidating all mining hardware.
There's also no need: 32 bits is plenty of precision. Hell, even 16 bits would do (assuming there's never more than a 65535s (about 18 hours) gap between two blocks). Just assume the "full" 64-bit time is the smallest one that makes sense, given its lower 32 bits.
On Thu, May 9, 2013 at 1:13 AM, Pieter Wuille <pieter.wuille@gmail•com> wrote:
Quoted text (6 lines)
> On Wed, May 08, 2013 at 09:08:34PM -0400, Jeff Garzik wrote: >> Guffaw :) The year 2038 is so far in the future that it is not really >> relevant, from that angle. > > "Meh". I think it's highly unlikely we'll break the block header format, as it > pretty much means invalidating all mining hardware.
Doesn't most mining hardware at the ASCI level start with a SHA256 midstate given that the nonce is at the end? Adding further information to the block should be possible at the beginning of the block without major changes to the mining hardware.
Quoted text (4 lines)
> There's also no need: 32 bits is plenty of precision. Hell, even 16 bits would > do (assuming there's never more than a 65535s (about 18 hours) gap between two > blocks). Just assume the "full" 64-bit time is the smallest one that makes > sense, given its lower 32 bits.
I feel somewhat uncomfortable about the "after-the-fact" auditing possible in this scenario. Besides the timestamping provided by the block headers appears to be useful in some payment protocols, not to mention in general.
On Thu, May 09, 2013 at 01:27:33AM +0000, John Dillon wrote:
Quoted text (8 lines)
> > There's also no need: 32 bits is plenty of precision. Hell, even 16 bits would > > do (assuming there's never more than a 65535s (about 18 hours) gap between two > > blocks). Just assume the "full" 64-bit time is the smallest one that makes > > sense, given its lower 32 bits. > > I feel somewhat uncomfortable about the "after-the-fact" auditing possible in > this scenario. Besides the timestamping provided by the block headers appears > to be useful in some payment protocols, not to mention in general.
Remember that interpreting the timestamp on a block for the purposes of timestamping is a lot more subtle than it appears at first.
Any node will accept a block with a timestamp no more than two hours ahead of what it thinks the current time is. That time is adjusted by the median of the timestamps reported by your peers. For instance the RPC call getinfo returns, among other things:
{
"timeoffset" : -1,
}
That is saying my node's wall clock time is 1 second behind the median reported by it's peers - pretty good!
Naively you might think this means block timestamps are accurate to within 2 hours right? Well, it's not so simple. Nodes will accept any block with any timestamp *after* the median of the last 11 blocks. From CBlock::AcceptBlock():
// Check timestamp against prev
if (GetBlockTime() <= pindexPrev->GetMedianTimePast())
return state.Invalid(error("AcceptBlock() : block's timestamp is too early"));
So in theory a miner could prevent that block from moving forward, although if they do they drive up the difficulty, so all miners have an incentive to set the timestamp accurately.
There are two types of timestamps possible: proofs that data existed before a time, and proofs that data existed after. With the former type the *later* the proof says the data existed, the more conservative the assumptions behind the proof. So simply adding two hours to the block's timestamp is relatively reasonable. (this assumes the attack managed to mine a single block, and all nodes have accurate clocks)
The latter type, where you prove data existed after a given time, is a much more tricky thing to apply. The genesis block is a great example with that famous newspaper headline:
The Times 03/Jan/2009 Chancellor on brink of second bailout for
banks
As I mentioned in my other (private) email to you a few minutes ago, the sig of my emails has the latest block hash in each one. The basic idea is called a random beacon; NIST has a good description and a project to create one:
http://www.nist.gov/itl/csd/ct/nist_beacon.cfm
Now technically speaking a random beacon is actually a more sophisticated concept than just timestamping, the random beacon's value is public and distributed widely, but for timestamping the idea is basically to have an unpredictable number known to have been produced at a certain time.
So you know this email was written after block #235235, timestamp 2013-05-09 01:21:52 right? Not so fast. All you actually know is the PGP *signature* was created after that time, because the actual text of the email is independent of the beacon nonce. (dunno if I have the correct terminology here FWIW)
For a blockchain it's easy enough, the blocks naturally depend on a genesis block, but applying the concept more generally is tricky and application dependent; consider for example proving you created a keypair after some data, which might be a useful thing to prove if the secret key was created in some tamperproof hardware that you know has left the factory and is in your possesion. It's easy to see how to do this with ECC: just use the same techniques as in HD wallets to derive keys.
To use the blockchain as a secure random beacon you need to make two assumptions, 50% of the hashing power is controlled by honest miners, and those honest miners have accurate clocks. With those assumptions you can work out what is the minimum possible time the block could have been accepted by the GetMedianTimePast() function and you are good to go.
What do people do in practice? Well look at http://vog.github.io/bitcoinproof/, they just give the timestamp and nothing else. Same for OpenTimestamps. (although I'm adding this email to my notes, half the reason it's so detailed...)
Back to the block header time... Frankly, the easiest thing to do is just have a flag day where blocks after a certain height are considered to have a different epoch from the standard 1970 one when computing their time. Boring, but it works fine and only needs to be updated every few decades.
You're midstate idea is very clever though and could come in handy in the future for different purposes. Eventually we should discuss this with the ASIC manufacturers - if it can be implemented as a firmware or FPGA upgrade in the field all the better.
On Thu, May 9, 2013 at 1:57 AM, Peter Todd <pete@petertodd•org> wrote:
Quoted text (2 lines)
> Remember that interpreting the timestamp on a block for the purposes of > timestamping is a lot more subtle than it appears at first.
I actually just meant how Pieter Wuille was talking about a blocktime accurate to only within 18 hours. :) But it is a nice writeup!
In any case, for many things simple relative ordering is enough rather than absolute time.
Quoted text (6 lines)
> There are two types of timestamps possible: proofs that data existed > before a time, and proofs that data existed after. With the former type > the *later* the proof says the data existed, the more conservative the > assumptions behind the proof. So simply adding two hours to the block's > timestamp is relatively reasonable. (this assumes the attack managed to > mine a single block, and all nodes have accurate clocks)
Nope. The attacker can make the timestamp on the block they mine as little as the minimum from GetMedianTimePast(), and adding two hours to that number could easily be well before true time.
What you probably need to do is some sort of median time calculation for the blocks around your timestamp. The proof becomes probabalistic based on the % of hashing power the attacker controls and in turn depends on if the time they created their timestamp was of their own choosing.
IE, if you just want to create an inaccurate timestamp, but don't care when, you can just mine blocks and wait until you get lucky. If you need to create an inaccurate timestamp *now* the problem is much harder.
But all this analysis can be developed later, and data timestamped now. :)
Quoted text (5 lines)
> Back to the block header time... Frankly, the easiest thing to do is > just have a flag day where blocks after a certain height are considered > to have a different epoch from the standard 1970 one when computing > their time. Boring, but it works fine and only needs to be updated every > few decades.
Oh, right, yes, that is a much more simple idea and far less prone to bugs.
Many SPV clients wouldn't even need upgrades if they don't acturally validate the blocks they receive and just look for the biggest PoW.
Quoted text (5 lines)
> > You're midstate idea is very clever though and could come in handy in > the future for different purposes. Eventually we should discuss this > with the ASIC manufacturers - if it can be implemented as a firmware or > FPGA upgrade in the field all the better.
Yes
Anyway, you are being distracted from what we were talking about before, get back to work!
On Thu, May 09, 2013 at 02:33:11AM +0000, John Dillon wrote:
Quoted text (9 lines)
> On Thu, May 9, 2013 at 1:57 AM, Peter Todd <pete@petertodd•org> wrote: > > Remember that interpreting the timestamp on a block for the purposes of > > timestamping is a lot more subtle than it appears at first. > > I actually just meant how Pieter Wuille was talking about a blocktime accurate > to only within 18 hours. :) But it is a nice writeup! > > In any case, for many things simple relative ordering is enough rather than > absolute time.
Ah, shoot, I just realized we both got missed Pieter's point entirely: he means to change the meaning of the header timestamp to be relative time passed since the last block...
Well, it was a nice writeup! Thanks for the correction re: probabalistic; you are absolutely correct.
On Wed, May 08, 2013 at 10:42:44PM -0400, Peter Todd wrote:
Quoted text (3 lines)
> Ah, shoot, I just realized we both got missed Pieter's point entirely: > he means to change the meaning of the header timestamp to be relative > time passed since the last block...
No, though that's also a possibility, but a backward-incompatible one.
What I mean is have a well-defined 64-bit timestamp for each block, but only put the lowest 32 bit in the header. Under the condition:
* There is never a gap of more than 136 years between two blocks.
The actual 64-bit timestamp can be deterministically derived from the header, by prefixing it with the lowest 32-bit value that does not cause the result to violate the at-least-above-the-median-of-the-previous-11-blocks rule.
2038 issues only apply to use of signed timestamps, I thought we treat this field as unsigned? Is it really a big deal?
On Thu, May 9, 2013 at 1:12 PM, Pieter Wuille <pieter.wuille@gmail•com> wrote:
Quoted text (30 lines)
> On Wed, May 08, 2013 at 10:42:44PM -0400, Peter Todd wrote: >> Ah, shoot, I just realized we both got missed Pieter's point entirely: >> he means to change the meaning of the header timestamp to be relative >> time passed since the last block... > > No, though that's also a possibility, but a backward-incompatible one. > > What I mean is have a well-defined 64-bit timestamp for each block, but > only put the lowest 32 bit in the header. Under the condition: > > * There is never a gap of more than 136 years between two blocks. > > The actual 64-bit timestamp can be deterministically derived from the > header, by prefixing it with the lowest 32-bit value that does not > cause the result to violate the > at-least-above-the-median-of-the-previous-11-blocks rule. > > -- > Pieter > > ------------------------------------------------------------------------------ > Learn Graph Databases - Download FREE O'Reilly Book > "Graph Databases" is the definitive new guide to graph databases and > their applications. This 200-page book is written by three acclaimed > leaders in the field. The early access version is available now. > Download your free book today! http://p.sf.net/sfu/neotech_d2d_may > _______________________________________________ > Bitcoin-development mailing list > Bitcoin-development@lists•sourceforge•net > https://lists.sourceforge.net/lists/listinfo/bitcoin-development
On Thu, May 9, 2013 at 11:40 AM, Mike Hearn <mike@plan99•net> wrote:
Quoted text (2 lines)
> 2038 issues only apply to use of signed timestamps, I thought we treat > this field as unsigned? Is it really a big deal?
Not a big deal at all, no.