3 ms·

> This is usually considered too expensive in the context of non-coin/-blockchain p2p networks because you want nodes to be able to run on embedded and other re

by jude- 10y ago

> This is usually considered too expensive in the context of non-coin/-blockchain p2p networks because you want nodes to be able to run on embedded and other resource-constrained devices. The O(log n) node state and bootstrap cost limits are quite important. Otherwise it would be akin to asking every mobile phone to keep up to date with the full BGP route set. Funny you should mention BGP. We have been approached by researchers at Princeton who are interested in doing something like that, using Blockstack (but to be fair, they're more interested in giving each home router a copy of the global BGP state). I totally hear you regarding the costly bootstrapping. In Blockstack, for example, we expect most nodes to sync up using a recent signed snapshot of the node state and then use SPV headers to download the most recent transactions. It's a difference between minutes and days for booting up. > Of course, but for some applications that is more than good enough. If your adversary can bring enough resources to bear to take over 1/3rd of your network he might as well DoS any target he wants. So you would be facing massive disruption anyway. Yes. The reason I brought this up is that in the context of public DHTs, it's feasible for someone to run many Sybil nodes. There's some very recent work out of MIT for achieving BFT consensus in open-membership systems, if you're interested: https://arxiv.org/pdf/1607.01341.pdf https://arxiv.org/pdf/1607.01341.pdf > I mean blockchains lose some of their security guarantees too once someone manages to dominate 1/2 of the mining capacity. Same order of magnitude. It's basically the design domain "secure, up to point X". In Bitcoin specifically, the threshold for tolerating Byzantine miners is 25% hash power. This was one of the more subtle findings from Eyal and Sirer's selfish mining paper. > The things you list don't really replace the purpose of a DHT. A dht is a key-value store for many keys and a routing algorithm to find them in a distributed environment. What you listed just gives you a bunch of nodes, but no data lookup capabilities. Essentially you're listing things that could be used to bootstrap into a DHT, not replacing the next layer services provided by a DHT. If the p2p application's steady-state behavior is to run its own overlay network and use the DHT only for bootstrapping, then DHT dependency can be removed simply by using the systems that bootstrap the DHT in order to bootstrap the application. Why use a middle-man when you don't have to?

the8472 10y ago
> If the p2p application's steady-state behavior is to run its own overlay network and use the DHT only for bootstrapping, then DHT dependency can be removed simply by using the systems that bootstrap the DHT in order to bootstrap the application. Why use a middle-man when you don't have to? It seems like we have a quite different understanding how DHTs are used, probably shaped by different use-cases. Let me see if I can summarize yours correctly: a) over time nodes will be interested or have visited in a large proportion of the keyspace b) it makes sense to eventually replicate the whole dataset c) the data mutation rate is relatively low d) access to the keyspace is extremely biased, there is some subset of keys that almost all nodes will access. Is that about right? In my case this is very different. Node turnover is high (mean life time <24h), data is volatile (mean lifetime <2 hours), nodes are only ever interested in a tiny fraction of the keyspace (<0.1%), nodes access random subsets of the keyspace, so there's little overlap in their behavior. The data would become largely obsolete before you even replicated half the DHT unless you spent a lot of overhead on keeping up with hundreds of megabytes of churn per hour and you would never use most of it. So for you there's just "bootstrap dataset" and then "expend a little effort to keep the whole replica fresh". For me there's really "bootstrap into the dht", "maintain (tiny) routing table" and then "read/write random access to volatile data on demand, many times a day". This is why the solutions you propose are no solutions for a general DHT which can also cope with high churn.
- jude- 10y ago
  > It seems like we have a quite different understanding how DHTs are used, probably shaped by different use-cases. Let me see if I can summarize yours correctly: a) over time nodes will be interested or have visited in a large proportion of the keyspace b) it makes sense to eventually replicate the whole dataset c) the data mutation rate is relatively low d) access to the keyspace is extremely biased, there is some subset of keys that almost all nodes will access. Is that about right? Agreed on (a), (b), and (c). In (a), the entire keyspace will be visited by each node, since they have to index the underlying blockchain in order to reach consensus on the state of the system (i.e. each Blockstack node is a replicated state machine, and the blockchain encodes the sequence of state-transitions each node must make). (d) is probably correct, but I don't have data to back it up (e.g. because of (b), a locally-running application node accesses its locally-hosted Blockstack data, so we don't ever see read accesses). > In my case this is very different. Node turnover is high (mean life time <24h), data is volatile (mean lifetime <2 hours), nodes are only ever interested in a tiny fraction of the keyspace (<0.1%), nodes access random subsets of the keyspace, so there's little overlap in their behavior. The data would become largely obsolete before you even replicated half the DHT unless you spent a lot of overhead on keeping up with hundreds of megabytes of churn per hour and you would never use most of it. Thank you for clarifying. Can you further characterize the distribution of reads writes over the keyspace in your use-case? (Not sure if you're referring to the Bittorrent DHT behavior in your description, so apologies if these questions are redundant). For example: * Are there a few keys that are really popular, or are keys equally likely to be read? * Do nodes usually read their own keys, or do they usually read other nodes' keys? * Is your DHT content-addressable (e.g. a key is the hash of its value)? If so, how do other nodes discover the keys they want to read? * If your DHT is not content-addressable, how do you deal with inconsistent writes during a partition? More importantly, how do you know the value given back by a remote node is the "right" value for the key?

by jude- 10y ago

the8472 10y ago

> If the p2p application's steady-state behavior is to run its own overlay network and use the DHT only for bootstrapping, then DHT dependency can be removed simply by using the systems that bootstrap the DHT in order to bootstrap the application. Why use a middle-man when you don't have to? It seems like we have a quite different understanding how DHTs are used, probably shaped by different use-cases. Let me see if I can summarize yours correctly: a) over time nodes will be interested or have visited in a large proportion of the keyspace b) it makes sense to eventually replicate the whole dataset c) the data mutation rate is relatively low d) access to the keyspace is extremely biased, there is some subset of keys that almost all nodes will access. Is that about right? In my case this is very different. Node turnover is high (mean life time <24h), data is volatile (mean lifetime <2 hours), nodes are only ever interested in a tiny fraction of the keyspace (<0.1%), nodes access random subsets of the keyspace, so there's little overlap in their behavior. The data would become largely obsolete before you even replicated half the DHT unless you spent a lot of overhead on keeping up with hundreds of megabytes of churn per hour and you would never use most of it. So for you there's just "bootstrap dataset" and then "expend a little effort to keep the whole replica fresh". For me there's really "bootstrap into the dht", "maintain (tiny) routing table" and then "read/write random access to volatile data on demand, many times a day". This is why the solutions you propose are no solutions for a general DHT which can also cope with high churn.

jude- 10y ago
> It seems like we have a quite different understanding how DHTs are used, probably shaped by different use-cases. Let me see if I can summarize yours correctly: a) over time nodes will be interested or have visited in a large proportion of the keyspace b) it makes sense to eventually replicate the whole dataset c) the data mutation rate is relatively low d) access to the keyspace is extremely biased, there is some subset of keys that almost all nodes will access. Is that about right? Agreed on (a), (b), and (c). In (a), the entire keyspace will be visited by each node, since they have to index the underlying blockchain in order to reach consensus on the state of the system (i.e. each Blockstack node is a replicated state machine, and the blockchain encodes the sequence of state-transitions each node must make). (d) is probably correct, but I don't have data to back it up (e.g. because of (b), a locally-running application node accesses its locally-hosted Blockstack data, so we don't ever see read accesses). > In my case this is very different. Node turnover is high (mean life time <24h), data is volatile (mean lifetime <2 hours), nodes are only ever interested in a tiny fraction of the keyspace (<0.1%), nodes access random subsets of the keyspace, so there's little overlap in their behavior. The data would become largely obsolete before you even replicated half the DHT unless you spent a lot of overhead on keeping up with hundreds of megabytes of churn per hour and you would never use most of it. Thank you for clarifying. Can you further characterize the distribution of reads writes over the keyspace in your use-case? (Not sure if you're referring to the Bittorrent DHT behavior in your description, so apologies if these questions are redundant). For example: * Are there a few keys that are really popular, or are keys equally likely to be read? * Do nodes usually read their own keys, or do they usually read other nodes' keys? * Is your DHT content-addressable (e.g. a key is the hash of its value)? If so, how do other nodes discover the keys they want to read? * If your DHT is not content-addressable, how do you deal with inconsistent writes during a partition? More importantly, how do you know the value given back by a remote node is the "right" value for the key?