1

Topic: hashgraph effective multi master replication.

https://hashgraph.com/the Author of algorithm does a series of rather courageous announcements. And studying algorithm I to it I trust. The code did not look. 1) steadily works if 2/3 nodes work correctly and can transfer messages each other. Remaining can do everything, including frank sabotage. Correctness of operation is proved . 2/3 fair nodes it is a theoretical minimum. 1/2 about which adherents  tell error or intended boxes. 2) throughput of hundred thousand messages in a second. It is restricted by a network and speed of operation of the slowest node. And if it chokes, the remaining network does not note loss of the fighter. 3) a time delay of a fraction of a second. Grows proportionally Log (N) where N an amount of nodes in a network. 4) the Volume of the control footing necessary for consensus acceptance is insignificant in comparison with pay load volume. The algorithm interface can be reduced to a black box which copy lives on each node and creates magic. The box has a queue of outgoing messages. In it the client puts the message which wants to send and a temporal label which according to the client should be at this message. The network can produce to the message other temporal label. And queue of entering messages having following properties: 1) If one client received the message - means this message all clients in a network received. 2) if the client did not receive the message - means this message anybody did not receive. 3) all clients receive messages in one order. Those if one client sees m1, m2, m3 that and all remaining see the same. 4) each message has an identical temporal label on all nodes. The temporal label is coordinated by all network. Having message queue dataful properties multi master replication everything becomes it is trivial. My observations: the Author it is strengthened it is protected from spiteful hackers. Because of it at it all is covered by a thick layer of cryptography. But it not always is necessary. And to the algorithm it is not necessary for operation. If operation goes strictly in the entrusted network the cryptography can be thrown out. Thereby increasing throughput also reduces a time delay. The time delay is counted for a case when nodes transfer messages in a random way. If in a network to add pseudo-leaders these are especially popular nodes with which other nodes communicate especially often. No additional functionality except popularity at pseudo-leaders is present. That can be reduced a time delay to Log (Log (N)). If the pseudo-leader one and remaining nodes put in order to it each message, and he sends all messages to each node the time delay becomes a constant. In case of death of pseudo-leaders the time delay degrades to Log (N). Also I have a suspicion that presence of pseudo-leaders accelerates consensus acceptance. Thereby reducing time delays even more strongly. But it needs to be checked separately. On remaining properties of a network pseudo-leaders of influence do not render. But here it is necessary to be careful that pseudo-leaders did not die from loading. In case of a casual message exchange loading on nodes is arranged uniformly.... <<RSDN@Home 1.0.0 alpha 5 rev. 0>>

2

Re: hashgraph effective multi master replication.

Hello, WolfHound, you wrote: WH> 2) Throughput of hundred thousand messages in a second. It is restricted by a network and speed of operation of the slowest node. And if it chokes, the remaining network does not note loss of the fighter. In this phrase is internal . I.e., or the slowest node restricts throughput, or the group does not note loss of the fighter. And what happens to already sent and delivered messages? They eternally dangle in a network, inflating basis indefinitely, or at some instant burn down? What happens to nodes which were connected to a network not from the very beginning? They receive the messages sent before their connection, or already are not present? WH> the algorithm Interface can be reduced to a black box which copy lives on each node and creates magic. WH> the box has a queue of outgoing messages. In it the client puts the message which wants to send and a temporal label which according to the client should be at this message. The network can produce to the message other temporal label. And what this temporal label generally does? WH> and queue of entering messages having following properties: WH> 1) If one client received the message - means this message all clients in a network received. I do not understand this statement. If the client temporarily had problems with communication, or with the throughput how and when it receives messages which were received already by other clients? WH> 3) all clients receive messages in one order. Those if one client sees m1, m2, m3 that and all remaining see the same. I suppose, if client A sends messages as it should be m1, m2, m3 all remaining receive them in the same order. But here if two clients, A and B send, agreed among themselves, messages in any order the network does not guarantee that the order of delivery of messages from different sources will be observed. I am right? WH> 4) each message has an identical temporal label on all nodes. The temporal label is coordinated by all network. If message contents same on all nodes, whether that we receive the same, simply including a temporal label as one of fields in the message? WH> having message queue dataful properties multi master replication everything becomes it is trivial.

3

Re: hashgraph effective multi master replication.

Hello, Pzz, you wrote: Pzz> In this phrase is internal . I.e., or the slowest node restricts throughput, or the group does not note loss of the fighter. No. If we  one of nodes it not to eat well. But the remaining network will live all the same. Pzz> and what happens to already sent and delivered messages? They eternally dangle in a network, inflating basis indefinitely, or at some instant burn down? Both variants are possible. Depends on the application-oriented task. For example, if we have some copies of relational database. And we drive through this algorithm of transaction which this DB change. That we can on one of mirrors stop record in basis. To make . To tell networks that it is possible all messages which were to  to delete. Again to resolve record in basis. It rolls all changes which happened during time . On the other hand, if at us financial or other essential information there lies to store history till the end of times the most logical decision. Pzz> that happens to nodes which were connected to a network not from the very beginning? They receive the messages sent before their connection, or already are not present? Both variants are possible. Depends on the application-oriented task. Pzz> and what this temporal label generally does? Shows, when the message has been created. Number and time. For what it is necessary besides depends on the application-oriented task. Pzz> I do not understand this statement. If the client temporarily had problems with communication, or with the throughput how and when it receives messages which were received already by other clients? From other clients when communication will be recovered. Pzz> I suppose, if client A sends messages as it should be m1, m2, m3 all remaining receive them in the same order. Correctly. Pzz> But here if two clients, A and B send, agreed among themselves, messages in any order the network does not guarantee that the order of delivery of messages from different sources will be observed. I am right? The network guarantees that if one node sees a1, b1, b2, a2, b3, a3 all remaining nodes see that too most. A and B can ask algorithm that the relative order of their messages was defined. Here a principle idle time: if we already received the message we cannot send other message so that it appeared before that message that we already received. If B wants to send b1 after a1 to it enough to wait while to it reaches a1. Further A will wait, when it reaches b2 before to send a2 . But between their messages can puts others. If A and B do not want to see remaining messages they can always ignore them. Though if they very much want, they can, communicating through other channel to collect messages in a heap and to make so that they came successively. But for this purpose it is necessary to collect all messages on one node and this node already should put in order all pack. What for it can be necessary... <<RSDN@Home 1.0.0 alpha 5 rev. 0>>

4

Re: hashgraph effective multi master replication.

Hello, WolfHound, you wrote: WH> https://hashgraph.com/WH> the Author of algorithm does a series of rather courageous announcements. And studying algorithm I to it I trust. The code did not look. WH> correctness of operation is proved . I something cannot find to unwinding these proofs on the specified site.

5

Re: hashgraph effective multi master replication.

Hello, Glory, you wrote: I something cannot find to unwinding these proofs on the specified site. Section WHITEPAPERS on a foreground.... <<RSDN@Home 1.0.0 alpha 5 rev. 0>>

6

Re: hashgraph effective multi master replication.

Hello, WolfHound, you wrote: Now I look course on distr. computing on edx, and something not so  than this hashgraph differs from Leader-based Sequence Paxos (or raft', probably)? What there on  to a step is added ? All results 20 summer prescription.

7

Re: hashgraph effective multi master replication.

Hello, Pzz, you wrote: Pzz> I suppose, if client A sends messages as it should be m1, m2, m3 all remaining receive them in the same order. But here if two clients, A and B send, agreed among themselves, messages in any order the network does not guarantee that the order of delivery of messages from different sources will be observed. I am right? All  algorithms work under a condition fifo the channel, i.e. tcp . I.e. messages will guarantee the order against one client, from several - are not present.

8

Re: hashgraph effective multi master replication.

Hello, Sharov, you wrote: S> Now I look course on distr. computing on edx, and something not so  than this hashgraph differs from Leader-based Sequence Paxos (or raft', probably)? Absence of the leader. All nodes make one decision completely independently. S> what there on  to a step is added ? All results 20 summer prescription. The cryptography for consensus acceptance is not necessary. It is necessary only if nodes each other do not trust.... <<RSDN@Home 1.0.0 alpha 5 rev. 0>>

9

Re: hashgraph effective multi master replication.

Hello, Sharov, you wrote: S> All  algorithms work under a condition fifo the channel, i.e. tcp . I.e. messages will guarantee the order against one client, from several - are not present. Not-not-not. On idea, all  just about it: the network gradually comes to some consensus about the order of messages, is simple "" as  on time operation appears as protection from  participants of the judgements in a consensus-turn. And further already the network samo-is balanced, merging queues of subnets and forcing out less powerful judgements from resultant queue.