Mercurial > lbo > hg > clusterconsensus
changeset 7:a8a747e3e568
Extend README
author | Lewin Bormann <lbo@spheniscida.de> |
---|---|
date | Sat, 08 Oct 2016 00:38:17 +0200 |
parents | 53f946107ec1 |
children | d0c8c9688151 |
files | README.md |
diffstat | 1 files changed, 22 insertions(+), 17 deletions(-) [+] |
line wrap: on
line diff
--- a/README.md Sat Oct 08 00:10:30 2016 +0200 +++ b/README.md Sat Oct 08 00:38:17 2016 +0200 @@ -11,39 +11,44 @@ ### Bootstrapping -* A set of participants is started. Participant `A`'s `ConsensusServer` receives a StartParticipation request. - This request is `StartParticipation(0, 0, Member{addr: "A"}, [Member{addr: "A"}], []byte{})` (i.e., an empty - StartParticipation request) -* Now one participant is running, and sees that it itself is the master. It now finds out in some way which other - participants to add, and calls `AddMember()` on itself. The library will arrange a process by which a new member - is added to the consensus round. -* Repeat this for other members +* A set of participants is started. +* On the participant that has been selected to become the first master, `InitMaster()` is called. +* After `InitMaster()`, `AddParticipants()` adds more participants to the cluster. The list of participants + could for example come from a configuration file or command line flags. +* Then, the cluster should be up and running. Now, on any participant, `Submit()` can be called to submit + changes to the state machine. Those changes are replicated through the master. (see next section) ### Normal operation We assume that the current instance is 12, and the sequence number in that instance is 34. * One participant wants to modify the state. It calls `Submit()` on the `Participant` with one or more changes. -* If the Participant is not master, it will send the request to the master using the `Submit()` method on the stub. -* If the Participant is the master, or has received a `Submit` request, it will proceed by coordinating the change. + * If the Participant is not master, it will send the request to the master using the `Submit()` method on the stub. + * If the Participant is the master, or has received a `Submit` request, it will proceed by coordinating the change. * First, all non-master participants are sent `Accept(12, 35, []Change{ *some change* })`. -* This leads to the non-master participants *staging* that change. -* The next time the master wants to apply a change, it sends `Accept(12, 36, []Change{ *some new change* })`. This leads - to non-master participants "committing" (i.e. applying) all staged changes before sequence `36` to the state machine. + * The request is sent using the `ConsensusClient` stub that was returned by the `ClientFactory` implementation. +* This leads to the non-master participants *staging* that change into a special area. The change is not yet applied to the state machine. +* The next time the master wants to apply another change, it sends `Accept(12, 36, []Change{ *some new change* })`. This leads + to non-master participants "committing" (i.e. applying) all staged changes before sequence `36` to the state machine, including + our change `35` from before. ### Master has crashed -Current instance: 22 Current sequence: 44 +Current instance: 22; Current sequence: 44 * A participant, "B", wants to modify the state, but can't reach the master to `Submit()` a change. It proceeds by becoming a candidate. It sends `Prepare(23, Member{ addr: "B" })` to all known members. If a majority replies within - the deadline with a positive vote, then the `B` is the new master. + the deadline with a positive vote, then `B` is the new master. * In order for all participants to know about this change, `B` submits a request `Accept(23, 1, []Change{})` which confirms - the previous `Prepare()` call (because of the new instance in the `Accept` request) -* Now everyone knows of the new master, and the cluster continues in *Normal operation*. + the previous `Prepare()` call, and makes all non-master participants acknowledge the new master. +* Now everyone knows of the new master, and the cluster continues as in *Normal operation*. ### Adding/removing members * Membership changes are relatively straight-forward; they are just special changes that don't use `Accept()`, but rather the `AddMember()` and `RemoveMember()` methods in `ParticipantStub`. - +* Membership changes are staged the same way as normal changes; memberships only actually change once a membership change + has been committed by an `Accept()` call with a higher sequence number. +* Membership adds are communicated to all old members. After successfully staging the change (i.e., getting a majority vote), + the new member is started using `StartParticipation()`. +* Membership removals are communicated to all members, including the one to be removed. Upon