Raft guarantees safety (consistency) no matter which servers can
communicate with which others (outside of Byzantine faults and
amnesia).
Regarding availability, the basic leader election algorithm assumes
that each server is either communicating with the others or isn't.
There are two extensions that can help with asymmetric cases:
1. A server can refuse to update its term or grant its vote if it's
recently heard from a current leader. See the top of page 42 in my
dissertation:
https://github.com/ongardie/dissertation#readme . With
this, while a server can't hear from a valid leader, it won't be able
to disrupt that leader.
2. Servers can go through a pre-vote phase before incrementing their
terms when starting elections: see section 9.6 in my dissertation.
With this, when a server regains connectivity with a leader, it won't
disrupt that leader.
Given both of these, I believe Raft will be more robust in the face of
asymmetric network links. However, I don't have a formal model for
which class of issues it would and wouldn't be able to handle (that'd
be interesting to think about).
-Diego
On Wed, Feb 11, 2015 at 2:47 AM, Vasileios Anagnostopoulos
<
fithi...@gmail.com> wrote:
> Yes,
>
> but the actual intention was "Computers in different geographic regions
> interconnected with Internet". So WAN was my intention.
>
> On Tue, Feb 10, 2015 at 7:25 PM, Kijana Woodard <
kijana....@gmail.com>
> wrote:
>>
>> By MAN, do you mean "Metropolitan Area Network"?
>>
>>
http://education-portal.com/academy/lesson/computer-networks-and-distributed-processing-pan-lan-wan-man.html
>>
>> On Mon, Feb 9, 2015 at 8:14 AM, Oren Eini (Ayende Rahien)
>> <
aye...@ayende.com> wrote:
>>>
>>> But all nodes are up and B & C can talk to A?
>>>
>>>
>>>
>>>
>>> On Mon, Feb 9, 2015 at 4:14 PM, Oren Eini (Ayende Rahien)
>>> <
aye...@ayende.com> wrote:
>>>>
>>>> What do you mean, MAN, that you have two nodes (out of three) that can't
>>>> talk to one another?
>>>>
>>>>
>>>>
>>>>
>>>>