Tom, I too come from a large scale highly-available software design background. I believe the problem Jan is addressing by having the DP make the decision instead of the server is to avoid what is commonly referred to as a "split brain" situation. The challenge with having the servers make the decision is how to avoid the situation where they both think they are the "master/primary" and send packets confusing the receivers. In many software architectures this problem is addressed by having the "master/primary" exclusively acquire some resource. (For example, in database systems where data corruption can be catastrophic this exclusive resource can be a SCSI device or some other I/O fencing technique.)
I would not really be able to suggest best alternative for the WHIII without fully knowing the underlying architecture. Depending on some of the design already there, combining your suggestion of letting the server decide and Jan's let the client DPs decide if they receive packets from multiple servers. Sometimes in these cases we make an arbitrary decision on which to listen to -- it might be based on the server with the lowest address/serial #/boot time/etc.
I have a great deal of large scale high available architecture design experience. (I almost always seek to have no single points of failure [SPOFs] and no manual intervention required to minimize any human-introduced delay.) If you would like to talk more about the subject of availability feel free to contact me offline. _________________ Kevin Montagne Litkam, Inc 713-397-1930 kevin (at) litkam.com
Tom, I too come from a large scale highly-available software design background. I believe the problem Jan is addressing by having the DP make the decision instead of the server is to avoid what is commonly referred to as a "split brain" situation. The challenge with having the servers make the decision is how to avoid the situation where they both think they are the "master/primary" and send packets confusing the receivers. In many software architectures this problem is addressed by having the "master/primary" exclusively acquire some resource. (For example, in database systems where data corruption can be catastrophic this exclusive resource can be a SCSI device or some other I/O fencing technique.)
I would not really be able to suggest best alternative for the WHIII without fully knowing the underlying architecture. Depending on some of the design already there, combining your suggestion of letting the server decide and Jan's let the client DPs decide if they receive packets from multiple servers. Sometimes in these cases we make an arbitrary decision on which to listen to -- it might be based on the server with the lowest address/serial #/boot time/etc.
I have a great deal of large scale high available architecture design experience. (I almost always seek to have no single points of failure [SPOFs] and no manual intervention required to minimize any human-introduced delay.) If you would like to talk more about the subject of availability feel free to contact me offline. _________________ Kevin Montagne Litkam, Inc 713-397-1930 kevin (at) litkam.com
