Re: endb 101

In article <[email protected]> you wrote:
> Peter T. Breuer schrieb:
> 
> >>machine A: md0 [hda3+nda0]
> >>                       |
> >>                       |__________________________
> >>                                                  |
> >>machine B: md0 [hda3+nda0], enbd-server exports hda3
> > 
> > 
> > I don't understand this diagram very well, because of the "machine B:
> > md0 [hda3+nda0]" part. What is that doing there? machine B is exporting
> > hda3, not building it into a raid array!
> 
> aha. i really though both machines build its own md0 out of
> 1) a local partition
> 2) an imported enbd

Well, they do - but not at the same time!

The whole idea of fallover is that you have a "floating IP address"
that moves between two machines, according to which of them is acting
as the server at that moment. That "primary" (A), when it breaks, passes the
IP over to the "secondary" (B), which has an up-to-date mirror copy of
the data to serve from.

When A comes back up again, it's likely the roles will remain reversed
until B breaks down for some reason.  Depends how you prefer to do it.

It's also normal to have TWO floating IPs, and TWO resources on each
of TWO servers. One resource on each server acts as a remote mirror
for a resource on the other server. When one server breaks, BOTH
floating IP addresses pass to the remaining server (well, it had one
already).

> so then it's like this:
> 
> machine-A: md0 [hda3+nda0]
> 
> machine-B:     exporting hda3, but were do the applications
>                 write to?

What applications? There should be a service running on the floating IP
on machine A which uses md0. It can be an nfs service if you like! All
access should be via that IP address, and will thus be to machine A,
until it breaks.

> so, machine-A has md0, the applications write to md0, if Machine-B is 
> down, Machine-A's md0 can still work in degraded mode.

Yes.

> if Machine-A is down, Machine-B has to serve files. does it serve them 
> from hda3?

Yes. Or if it likes it can start a trivial raid mirror using hda3 and
waiting for machine A to come back up, which it can then use as the 
second half of the raid mirror.


> so machine-A is the only machine with a md0 made from a local partition 
> and an nda0 from some other machine.

While it is running, yes.


> i *thought* all machines could 
> carry and md0. i was wrong. hm.

They can, and you are not wrong. There is no reason for B to use a
mirror (in degraded mode) while A is running, but it can. It would
make things a little easier for when A comes back up again.

> > The intention is key. I think you want to build a failover arrangement?
> 
> yes.
> 
> > In that case, you are doing fine. Now you export md0 from A,
> 
> "export md0 from A"? you mean, i "use md0 on machine A, which is made 
> from local hda3+nda0", right?

Yes - you use it as the basis of whatever service you are running via
the floating IP, which currently resides on A.

> > and when A breaks you fail over to exporting hda3 from B instead. 
> 
> if A breaks, B takes over and B's hda3 (which is exported by 
> enbd-server, but not used by A now) is serving the files.

Yes.

> > If B breaks instead you still have the hda3 part of the raid on A
>  > working.
> 
> yes. i think i got it now, if you can please confirm the statement above.

Sure. I think you have it too!

> 
> > There are prepared failover scripts for heartbeat in the etc/ directory
> > in the archive.
> 
> yes, that'll be a tricky part too.
> 
> > It's fine - it's even convenient, as you can switch off kernel
> > buffering (use the switch that does open O_DIRECT on the server) and
> > rely on the db servers to sync between distributed copies of the db.
> 
> ok. i'm thinking about one issue here, but it's not really 
> enbd-specific: if the db-server (mysqld) is running on A and B all the 
> time and now A goes down, B takes over. fine. but in the time A *and* B 
> are running, only *one* machines should be allowed to write, yes? (i 

Indeed. Only one is normally capable of being accessed, because only
one has the floating IP address through which communication occurs.

But with a DB server that understands distribution, I think it doesn't
matter - not that I have personal experience.

> guess i'll only start mysqld on B when A is really down).

That would be the standard way. But mysqld may understand the concept
of running two places at once! Oracle does, I believe.

Peter