"Luis Claudio R.Goncalves" wrote:
>
> Hi!
>
> I was reading again your email and our chat log and... I think, as we're
> using the OtherNodeState as part of decision, we will have the same startup
> procedures as we have in heartbeat.
Did you mean the same kinds of race conditions we ran into with
nice_failback?
If so, it's a little different, since if both are down, each side must wait
until both sides are up before it can proceed. This is a vital difference,
because you can't have both sides trying to come up independently at the
same time because you don't yet know who has the "good bits". In the case
of heartbeat, either side could easily take over, so both would try to. In
this case, only one is qualified, you must wait for both sides to come up
before proceeding.
In no case can a machine coming up by itself take anything over without
communicating with the other side.
> Not strictly the same bu we do have to take care of some conditionms
> that may become races.
The possible race conditions include:
Both sides coming up at once: Solved because both must wait for the other
before
proceeding
One side coming up a little after the other:
Each side must wait to rendevous before continuing.
One side already being up and the second coming up:
Subcase other is PRI: No problem, he's active and you slave to him.
Subcase other is SEC: No problem, he's been waiting for you.
There ARE some interesting things to think about, like one side tries to
become master, but cannot because he doesn't know if he has the good bits or
not. Then the "become master" script must hang until the other side comes
up. It may also have to hang until it does a slow sync from the other
side. But again, the communication is wonderfully synchronous: Any side
which is down will never come up until it hears from the other side. If the
other side is up, then by definition it is either primary and has the "good
bits", or it is secondary and is just waiting for you to come along and
find out who has the "good bits".
There are probably other subtle problems hanging out here waiting to bite
one in the hindquarters.
-- Alan Robertson
alanr@example.com
>
> Hi!
>
> I was reading again your email and our chat log and... I think, as we're
> using the OtherNodeState as part of decision, we will have the same startup
> procedures as we have in heartbeat.
Did you mean the same kinds of race conditions we ran into with
nice_failback?
If so, it's a little different, since if both are down, each side must wait
until both sides are up before it can proceed. This is a vital difference,
because you can't have both sides trying to come up independently at the
same time because you don't yet know who has the "good bits". In the case
of heartbeat, either side could easily take over, so both would try to. In
this case, only one is qualified, you must wait for both sides to come up
before proceeding.
In no case can a machine coming up by itself take anything over without
communicating with the other side.
> Not strictly the same bu we do have to take care of some conditionms
> that may become races.
The possible race conditions include:
Both sides coming up at once: Solved because both must wait for the other
before
proceeding
One side coming up a little after the other:
Each side must wait to rendevous before continuing.
One side already being up and the second coming up:
Subcase other is PRI: No problem, he's active and you slave to him.
Subcase other is SEC: No problem, he's been waiting for you.
There ARE some interesting things to think about, like one side tries to
become master, but cannot because he doesn't know if he has the good bits or
not. Then the "become master" script must hang until the other side comes
up. It may also have to hang until it does a slow sync from the other
side. But again, the communication is wonderfully synchronous: Any side
which is down will never come up until it hears from the other side. If the
other side is up, then by definition it is either primary and has the "good
bits", or it is secondary and is just waiting for you to come along and
find out who has the "good bits".
There are probably other subtle problems hanging out here waiting to bite
one in the hindquarters.
-- Alan Robertson
alanr@example.com