I think I should be able to help here a little then. Applying the huge
patch did not work any more on the current trunk for me (did not apply
cleanly).
But if you'd describe of what kind of support you'd need exactly (which
tests, checking which code etc), I should be able to spend some time on
that.
Am 16.03.2012 19:37, schrieb ext Alexander Shraer:
I think this is why when you're doing rolling restarts /
reconfiguration you should never have two different servers that have
any chance of being up at the same time with the same id.
With 107 you'd have to remove the server and add a new server with
some different id (choosing the new id is left to the user).
In terms of support with 107 we need all the help we can get :)
Currently there are two parts of it in pretty good shape that I'm
hoping to integrate soon: 1355 and 1411.
Comments or testing of 1411 would be very helpful at this point. Also,
if you wish, you can check out the latest patch for 107 (that patch is
not going to be integrated - instead I'm trying to get it in piece by
piece, but still, you can try it and see if it works for you or if you
have comments. You can also help by writing tests for it).
Best Regards,
Alex
On Fri, Mar 16, 2012 at 2:56 AM, Christian Ziech
<[email protected]> wrote:
Under normal circumstances the ability to detect failures correctly should
be given. The scenario I'm aware of includes one zookeeper system would be
taken down for a reason and then possibly just rebooted or even started from
scratch elsewhere. In both cases however the new host would have the old dns
name but most likely a different IP. But of course that only applies to us
and possibly not to all of the users.
When thinking about the scenario you described I understood where the
problem lies. However wouldn't the same problem also be relevant the way
zookeeper is implemented right now? Let me try to explain why (possibly I'm
wrong here since I may miss some points on how zookeeper servers works
internally - corrections are very welcome):
- Same scenarios as you described - nodes A with host name a, B host name b
and C with host name c
- Also same as in your scenario C is due to some human error falsely
detected as down. Hence C' is brought up and is assigned the same DNS name
as C
- Now rolling restarts are performed to bring in C'
- A resolves c correctly to the new IP and connects to C' but B still
resolves the host name c to the original address of C and hence does not
connect (I think some DNS slowness is also required for your approach in
order for the host name c being resolved to the original IP of C)
- now the rest of your scenario happens: Update U is applied, C' gets slow,
C recovers and A fails.
Of course also this approach requires some DNS craziness but if I did not
make a mistake in my thoughts it should still be possible.
PS: Wouldn't your scenario not also invalidate the solution of the hbase
guys using amazons elastic ips to solve the same problem (see
https://issues.apache.org/jira/browse/HBASE-2327)?
PS2: If the approach I had in mind is not valid, do you guys already have a
plan for when 3.5.0 would be released or could you guys be supported in some
way so that zookeeper-107 makes it sooner into a release?
Am 16.03.2012 04:43, schrieb ext Alexander Shraer:
Actually its still not clear to me how you would enforce the 2x+1. In
Zookeeper we can guarantee liveness (progress) only when x+1 are connected
and up, however safety (correctness) is always guaranteed, even if 2 out of
3 servers are temporarily down. Your design needs the 2x+1 for safety, which
I think is problematic unless you can accurately detect failures (synchrony)
and failures are permanent.
Alex
On Mar 15, 2012, at 3:54 PM, Alexander Shraer<[email protected]> wrote:
I think the concern is that the old VM can recover and try to
reconnect. Theoretically you could even go back and forth between new
and old VM. For example, suppose that you have servers
A, B and C in the cluster, A is the leader. C is slow and "replaced"
with C', then update U is acked by A and C', then A fails. In this
situation you cannot have additional failures. But with the
automatic replacement thing it can (theoretically) happen that C'
becomes a little slow, C connects to B and is chosen as leader, and
the committed update U is lost forever. This is perhaps unlikely but
possible...
Alex
On Thu, Mar 15, 2012 at 1:35 PM,<[email protected]> wrote:
I agree with your points about any kind of VMs having a hard to predict
runtime behaviour and that participants of the zookeeper ensemble running on
a VM could fail to send keep-alives for an uncertain amount of time. But I
don't yet understand how that would break the approach I was mentioning:
Just trying to re-resolve the InetAddress after an IO exception should in
that case still lead to the same original IP address (and eventually to that
node rejoining the ensemble).
Only if that host name (the old node was using) would be re-assigned to
another instance this step of re-resolving would point to a new IP (and
hence cause the old server to be replaced).
Did I understand your objection correctly?
________________________________________
Von: ext Ted Dunning [[email protected]]
Gesendet: Donnerstag, 15. März 2012 19:50
Bis: [email protected]
Cc: [email protected]
Betreff: Re: Zookeeper on short lived VMs and ZOOKEEPER-107
Alexander's comment still applies.
VM's can function or go away completely, but they can also malfunction
in more subtle ways such that they just go VEEEERRRRY slowly. You
have to account for that failure mode. These failures can even be
transient.
This would probably break your approach.
On 3/15/12, [email protected]<[email protected]> wrote:
Oh sorry there is a slight misunderstanding. With VM I did not mean the
java
vm but the Linux vm that contains the zookeeper node. We get notified
if
that goes away and is repurposed.
BR
Christian
Gesendet von meinem Nokia Lumia 800
________________________________
Von: ext Alexander Shraer
Gesendet: 15.03.2012 16:33
An: [email protected]; Ziech Christian (Nokia-LC/Berlin)
Betreff: Re: Zookeeper on short lived VMs and ZOOKEEPER-107
yes, by replacing x at a time from 2x+1 you have quorum intersection.
i have one more question - zookeeper itself doesn't assume perfect
failure detection, which your scheme requires. what if the VM didn't
actually fail but just slow and then tries to reconnect ?
On Thu, Mar 15, 2012 at 2:50 AM, Christian Ziech
<[email protected]> wrote:
I don't think that we could be running into a split brain problem in
our
use
case.
Let me try to describe the scenario we are worried about (assuming an
ensemble of 5 nodes A,B,C,D,E):
- The ensemble is up and running and in sync
- Node A with the host name "zookeeperA.whatever-domain.priv" goes
down
because the VM has gone away
- That removal of the VM is detected and a new VM is spawned with the
same
host name "zookeeperA.whatever-domain.priv" - let's call that node A'
- Node A' zookeeper wants to join the cluster - right now this gets
rejected
by the others since A' has a different IP address than A (and the old
one
is
"cached" in the InetSocketAddress of the QuorumPeer instance)
We could ensure that at any given time there is only at most one node
with
host name "zookeeperA.whatever-domain.priv" known by the ensemble and
that
once one node is replaced, it would not come back. Also we could make
sure
that our ensemble is big enough to compensate for a replacement of
more
than
x nodes at a time (setting it to x*2 + 1 nodes).
So if I did not misestimate our problem it should be (due to the
restrictions) simpler than the problem to be solved by zookeeper-107.
My
intention is basically by solving this smaller discrete problem to not
need
to wait for that zookeeper-107 makes it into a release (the assumption
is
that a smaller fix has a possibly a chance to make it into the 3.4.x
branch
even).
Am 15.03.2012 07:46, schrieb ext Alexander Shraer:
Hi Christian,
ZK-107 would indeed allow you to add/remove servers and change their
addresses.
We could ensure that we always have a more or less fixed quorum of
zookeeper servers with a fixed set of host names.
You should probably also ensure that a majority of the old ensemble
intersects with a majority of the new one.
Otherwise you have to run a reconfiguration protocol similarly to
ZK-107.
For example, if you have 3 servers A B and C, and now you're adding D
and
E
that replace B and C, how would this work ? it is probable that D
and E
don't have the latest state (as you mention) and A is down or doesn't
have
the latest state too (a minority might not have the latest state).
Also,
how
do you prevent split brain in this case ? meaning B and C thinking
that
they
are still operational ? perhaps I'm missing something but I suspect
that
the
change you propose won't be enough...
Best Regards,
Alex
On Wed, Mar 14, 2012 at 10:01 AM, Christian Ziech
<[email protected]<mailto:[email protected]>> wrote:
Just a small addition: In my opinion the patch could really boil
down to add a
quorumServer.electionAddr = new
InetSocketAddress(electionAddr.getHostName(),
electionAddr.getPort());
in the catch(IOException e) clause of the connectOne() method of
the QuorumCnxManager. In addition on should perhaps make the
electionAddr field in the QuorumPeer.QuorumServer class volatile
to prevent races.
I haven't checked this change yet fully for implications but doing
a quick test on some machines at least showed it would solve our
use case. What do the more expert users / maintainers think - is
it even worthwhile to go that route?
Am 14.03.2012 17:04, schrieb ext Christian Ziech:
LEt me describe our upcoming use case in a few words: We are
planning to use zookeeper in a cloud were typically nodes come
and go unpredictably. We could ensure that we always have a
more or less fixed quorum of zookeeper servers with a fixed
set of host names. However the IPs associated with the host
names would change every time a new server comes up. I browsed
the code a little and it seems right now that the only problem
is that the zookeeper server is remembering the resolved
InetSocketAddress in its QuorumPeer hash map.
I saw that possibly ZOOKEEPER-107 would also solve that
problem but possibly in a more generic way than actually
needed (perhaps here I underestimate the impact of joining as
a server with an empty data directory to replace a server that
previously had one).
Given that - from looking at ZOOKEEPER-107 - it seems that it
will still take some time for the proposed fix to make it into
a release, would it make sense to invest time into a smaller
work fix just for this "replacing a dropped server without
rolling restarts" use case? Would there be a chance that a fix
for this makes it into the 3.4.x branch?
Are there perhaps other ways to get this use case supported
without the need for doing rolling restarts whenever we need
to replace one of the zookeeper servers?
