[infinispan-dev] Using infinispan as quorum-based nosql

Thu Jun 6 04:49:15 EDT 2013

Hello.

We are using pessimistic transaction mode. In this case everything's already locked by the time of prepare, is not it?
As of merge, for quorum mode it's simple - take data from quorum. I think I will try to simply suppress sending data from non-quorum members on merge. Because currently everyone sends it's data and it creates complete mess with unsynchronized data after merge (depending on the timing).

Best regards, Vitalii Tymchyshyn

________________________________

From: infinispan-dev-bounces at lists.jboss.org [mailto:infinispan-dev-bounces at lists.jboss.org] On Behalf Of Dan Berindei
Sent: Wednesday, June 05, 2013 12:04 PM
To: infinispan -Dev List
Subject: Re: [infinispan-dev] Using infinispan as quorum-based nosql

On Mon, Jun 3, 2013 at 4:23 PM, <vitalii.tymchyshyn at ubs.com> wrote:

	Hello.

	Thanks for your information. I will subscribe and vote for the issues noted.
	In the meantime I've implemented hacky JgroupsTransport that "downgrades" all (but CacheViewControlCommand and StateTransferControlCommand) SYNCHRONOUS invokeRemotely calls to SYNCHRONOUS_IGNORE_LEAVERS and checks if required number of answers was received with a filter (I've tried to use original invokeRemotely return value but it often returns some strange value, like empty map). It seems to do the trick for me. But I am still not sure if this has any side effects.

Indeed, I started working on a solution, but I over-engineered it and then I got side-tracked with other stuff. Sorry about that.

The problem with using SYNCHRONOUS_IGNORE_LEAVERS everywhere, as I found out, is that you don't want to ignore the primary owner of a key leaving during a prepare/lock command (or the coordinator, in REPL mode prior to 5.3.0.CR1/ISPN-2772). If that happens, you have to retry on the new primary owner, otherwise you can't know if the prepare command has locked the key or not.

A similar problem appears in non-transactional caches with supportsConcurrentUpdates=true: there the primary owner can ignore any of the backup owners leaving, but the originator can't ignore the primary owner leaving.

	For now I can see merge problem in my test: different values are picked during merge. I am going to dig a little deeper and follow up. But it's already a little strange for me, since the test algorithm is:
	1)Assign "old" value to full cluster (it's REPL_SYNC mode)
	2)Block coordinator
	3)Writer "new" value to one of two remaining nodes. It's syncrhonized to second remaining node
	4)Unblock coordinator
	5)Wait (I could not find a good way to wait for state transfer but wait in this case).
	6)Check the value on coordinator

	And in my test I am randomly getting "old" or "new" in assert. I am now going to check why. May be I will need to "reinitialize" smaller cluster part to ensure data is taken from the quorum part of the cluster.

We don't handle merges properly. See https://issues.jboss.org/browse/ISPN-263 and the discussion at http://markmail.org/message/meyczotzobuva7js

What happens right now is that after a merge, all the caches are assumed to have up-to-date data, so there is no state transfer. We had several ideas floating around on how we could force the smaller partition to receive data from the quorum partition, but I think with the public API your best option is to stop all the caches in the smaller partition after the split and start them back up after the merge.

Cheers

Dan

	Best regards, Vitalii Tymchyshyn

	-----Original Message-----
	From: infinispan-dev-bounces at lists.jboss.org [mailto:infinispan-dev-bounces at lists.jboss.org] On Behalf Of Galder Zamarreno
	Sent: Monday, June 03, 2013 9:04 AM
	To: infinispan -Dev List
	Subject: Re: [infinispan-dev] Using infinispan as quorum-based nosql

	On May 30, 2013, at 5:10 PM, vitalii.tymchyshyn at ubs.com wrote:

	> Hello.
	>
	> We are going to use Infinispan in our project as NoSQL solution. It
	> performs quite well for us, but currently we've faced next problem.
	> Note: We are using Infinispan 5.1.6 in SYNC_REPL mode in small cluster.
	> The problem is that when any node fails, any running transactions wait
	> for Jgroups to decide if it've really failed or not and rollback
	> because of SuspectException after that. While we can live with a
	> delay, we'd really like to skip rolling back. As for me, I actually
	> don't see a reason for rollback because transactions started after
	> leave will succeed. So, as for me, previously running transactions
	> could do the same.

	We're aware of the problem (https://issues.jboss.org/browse/ISPN-2402).

	@Dan, has there been any updates on this?

	> The question for is if node that left will synchronize it's state
	> after merge (even if merge was done without infinispan restart). As
	> for me, it should or it won't work correctly at all.

	This is not in yet: https://issues.jboss.org/browse/ISPN-263

	> So, I've found RpcManager's ResponseMode.SYNCHRONOUS_IGNORE_LEAVERS
	> and think on switching to it for RpcManager calls that don't specify
	> ResponseMode explicitly. As for me, it should do the trick. Also, I am
	> going to enforce Quorum number of reponses, but that's another story.
	> So, how do you think, would it work?

	^ Not sure if that'll work. @Dan?

	> P.S. Another Q for me, how does it work now, when SuspectException is
	> thrown from CommitCommand broadcasting. Af far as I can see, commit is
	> still done on some remote nodes (that are still in the cluster), but
	> rolled back on local node because of this exception. Am I correct?

	^ How Infinispan reacts in these situations depends a lot on the type of communications (synchronous or asynchronous) and the transaction configuration. Mircea can provide more details on this.

	Cheers,

	> This
	> can cause inconsistencies, but we must leave with something in
	> peer-to-peer world :) The only other option is to switch from
	> write-all, read-local to write-quorum, read-quorum scenario that is
	> too complex move for Infinispan as for me.
	>
	> Best regards, Vitalii Tymchyshyn
	>
	> Please visit our website at
	> http://financialservicesinc.ubs.com/wealth/E-maildisclaimer.html
	> for important disclosures and information about our e-mail policies.
	> For your protection, please do not transmit orders or instructions by
	> e-mail or include account numbers, Social Security numbers, credit
	> card numbers, passwords, or other personal information.
	>
	> _______________________________________________
	> infinispan-dev mailing list
	> infinispan-dev at lists.jboss.org
	> https://lists.jboss.org/mailman/listinfo/infinispan-dev

	--
	Galder Zamarreño
	galder at redhat.com
	twitter.com/galderz

	Project Lead, Escalante
	http://escalante.io

	Engineer, Infinispan
	http://infinispan.org

	_______________________________________________
	infinispan-dev mailing list
	infinispan-dev at lists.jboss.org
	https://lists.jboss.org/mailman/listinfo/infinispan-dev
	Please visit our website at
	http://financialservicesinc.ubs.com/wealth/E-maildisclaimer.html
	for important disclosures and information about our e-mail
	policies. For your protection, please do not transmit orders
	or instructions by e-mail or include account numbers, Social
	Security numbers, credit card numbers, passwords, or other
	personal information.

	_______________________________________________
	infinispan-dev mailing list
	infinispan-dev at lists.jboss.org
	https://lists.jboss.org/mailman/listinfo/infinispan-dev

-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://lists.jboss.org/pipermail/infinispan-dev/attachments/20130606/4a53a795/attachment-0001.html 
-------------- next part --------------
Please visit our website at 
http://financialservicesinc.ubs.com/wealth/E-maildisclaimer.html 
for important disclosures and information about our e-mail 
policies. For your protection, please do not transmit orders 
or instructions by e-mail or include account numbers, Social 
Security numbers, credit card numbers, passwords, or other 
personal information.