On Jul 20, 2010, at 6:14 PM, Mircea Markus wrote:
On 20 Jul 2010, at 14:07, Galder Zamarreño wrote:
> Hi,
>
> First of all, thanks Mircea for writing this up. These are my comments:
>
> I'm not sure I understand the meaning of or the point you're trying to make
with: "Further on, it is possible (and not difficult) to build local transactions on
top of global transaction: Infinispan's batching API does just that
internally.".
Infinispan batching is a way to execute local transactions.
cache.startBatch(); //start local transactions
//do stuff
cache.endBatch(true); //this would commit/rollback the local transaction
Batching functionality (i.e. local transaction) is implemented by starting a JTA
transaction(i.e. global transaction) within BatchingInterceptor.
Sure, but what I'm not sure I understand this in the context of the section.
You're trying to decide whether Hot Rod should use global or local transactions.
It's clear that it needs global transactions, or JTA transactions. I think talking
about batching confuses things here. Batching AFAIK should be used to batch replication
messages independent of JTA transactions (
http://community.jboss.org/wiki/Batching). I
don't think it's a good idea to tie up batching as a way to build transactions cos
that's not the use case of batching. If you need transactions, use JTA transactions.
>
> Transactionable clients doesn't sound very well to me. I'd go for
'transactional clients'.
+1
>
> I don't think we should support this: "Through HotRod, operations associated
with same transaction might be dispatched to multiple nodes. ". I think this is
rather messy and will cause problems. Think of sticky sessions. Instead, I think
transactional client implementations will need a new load balance policy which is
transaction sticky. IOW, if you call begin tx on node A, you want the rest of transaction
operations to be directed there.
Good point. Thinking some more we can pool the connection to the server so that we
won't keep a TCP connection for the entire duration of the transaction, which would be
bad.
> Otherwise, it gets very messy if the prepare lands on node B and commit on node C.
So, wherever the beginTx lands, that's the node that should be used for the duration
of the transaction. IOW, my vote is definitely for solution 1 which is simpler and avoids
potential lock ups resulting from sending operations in the same tx to diff nodes.
I've just started t o like 1 more as well :) Just to clarify one thing with 2: the tx
would reside on one server only. If the prepare lands on B it is forwarded (through
something like a FrowardCommand) to C where it would be executed. More complex though.
Indeed more complex.
>
> How are we gonna deal with situations where client sends a commitTx which is applied
correctly in the target server and any other involved members in the cluster, but
there's a failure when commitTx response is sent back to client?
XAResource on the client won't confirm the transaction commit to the
TransactionManager(XAResouce.commit would throw an XAException with an specific exception
code). From there on it is with the TXManager.
To which TransactionManager? A transaction manager running on the client or the server?
Another question, if there's an XAResource on the client, the client itself must have
a TM running there?
> The client could think that the commit failed but this worked
fine on the server.
>
> I think we need something other than client intelligence for determining whether a
transaction is present or not for the following reason: Imagine that as part of
transactional operation the server figures out that the client has a stale view. If client
sends 4 as client intelligence, what is the server gonna reply in the topology change
header? Is it gonna reply with no cluster info? or hash aware topology header? I get the
feeling that we're trying to use client intelligence as way to signal that the
operation sent is transactional: "Base on client's intelligence, the server
should be able to determine weather these fields are present or not." and I think
this is not correct. Let's leave client intelligence as it is and let's not try to
give it a different meaning.
Yes, I agree.
>
> Instead, let's use [tx_id length] to signal transactions. First of all, I think
[tx_id length] [tx_id] should be part of the header since it's something common to all
operations,
but ping, but ping is not really relevant :)
Yeah, ping is exceptional.
> rather than appending it at the end of the command where we store
command specific information. On top of that, [tx_id length] can easily be used to signal
a transaction. If [tx_id length] is 0, no tx is being sent. If not 0, a tx_id follows and
hence the operation is transactional.
>
Point taken. This flag is in trunk and is ignored for now, as we've discussed.
Thanks a lot for the feedback, this changes the design significantly. I'll update the
doc and let you know.
You're welcome :)
> Cheers,
>
> On Jul 20, 2010, at 11:45 AM, Mircea Markus wrote:
>
>> Hi,
>>
>> I'be just added a design draft for transactions over hotrod[1]. Feedback
appreciated!
>>
>> Cheers,
>> Mircea
>>
>> [
1]http://community.jboss.org/wiki/TransactionsOverHotRod
>> _______________________________________________
>> infinispan-dev mailing list
>> infinispan-dev(a)lists.jboss.org
>>
https://lists.jboss.org/mailman/listinfo/infinispan-dev
>
> --
> Galder Zamarreño
> Sr. Software Engineer
> Infinispan, JBoss Cache
>
>
> _______________________________________________
> infinispan-dev mailing list
> infinispan-dev(a)lists.jboss.org
>
https://lists.jboss.org/mailman/listinfo/infinispan-dev
_______________________________________________
infinispan-dev mailing list
infinispan-dev(a)lists.jboss.org
https://lists.jboss.org/mailman/listinfo/infinispan-dev
--
Galder Zamarreño
Sr. Software Engineer
Infinispan, JBoss Cache