I think you misunderstood what I"m getting at. Consider this normal sequence of events under the proposed enlistment strategy:
System1 creates a new transaction
System1 invokes EJB1 on System2 with XID1
System2 examines EJB1's transaction mode
System2 inflows XID1
System2 executes EJB1
System2 returns with "EnlistMe" flag
System1 enlists System2
...other work happens...
System1 prepares XID1
System2 prepares XID1 & returns XA_OK
System1 commits
System2 commits
Now what happens if the System1-System2 connection is broken before EJB1 returns? System2 is never enlisted, System1 may prepare & commit the XID1 transaction without ever knowing about System2, which now has an orphaned transaction which has performed some unknown amount of work under the same GTID as the now-committed transaction. System2's transaction times out, the work is rolled back, and chaos ensues.
So this means that I have to have System2 tell System1 to enlist *before* "System2 executes EJB1" rather than after, right?--
On 02/09/2016 10:41 AM, Tom Jenkinson wrote:
My understanding would be that the EJB transport would detect the
transaction.commit call (although it should have an outstanding EJB call so
I don't think that it would be possible for a client to call commit) and
throw an exception. It could be that EJB transport would register a
synchronization beforeCompletion to do this.
On 9 February 2016 at 14:41, David M. Lloyd <david.lloyd@redhat.com> wrote:
I have a follow-up question about delaying enlistment (I'm getting to this
point in the code).
I'm worried about a scenario wherein the subordinate server begins the
transaction, does some work in the EJB, and then tries to return, but the
connection to the server has been partitioned indefinitely. The server
*may* opt to continue and commit the transaction despite the subordinate
failure, in which case the subordinate, not receiving any control messages
from the root coordinator, would roll back the transaction at the timeout,
while the server (not knowing about the subordinate server) would continue
and commit the transaction, not knowing what work (if any) the subordinate
server had performed.
Is this a real problem? Perhaps I do need an "enlist me" callback to the
server after all, which executes before the EJB imports and resumes the
subordinate transaction?
On 02/02/2016 08:27 AM, Tom Jenkinson wrote:
No problem :)--
On 2 February 2016 at 13:49, David M. Lloyd <david.lloyd@redhat.com>
wrote:
Ah okay, cool. This is an easy (and, in hindsight, rather obvious)
enhancement that I can build into the new protocol with a minimum of
effort.
Thanks!
On 02/02/2016 03:01 AM, Tom Jenkinson wrote:
Hi David,
- DML
I was referring to case 1 when the transaction is inflowed into a second
server. For JTS the type of transaction we create is a
ServerTopLevelAction and this in its ctor calls back to the remote
server:
https://github.com/jbosstm/narayana/blob/master/ArjunaJTS/jts/classes/com/arjuna/ats/internal/jts/orbspecific/interposition/resources/arjuna/ServerTopLevelAction.java#L121
If the transport can do that without the TMs assistance then that works
for me :)
I don't think we should optimize for case 2. The incidents where a
transaction is created and not used should be really low.
Thanks,
Tom
On 1 February 2016 at 15:43, David M. Lloyd <david.lloyd@redhat.com
<mailto:david.lloyd@redhat.com>> wrote:
I was thinking about this a bit. It seems to me that there are two
"levels" of this that could be explored:
1. A transaction was made available to the server, but the EJB on
the server does not use the caller's transaction context, so the
EJB
code never actually has to inflow the transaction. The EJB code
would be able to make this determination without any help from the
TM.
2. A transaction was made available, and the EJB resumed it, but no
resources were actually enlisted, or perhaps resources were
enlisted
but not actually used, resulting in the same effect but relying on
the TM to provide this information.
I guess when you refer to a callback from Narayana, (2) is what
you're referring to? When would this information be available?
Maybe as some special result of suspending the transaction?
On 01/29/2016 12:27 PM, David M. Lloyd wrote:
That's an interesting idea. So in effect, the remote EJB would
tell the
caller "you sent me a transaction ID, but in the end, I didn't
use it"?
I would need to think about how this might work in the
presence of
multiple concurrent invocations on the same transaction.
Either way though, I think it would still be beneficial for
clients to
be able to explicitly annotate a client method (or otherwise
establish a
policy) such that it causes transactions to be propagated (or
not), or
to enforce transaction-related preconditions. The interceptor
that
implements this feature doesn't actually have protocol
awareness:
it
just examines the current environment, and decides whether to
attach the
transaction to the invocation context.
On 1/29/16 10:42 AM, Tom Jenkinson wrote:
One option that I would favour is to go down the JTS route
where the
subordinate calls back on the parent to tell it to register
it in the
transaction. This could be a new JBoss Remoting API that I
can invoke
from Narayana. The call would not necessarily be a remote
call, it would
invoke back into the JBR transport to tell it that when it
returns to
the parent it needs to enlist (or not).
On 29 January 2016 at 15:47, David M. Lloyd
<david.lloyd@redhat.com <mailto:david.lloyd@redhat.com>
<mailto:david.lloyd@redhat.com
<mailto:david.lloyd@redhat.com>>> wrote:
As you may know, WildFly supports a feature wherein an
EJB client
which
is invoking an EJB on a remote server has the option
to
propagate its
local transaction to the remote server, treating the
remote server
as a
subordinate and coordinating the transaction's
two-phase commit among
the resultant graph of servers. This feature has
always been
limited in
that, when enabled, transactions are always
propagated,
regardless of
the peer EJB's transaction policy, or of whether the
peer even has a
transaction manager.
So, for the invocation rework which I anticipate will
be included in
WildFly 11, I've introduced a new client-side
annotation intended
to be
associated with the EJB interface which informs the
client library
what
to do for transaction propagation for that interface.
In addition, I
intend to configuration strategies which will allow
the
default
mode to
be specified in various ways (per-thread, globally,
and
by target
interface/method all come to mind), for cases where
the
EJB's remote
interface cannot be easily modified for some reason.
I
expect to
also
broaden these configuration strategies to apply to all
client-side
EJB
interface/methods configuration items [3].
The first part of this change is the addition of a new
annotation
called
@ClientTransaction [1], which accepts as a value an
enum called
ClientTransactionPolicy [2]. The latter specifies
whether a local
transaction is required or forbidden for the method or
interface, and
also specifies whether the transaction is propagated
or
not
propagated.
I've added copious amounts of JavaDoc in order to
establish
exactly what
the behavior of each mode is, as well as to specify
how
each mode
interacts with the various modules that are configured
via the
standard
javax.ejb.TransactionAttributeType enum.
[1]
https://github.com/jbossas/jboss-ejb-client/blob/master/src/main/java/org/jboss/ejb/client/annotation/ClientTransaction.java
[2]
https://github.com/jbossas/jboss-ejb-client/blob/master/src/main/java/org/jboss/ejb/client/annotation/ClientTransactionPolicy.java
[2] (raw)
https://raw.githubusercontent.com/jbossas/jboss-ejb-client/master/src/main/java/org/jboss/ejb/client/annotation/ClientTransactionPolicy.java
[3] for a list, see:
https://github.com/jbossas/jboss-ejb-client/tree/master/src/main/java/org/jboss/ejb/client/annotation
--
- DML
_______________________________________________
wildfly-dev mailing list
wildfly-dev@lists.jboss.org
<mailto:wildfly-dev@lists.jboss.org>
<mailto:wildfly-dev@lists.jboss.org
<mailto:wildfly-dev@lists.jboss.org>>
https://lists.jboss.org/mailman/listinfo/wildfly-dev
--
- DML
--
- DML
- DML