One of the fortunate things of being apart of the Oracle ACE Director program is behind the scenes, we and Oracle staff are often chatting and helping each other out, which I'm incredibly appreciative of. In this case I must raise my hat to Steven Davelaar and John Stegeman for their out of hours assistance in one form or another.
Regarding my last post, John made a reasonable point that I was making/drawing my conclusions far too early, that in fact I needed to test the complete task flow transaction life cycle to see if the behaviour of the task flows has changed in 18.104.22.168.0. In particular John's good memory led him back to this OTN forum post by Steve Muench that stated:
In fact, in a future release we will likely be changing that implementation detail so that the AMs are always used from their own AM pool, however they will share a transaction/connection.Steve's point from late 2010, and the one that John was re-affirming is that even though the underlying implementation may change, the end effect from the Bounded Task Flow (BTF) programmer's point of view, everything can still work the same. And this is what I needed to check, looking at what AM methods are called is not enough. I needed to check the actual database connections and transaction behaviour.
From my post I was concerned without auto AM nesting, a page comprised of several BTFs with separate AMs would spawn as many database connections compromising the scalability of the application. From the logs I thought this was the case as I could see two root AMs created and a separate (ie. two) call to prepareSession() for each. My assumption being this meant two connections were being raised with the database under 22.214.171.124.0, where alternatively using 11.1.1.X.0 is was only one database connection using the auto AM nesting feature.
However a query on the v$session table in the database using the 126.96.36.199.0 solution:
SELECT * FROM v$session WHERE username = 'HR';
...showed only 1 connection. So regardless that there is 2 roots AMs instantiated under 188.8.131.52.0, they share connections (and therefore transactions too). In other words, while the end result is the same, the underlying implementation has changed.
I don't have a snazzy name for this new implementation vs the older auto AM nesting, so I figure we should call it Bruce to keep it simple (with apologies to Monty Python).
The only discrepancies between implementations we can see is that the prepareSession() and similar AM methods that deal with the connection or transaction (eg. afterConnect(), afterCommit()) are now called on the secondary AM as it's treated as a root AM rather than a nested AM. This was not the behaviour under 11.1.1.X.0 as nested AMs delegate that responsibility back to the root AM. This in turn may cause you a minor hiccup if you've overridden these method in a framework extension of the AppModuleImpl as they'll now be called across all your AMs, including those who used to be auto nested.
In returning to Steve Muench's point:
In fact, in a future release we will likely be changing that implementation detail so that the AMs are always used from their own AM pool, however they will share a transaction/connection.Via http://localhost:7101/dms/Spy I've verified this is the case with Bruce, where under 11.1.1.X.0 is used to be a single AM pool, but now under 184.108.40.206.0 there is 2 AM pools & 1 defined connection. The end effect and my primary concern from the previous blog post is now mute, the scalability of database connections is maintained. Bruce is a winner.
The interesting change under 220.127.116.11.0 is the 1 AM pool vs many AM pools. Ignoring if at design time create nested AMs under the 1 root AM, traditionally with the runtime auto nesting AM feature you'd still have 1 root AM pool. Now for each root AM you'll end up with an associated AM pool. If you're app is made up of hundreds of AMs that were nested and used the AM pool of their parent, you'll now end up with hundreds of AM pools. Whether this is a problem is hard to tell without load testing an application with this setup, but Steve Muench does comment in a follow up to the OTN forum post "The AM pool in an of itself is not additional overhead, no."
So potentially the midtier is less scalable as it needs to maintain more pools and process them, to what degree we don't know. Yet, more importantly, we now have a relatively more flexible solution in terms of tuning the AM pools, because previously
it was an all or nothing affair with 1 AM pool under the auto AM nesting 11.1.1.X.0 approach (again ignoring design time nesting AMs under the root AM), and now with Bruce we've lots of find grained AM pools to tune. As such each pool can be tuned where an AM pool for a little used BTF can be set to consume less resources over a BTF with an AM pool that is hit frequently.
So, besides a couple of minor implementation changes (and if you find more please post a comment), it looks like Bruce is a winner.
Again thanks must go to John Stegeman for his assistance in working through his issues, and Steven Davelaar for his offer of support.