I think the first two are definitely bugs. You lost me on the
discussion of 2128...
I'd like to see the discussion included on OpenEJB dev, also...
--kevan
On Jul 17, 2006, at 4:47 PM, Matt Hogstrom wrote:
I opened 3 JIRAs that affect CMP deployment, Session bean
performance and consistency.
The JIRA's in question are:
GERONIMO-2127 - Expose ability to use SELECT FOR UPDATE
GERONIMO-2129 - Allow user to specify pool size on Stateless
Session Beans
GERONIMO-2128 - Allow user to specify an Isolation Level on a CMP Bean
I think these items can be argued in two ways. Bugs and features.
Based on my experience with CMP beans what we have for Geronimo is
fully compliant with the J2EE specification. We pass the tests and
are compliant.
However, our current implementation does not allow for a user to
deploy an application that will ensure data consistency. For
instance, if one is using Oracle as the database, which operates at
Read Committed by default, two competing applications will possible
overlay data from one another with no notification at all. In
order to properly provide for ACID properties a SELECT FOR UPDATE
needs to be provided so one application can block another. I
consider this a bug since even though the implementation "is
compliant" it is also unusable unless your data is read-only or you
can guarantee no conflicts in some other way.
The second issue goes to consumability as well as accuracy.
Stateless session beans are traditionally used as facades and
wrappers for Tx. They are also used to store information that is
transient but expected to be longer lived than a single use. The
SLSB in OpenEJB has a pool size of one and will make some
applications perform poorly and perhaps malfunction.
In the case of SPECjAppServer it will do both. The SLSB is used in
that application as a temporary cache for keys used to insert into
a database. The current behaviour is that on every request a new
block of keys is retrieved from the Key database. For SPECj and
DayTrader it results in deadlocks and collided keys. The Pool
(which does exist but is fixed at a size of 1) will eliminate this
problem.
2128 is similar to 2127 in so far as using database isolation to
provide ACID properties it allows for multiple Isolation levels to
be used such that RDBMs such as DB2 or Derby can perform better.
Although this isn't required for ACID properties it does require a
schema change to OEJB 2.1.1.
All of these JIRA's can be implemented without disrupting current
applications so I believe we should include them in 1.1.1.
The changes are actually limited to OEJB and TranQL which are
components of G.
My vote is to include these JIRA's.
Matt
