[
http://opensource.atlassian.com/projects/hibernate/browse/HHH-3910?page=c...
]
Shawn Clowater commented on HHH-3910:
-------------------------------------
Steve, you have impeccable timing :D
I'm about to go back and see if I can squeeze out some extra performance during
flushing. I'm not convinced I can't get it done by just providing a custom impl
in my persister but the dirty check didn't seem to be as big a factor as I had thought
in the whole flushing scheme (I haven't quite fully profiled the time yet but the
dirty check was about 20% overall from what I had found in my testing).
With that aside, here's our scenario. We're already keeping track of modified
properties (for our auditing) so we figured we should be able to use that for our dirty
check. Our properties are tracked in a map via their names with their original values.
So ultimately rather than having to check equality on n number of properties using
reflection the thought is that I can bypass that altogether for 'clean' entities
if their map is empty. I had a run at this a few years back and there was one case where
an entity was getting updated and setting a field directly bypassing our audit and
slipping through the cracks.
We've got some monstrous processes and sometimes get hammered with flush times so
we're looking at squeezing out whatever we can (as well as trying to be smart with
batches and clearing the sessions, etc). I had just seen Ovidio struggling with the same
thing a few years back and was hopping on the bandwagon to see if we could eke anything
out.
I may very well roll our environment up to 4 to see if perf is any better, I think I saw
some initiatives to optimize cascade flushing but I may as well tune on what we'll
eventually be running on. Just not enough hours in a day :D
Add support for custom dirty checking during flush
--------------------------------------------------
Key: HHH-3910
URL:
http://opensource.atlassian.com/projects/hibernate/browse/HHH-3910
Project: Hibernate Core
Issue Type: Improvement
Components: core
Affects Versions: 3.3.1
Reporter: Ovidio Mallo
Attachments: DirtyCheckFailedAttempt.patch
Currently, Hibernate supports a special dirty checking on instrumented entities
in order to improve the flush performance. IMO, this optimization can often be
rather significant. However, the drawback is that you have to use bytecode
instrumentation in order to take advantage of this performance improvement which
might not be an option in some projects.
Therefore, I wanted to propose to extend the current dirty checking during flush
in such a way that the dirtyness information can also be directly provided by
clients. Thereby, I could think of two possible approaches to do this:
1. Introduce an interface which client entities might implement in case they
have some notion of dirtyness. The interface could look something like:
public interface DirtyAwareEntity {
boolean getMightBeDirty();
void setMightBeDirty(boolean mightBeDirty);
}
Using such an interface, Hibernate could easily check whether an entity might
be dirty during flush and it could also reset the dirty flag after flush just
as is currently done for instrumented classes. So this approach would probably
be rather easy to implement and very convenient for clients since they would
only have to implement that interface on the appropriate entities and set the
dirty flag when the entity is actually modified.
2. Add some hooks on event listeners and/or on the Interceptor for querying whether
an entity is dirty and for resetting the dirty flag. E.g. one could add the
following hook method to the DefaultFlushEntityEventListener class:
protected boolean requiresDirtyCheck(FlushEntityEvent event);
By default, this method would call EntityEntry#requiresDirtyCheck(Object entity)
as is done right now.
Resetting the dirty flag could maybe be done in Interceptor#postFlush() or some
dedicated method could be provided.
BTW, I know that currently there already is the Interceptor#findDirty() method which
already allows for some custom dirty checking but the problem from a performance
point of view is that this method requires the entity's property values as parameter
which are retrieved in DefaultFlushEntityEventListener#getValues() which is the most
expensive method during flush. This drawback of the findDirty() method has often been
noticed in comments on the news groups.
I personally think it would be nice if something could be done to improve the
performance of flushing in Hibernate since from what I read on the news groups and
the like, flushing still seems to often lead to performance problems in practice,
especially in larger projects where it is often not easy to avoid flushes or to
keep the numer of entities in the session cache small. In fact, we are having quite
some trouble with that in our project and having some custom dirty checking like the
one I'm proposing here would greatly help in our project and in other projects as
well, I guess.
--
This message is automatically generated by JIRA.
For more information on JIRA, see:
http://www.atlassian.com/software/jira