[Cayenne-dev] Cayenne 3.0 cache RefreshQuery for NamedQuery

Sorry for not replying to this earlier. Per my comment on CAY-1465, I
suggest just map entity event listeners to do cache invalidation. I am
personally doing that everywhere.

Andrus

Thx, for your reply.

Now, I see only events related with Entity manipulations. Is there a
way to add listener for NamedQuery execution?

As I understand Event listeners could be configured in some "static"
section while Cayenne configuration initialization...
It's not so bad until you have ~30 cashed tables and almost 100-200
queries to that tables.
Problem that it's had to monitor that all Update queries got their
listener when this information is decomposed between Java Code and XML
with SQL Templates.

Idea was to create such architecture where "Persistent" (DBMS related)
layer is separate from Java logic and caching could be handled by only
map.xml changes.

Evgeny.



2010/7/20 Andrus Adamchik <[email protected]>:

Ok, now I finally see what you are saying (sorry for being so
slow :-)) Yes, a listener will only be notified of entity events and
running a SQLTemplate won't cause such an event.

So how do we solve it? I would like to avoid tying a "query execution
event" directly to cache refresh. Maybe instead we can attach this new
type of events (at the beginning EJBQL/SQLTemplate events) to the same
set of listeners as entity events? A listener can be defined to do a
cache refresh of a certain group, or do something else entirely. Even
the events can be defined in terms of specific entities. E.g. "query X
generates 'post-update' event for entity Y" ... or not :-) Something
to think about...

The note below is a bit of an aside from the main topic. Just a
general observation.... Interestingly, I have a somewhat opposite
problem in my apps - I have lots of ad-hoc queries created dynamically
in the code, and a number of different object modification flows
working over the same set of entities. So I need different listeners
for the same entity events across different apps, and as a result I am
finding that mapping "events" in XML is too inflexible, and I am more
in favor of in-code configuration of those. I guess at the end we'll
allow both styles, and allow extra XML files to be merged in the
shared mapping to address that.

Andrus

I am glad that it gave us some "food for brain"! :)

But I should remind that my issue wasn't about adding SQLTemplate events...
I just show that events can't solve this issue. But even if it could,
I'm not sure what I would prefer... configuring listeners on
SQLTemplate events or just put "removeCacheGroup" method invocation
after each "performNotSelectionQuery".
Second one seems even easier... and already available in 3.0...

One more thing... would we pay some CPU cost on SQLTemplate event
initialization? Like iterating through Collection of Listeners...
One business operation could generate ~ 8000 SQLTemplate queries. I
don't wish to lose here any performance!....
Actually I'm looking for any solution that could speed up them.... but
it's another topic...

Evgeny.

P.S.
And one more off-top :)
For me: Most advantage of Cayenne above so popular Hibernate is it
simplicity. For simple things it does simple actions! And doest it
fast!
Maybe this project should stay on this way, and not to follow it's
neighbor overloading every operation with lot's of features just to be
super flexible super universal....?
Please, don't think that I am trying to dictate you where to lead
Cayenne. Just some opinion :)

I totally understand what you are saying about simplicity and
certainly don't want to sacrifice any of that here. Along these lines
I am trying to avoid adding special cases of mapping that are not
generic enough, which actually makes things more complex in the end.
Both for us, as we end up with too many if/else cross-feature
interactions throughout the stack, and for our users, as there are too
many abstractions and no consistent picture of what the framework is
about. We tried hard to get rid of a few of those in 3.0, such as
"derived DbEntities" and will keep doing it in 3.1 (such as merging
EJBQL and SelectQuery). So I am being extra cautious now to avoid the
next round :-) At the same time this shouldn't stop us from trying
and experimenting with things of course...

So back to the issue at hand. Performance... Indeed, there is some
overhead looking up entity listeners. I think it will be much much
less with query listeners - a single HashMap lookup vs. recursively
scanning through the entity class hierarchy (which we should optimize
too by caching all listeners per class hierarchy). BTW, with listeners
you may even optimize your 8000 SQLTemplates per transaction scenario,
as a listener can batch groups to refresh and then refresh them at
once at the end of the transaction, instead of doing it 8000 times.

The second aspect is simplicity of mapping. Current entity listeners
mapping is not entirely intuitive (mainly because of its JPA roots
with insane hierarchical listener rules), and I'd like to simplify
that in the future. One angle of attack is to start using annotations
for callbacks/listeners, so that we don't have to map the method names
in the Modeler at all. For queries it will be simpler even without
annotations, as initially we will have a single event type ("post-
commit"), and you will have to simply specify the listener class (or
multiple classes) per query and that's it.

So do you think this is still too complex? (This is an honest
question... when I looked closer at the problem, I don't feel that
much opposed to your original solution, just trying to see if we can
reuse an abstraction that we already have).

If all queries invalidate the same cache group, then yes, you should
totally use it.

Andrus

I'm totally (100%) against losing Performance even on 5% even on 1%!!!
So I am against adding listeners here....

Not sure if I understand you... If you are talking about group caching
per call(transaction) - it's not my case. Queries are mostly different
and even same queries could/should return different results because
there could be several inserts/updates between.... Try to catch this
problem in 32000+ line log :) Really not ready to try it...

My caching in first way will be used only for dictionary tables in
SHARED cache and not refreshed too often (not more then once per day).
In second I may try Context cashing for tables that are not changed
per operation(transaction). But it's not near feature optimization....

Honestly still don't wish to use listeners... in my case it will be
heap of listeners at one place. Even it will be several heaps (from
several modules) which should be collected in one.... Sory can't find
any place for this feature.... :(
For me it's much better to have such things at local places in code
(like it's can be done now) or in SQLTemplates (like was in my idea).


May be I will try explain my point of view.... So you will see it with
my eyes :)
Imagine:
1) You are going to Modeller
2) Creating simple Row SQL
3) See options about cache: NO Cache, LOCAL Cache, SHARED Cache

Cool! But it's actualy only for SELECT queries. So what with
INSERT/UPDATE/DELETE queries (that could be also in template)?
For INSERT/UPDATE/DELETE queries such options have no sense.
But.... They could invalidate cache!!!.. It's actually what they
usually does... Why not to extend this options for
INSERT/UPDATE/DELETE in reverse effect...?
So on SELECT we are caching on INSERT/UPDATE/DELETE we are refreshing
cache... for me it looks very intuitive and logically......


If still not looks logically for you, then it was bad idea.... )



Evgeny.