Re: Concurrency question
Re: Concurrency question
- Subject: Re: Concurrency question
- From: Ken Anderson <email@hidden>
- Date: Thu, 2 Aug 2007 08:41:39 -0400
This is pretty amazing Mike.
I'm having this exact problem right now, and have been banging my
head against the wall, saying "THIS IS NOT POSSIBLE!". Apparently...
it is.
Now that the problem is identified, is there a reasonable solution?
Ken
On Aug 2, 2007, at 7:58 AM, Mike Schrag wrote:
My reply was too big with the attachments, so I moved my response
into the wiki:
http://wiki.objectstyle.org/confluence/display/WO/
Programming__WebObjects-EOF-Using+EOF-Problems
The last one, "Strange Locking Problems" ... And for the lazy:
It would appear that there is, in our opinion, some bugs related to
optimistic locking within a single EOF stack. Essentially what it
boils down to is that it appears that the update database operation
that is created as a result of a call to .saveChanges() is backed
by the EODatabaseContext snapshot and NOT the "working" snapshot
inside in the EO in the editing context it came from. What this
means is that while changes are not merged until you .unlock()
and .lock() under normal circumstances, because the underlying
snapshot that EOF diffs your changes against on save is the DBC
snapshot, it's effectively inadvertently "merged" on commit. That
is to say that if another EC makes changes and saves, then you make
different changes and save, you will blow away their changes with
no sign of an optimistic locking exception because your snapshot IS
their snapshot now (meaning, it looks like just you are overwriting
their changes, versus the reality of the situation that you are
actually conflicting with their changes). After discussing this
some, we believe that if the update operation used a version of the
EO's backing snapshot instead that these weird behaviors would be
fixed and it would behave exactly like a normal conflicting update
if you were in two EOF stacks. The current behavior smells of bug,
but I'm curious if anyone a dissenting opinion on the topic. It's
certainly really complicated and nasty down in that code, so it's
possible there's some crazy justifiable reason for it.
Go to the wiki for the diagrams.
ms
On Aug 2, 2007, at 1:12 AM, Chuck Hill wrote:
The problem comes in when the modifications are made and saved
after your editing context has been locked in the RR loop. I
tricked ;-) Mike into looking at this today with. Looks like
tigers lurk here. Maybe Mike will comment.
Chuck
On Aug 1, 2007, at 2:34 PM, Pierre Bernard wrote:
You can simulate OL, by listening to merge notifications. If it
affects a modified object you can later on refuse to save.
Pierre
On Aug 1, 2007, at 7:18 PM, Chuck Hill wrote:
On Aug 1, 2007, at 5:00 AM, Miguel Arroz wrote:
Hi!
The contexts are locked. The problem is that it's not the
same context - it's a different context per thread, with local
copies of the same objects.
Synchronizing solves the sample problem, but as my real
problem is much more complex than this example, it starts to
get a little... ugly. Also, I still did not full understand
why, the objects in the thread that runs in second are not
updated with the data saved by the thread that run first. I
suspect that the cause of this is that the objects only receive
the notifications to update themselves after finishing the R-R.
Can anyone confirm this?
Yes, if you are locking properly (and I know you are), the
changes only get merged at the end of the RR loop. You can do
it yourself by unlocking and relocking the EC.
About being solving 2 different problems or not, it depends
on the level os abstraction you use to look at it. :) From my
point of view, I'm solving one problem - the problem of
concurrent updates. My app has now enough info to solve the
problem (OL fields, and how to retry). It would work *if* the
UPDATE WHERE data was fetched from the original data in the
context, and not in the row data.
It gets more complex than this. One quick fix is to have each
EC in its own EOF stack. But that can be rather memory expensive.
Chuck
On 2007/08/01, at 03:03, Ken Anderson wrote:
Miguel,
Is your editing context locked before calling incrementIt() ?
I would think that would solve your concurrencyssue her
If not, just synchronizing the method should solve the
problem. It may seem inelegant, but you really ARE solving 2
different problems...
Ken
On Jul 31, 2007, at 7:41 PM, Miguel Arroz wrote:
Hi!
I'm trying to understand what's the best way to do
something here.
Imagine that I need to get a object from the database,
modify some attribute based on itself and save it again.
So, we have the method:
public void incrementIt() {
if( aIsEven() ) { // 1
setB( b() + 1 ); // 2
}
setA( a() + 1 ); // 3
editingContext.saveChanges(); // 4
}
Well, it's easy to solve the problem of update conflicts
between two different instances of the app. Just tick the OL
lock for a and b fields, catch the evil expression, refault
the object, and recalculate (and retry to save). That's "easy".
Now, my problem is inside the same instance! Imagine that
this method runs at the same time and we have the following
run order, for threads X and Y, with the same object in two
different contexts (and imagine a = 3):
X 1
Y 1
X 3
X 4
Y 3
Y 4
This will produce wrong results, but it won't cause any
locking exception. Why?
1) Both threads get the object with a = 3.
2) Both threads do not run line 2 because 3 is not even.
3) The thread X increments a, and saves it. When saving,
the object at thread Y will have it's 'a' attribute updated,
assuming both objects are in the same coordinator.
4) The thread Y increments a again, and saves it. N
optimistic locking exception will be thrown, because the
coordinator snapshot was updated in the last commit, so the
SELECT FOR UPDATE will run OK.
This will cause a to be 5, but b did not increment as it
should. The problem is that, when saving, we are basing OL on
the row snapshot (that is updated during the process) and not
to the original value of the object when it was loaded into
the editing context.
Well, this may be solved using the "classic" Java
"syncronized" stuff, and locks and all that stuff. But this
is a bit stupid. I already solved the problem with OL for
the, theoretically, more difficult case of managing several
app instances. Do I have to solve it all over again, in a
different way, to deal with multiple updates on the same
instance? Isn't there a way to use OL just like I'm already
doing?
Yours
Miguel Arroz
Miguel Arroz
http://www.terminalapp.net
http://www.ipragma.com
_______________________________________________
Do not post admin requests to the list. They will be ignored.
Webobjects-dev mailing list (Webobjects-
email@hidden)
Help/Unsubscribe/Update your Subscription:
@anderhome.com
This email sent to email@hidden
Miguel Arroz
http://www.terminalapp.net
http://www.ipragma.com
_______________________________________________
Do not post admin requests to the list. They will be ignored.
Webobjects-dev mailing list (email@hidden)
Help/Unsubscribe/Update your Subscription:
40global-village.net
This email sent to email@hidden
--
Practical WebObjects - for developers who want to increase their
overall knowledge of WebObjects or who are trying to solve
specific problems.
http://www.global-village.net/products/practical_webobjects
_______________________________________________
Do not post admin requests to the list. They will be ignored.
Webobjects-dev mailing list (email@hidden)
Help/Unsubscribe/Update your Subscription:
email@hidden
This email sent to email@hidden
- - -
Houdah Software s. à r. l.
http://www.houdah.com
HoudahGeo: One-stop photo geocoding
HoudahSpot: Powerful Spotlight frontend
--
Practical WebObjects - for developers who want to increase their
overall knowledge of WebObjects or who are trying to solve
specific problems.
http://www.global-village.net/products/practical_webobjects
_______________________________________________
Do not post admin requests to the list. They will be ignored.
Webobjects-dev mailing list (email@hidden)
Help/Unsubscribe/Update your Subscription:
40mdimension.com
This email sent to email@hidden
_______________________________________________
Do not post admin requests to the list. They will be ignored.
Webobjects-dev mailing list (email@hidden)
Help/Unsubscribe/Update your Subscription:
40anderhome.com
This email sent to email@hidden
_______________________________________________
Do not post admin requests to the list. They will be ignored.
Webobjects-dev mailing list (email@hidden)
Help/Unsubscribe/Update your Subscription:
This email sent to email@hidden