Hi Florent, hi Jens,
Would that be a fair summary of your posts:
- We keep the path cache.
- We add a revalidation flag
(to the OperationContext?).
- We add an expiration time to the path-to-id mapping
(controlled by a session parameter?).
- I would like to add:
We add a general expiration time to objects in the cache
(controlled by a session parameter?).
If we agree on that I would rework the cache implementation.
@Florent: You can already deactivate caching for a getObjectByPath()
call by providing an OperationContext that has the "cache enabled" flag
set to false. That is not obvious, though.
@Jens: Whenever you call refresh() and the object is gone or you lost
the permission to see it, it will throw an exception. We can't change
the fact that the cache might return stale objects. Even if you load a
fresh object from the repository, it might have been changed on the
server a second later ... and we don't not know that until we load it
again or try to change it.
Cheers,
Florian
On 16/11/2010 13:45, Jens Hübel wrote:
HI Florian
Yes I agree to all your thoughts, but my idea was that the case where the
object changes but the path keeps stable is one that may more weird than some
of the others you mention. As those are not opencmis specific I am not sure if
we should promote some of your thoughts to the OASI TC...? In addition the
change token is also opaque to the client and maintained by the servers. This
means that the client cannot make any assumption about its meaning. My thought
was that if we reduce this just to check for equality might help in some cases.
But some of your thoughts go beyond the "path is no stable id" problem and are
inherent with caching in general. What is if I get an object by id, cache it and the ACL
changes so that I do not have access any longer? There is some risk that a property has
changed. What happens if the object gets deleted in the server?
We really need to carefully document the behavior of our client lib here at
least.
One pragmatic solution would be to cache every object with a timestamp when it
got cached. If the object is accessed from the cache we might set a
(configurable) timeout. If the timeout is exceeded we always refresh the object
from the server. This would give at least some guarantee that a stale object
will only live for a certain period of time (lets say 30mins or so). For highly
sensitive scenarios this timeout might be reduced or set to zero. The timestamp
should be associated with the key (e.g. path) and not the value of the cache
(object).
I fear there won't be a perfect solution as you already said...
Jens
-----Original Message-----
From: Florian Müller [mailto:[email protected]]
Sent: Dienstag, 16. November 2010 11:40
To: [email protected]
Subject: Re: getObjectByPath cache problem
The change token can only be used to detect changes within an object.
The problem here is that we are potentially dealing with two objects.
The root of the problem is that a path is not a stable key for an object.
An object can be updated in the repository without our knowledge. The
cache would then return an outdated object and everybody should be
prepared for that. refresh() reloads the current state from the
repository. That works fine if the object is retrieved through
getObject(). You can move the object around, unfile it, put it in
multiple folders and it still works. The object id is unambiguous.
The cache currently maps object paths to object ids. When you call
getObjectByPath() it will look up the id for this path and gets the
object from the cache. If you move the object to a different folder,
getObjectByPath() shouldn't find it anymore. The path of the object has
changed and the old path is now invalid. Note that the object hasn't
changed and therefore the change token hasn't either.
Since there is no notification from the repository, the path-to-id
mapping can't be corrected. The cache still thinks the object is
accessible through this path. So getObjectByPath() returns the object
although it should throw a CmisObjectNotFound exception.
Let's assume we create a new object in the place where the old object
was. The new object can now be accessed with the old path. Since the
outdated path-to-id mapping is still in place, getObjectByPath() returns
the old object and not the new one -- which clearly wrong.
The problem that we are facing here is that there is no reliable way to
keep the path-to-id mapping up-to-date. If we want to be correct, we
would have to ask the repository for the current id for the given path
every time getObjectByPath() is called (3) -- or not use the cache at
all (2).
- Florian
On 16/11/2010 08:23, Jens Hübel wrote:
Shouldn't the change token solve that? How do we deal with the change token for
other objects that are in the local cache? Ignore? Check on each access?
Configurable?
Jens
-----Original Message-----
From: Florian Müller [mailto:[email protected]]
Sent: Montag, 15. November 2010 21:17
To: [email protected]
Subject: getObjectByPath cache problem
Hi all,
I had another look at [1]. Unfortunately, it's an unsolvable problem.
I can think of three ways to cope with it:
1. We leave it like it is, although it is very, very confusing when you run
into this situation.
2. We don't cache by path. How would that affect applications?
3. When getObjectByPath is called we fetch the object id from the repository
and then get the object from the cache.
In the worst case, we would have to hit the repository twice.
Any opinions?
- Florian
[1] https://issues.apache.org/jira/browse/CMIS-260
