Hi Norbert, Thank you very much for your extensive answer.
For starters, we would be happy to have an option available for the course grained handling that you mention. I'd be interested to hear if there are any options besides GemStone available to be used with Voyage/MongoDB. Your mention of the Mongo cache being a weak dictionary is very interesting, I did not know that (it doesn't seem to be documented). Yes, the Seaside sessions seem to be the path through which these objects remain referenced. When I inspect all instances of our session class, I see old session objects hanging around that still reference our root application component with the rendering tree containing various references to our domain model objects. Can you tell me why the session objects don't get garbage collected? Should we manually clean up our session object somehow? I must admit I'm a bit out of my comfort zone here, after working in Java for close to 20 years ;-) We explicitly don't want to use any relational databases for our own application (perhaps only for integrating with customer data). I still haven't fully integrated the conceptual differences in my mental model, about how to work with objects in an image based/object persistence environment. I did look into using a Voyage repository filter, which is mentioned (or buried deep I should say ;-)) in this post: https://pharoweekly.wordpress.com/2017/02/20/consortium-action-13-17-feb/. What if we were to use a dynamic variable like this for each session instance? Then at least each user will have her/his own cache. But that doesn't answer the need for a kind of rollback mechanism within the same Voyage cache/session. And then there is your comment about potentially having multiple copies of an identical object... I wonder what patterns other Seaside 'enterprise' application developers are using. You are right, the term object transactionality doesn't make much sense ;-) We are not using Magritte (and probably won't). I don't know much about Magritte, but it feels like it might be incompatible with the behavioral, 'pure' object approach that we want to use. However, I am interested to investigate. Can you recommend any good documentation sources for learning Magritte from an application architecture perspective? Also, I'm interested to hear more about the modification tracking approach you are working on. Please drop me a personal note if you are willing to collaborate on this. Kind regards, Jonathan van Alteren Founding Member | Object Guild jvalte...@objectguild.com On 8 Oct 2019, 12:54 +0200, Norbert Hartl <norb...@hartl.name>, wrote: > Hi, > > > Am 08.10.2019 um 12:05 schrieb Jonathan van Alteren > > <jvalte...@objectguild.com>: > > > > Hello all, > > > > We are having some issues with using Voyage/Mongo for a customer project > > that I'd like to get your feedback on. > > > > The customer application is a form based business web application using > > Seaside with object persistence using Voyage with MongoDB on Pharo 7.0. The > > application is deployed on a dedicated Linux production server running > > MongoDB version 4.2. > > > > The application is used to manage meeting agendas and minutes. After > > opening the agenda view of a future meeting, the user can add an item to > > the agenda by clicking a button. This calls an item editor component which > > answers when a Save or Cancel button is clicked. The agenda view component > > itself also has a Save button, which performs a Voyage save of the object > > aggregate (agenda + items). > > > > We've encountered an issue where a user makes changes to an agenda, but > > does not click the Save button. Instead, the user closes the browser or > > uses the navigation to go to a different part of the application. When > > navigating back to the original agenda, the changes made previously (e.g. > > items added) are still being displayed, even though they were never > > explicitly saved. > > > > It does not matter if we select the agenda aggregate object instance using > > Voyage or access it in a different way. Changes to the state of the object > > are retained, even though a Voyage save was never sent to the agenda > > instance. The cause seems to be that the Voyage repository caches the > > object instance and thus on select, it returns an object that is in a > > different state than how it was persisted. > > > > This all seems to come down to object transactionality. > > > > We have a need to 'cancel' changes made to an object instance. Before > > working with Pharo/Smalltalk in a non-image based environment, I was used > > to do this by retrieving the original object from persistence again. This > > also allowed for a convenient way to detect changes to an object's state, > > which we are missing at the moment too. > > > > We know that moving to GemStone can help us with these issues, but our > > current planning of customer projects does not allow us to do this within > > the next 3 months. And we really need to find a way to tackle these issues . > > > > > > Your feedback is greatly appreciated! > > > this is none to be confusing to a lot of people. If you map a memory object > graph to a database there are no intrinsic points of transactions unless you > put them into your application model. Gemstone or any transaction based > handling won‘t help you if your use case is not as coarse grained as > discarding all modified data and not just some. Or you need to hop into > nested transactions which are supported by Gemstone IMHO but there is rules > to care about, too, if you want to use them. > > Regarding your application you modify data and this is kept. The cache in > mongo is a weak dictionary. This means that if you get the object with the > changes a second time it means these objects are referenced somewhere. > Seaside sessions seems to be the obvious thing here. If you have a reachable > object (e.g. server singleton -> handler -> seaside -> seaside session -> > business object) then mongo keeps the object in its cache because it needs to > be able to establish identity on further queries. Ein emoving all the changes > from an object you keep in memory wouldn‘t be really good if the persistence > layer would do it. > > The problem is obvious but the solution is not. We could make the cache > session or request based which would circumvent the problem. But if an object > would be attached to something outside the session/request which you cannot > forbid it causes real problems like multiple copies to an identical object. > > I don‘t know what you mean with object transactionality because it makes no > sense in my head. But you have basically two approaches to solve the problem. > If you would use magritte you would have it at hand. There each UI component > gets just a memento of the business object for modification. You need to > commit the UI component in order to modify the business object. So you can > commit on user interaction and separately on persistence which allows a huge > set of use cases. > The other one I‘m working on since a while is a proper modification tracking. > In this you change the objects directly but have a registry with the dirty > objects which you can act upon. I have a sample modification tracker that > uses readOnly object write barrier for objects. On write attempt it registers > the object and the write operation as command pattern. You can undo > modifications of a single object where all commmands are undone on that > object. Or you abort all modifications at once. It is basically working but > has the problems on concurrent access. As the mongo cache is image wide two > concurrent requests interfere with each other. > So you can see that the problem is not that easy to solve. > > TL;DR a memento approach might your best and cleanest option right now. > > Norbert > > > > Kind regards, > > > > > > Jonathan van Alteren > > > > Founding Member | Object Guild > > jvalte...@objectguild.com > > > >
