Re: [xwiki-devs] [proposal] Introduce a new persistence engine.

On Thu, Jan 20, 2011 at 17:15, Caleb James DeLisle &lt;calebdelisle(a)lavabit.com > > > On 01/20/2011 02:42 AM, Denis Gervalle wrote: >> On Wed, Jan 19, 2011 at 19:54, Caleb James DeLisle < > calebdelisle(a)lavabit.com >>>> >>> >>> >>> On 01/19/2011 12:13 PM, Vincent Massol wrote: >>>> Hi Caleb, >>>> >>>> I see you're excited, that's good! :) >>>> >>>> Some general comments: >>>> * This looks more like a design for a transaction module than for a >>> persistence engine. I don't see anything related to persistence in your >>> proposal below. Your proposal could work on stuff others than storage, >>> right? >>> Yes, this proposal only covers the transaction sub-module of the >>> persistence engine. The so far >>> un-proposed modules include xwiki-store-serialization, >>> xwiki-store-filesystem and a legacy >>> attachment storage module: xwiki-store-filesystem-attachments. >>> >>> Anything which requires transactions could use the TransactionRunnable >>> although I'm at a loss to >>> think of anything other than storage which would require transactions. >>> >>>> * I was expecting to see some Store/Storage/Persistence interfaces with >>> proposed APIs and explanation on how they could be implemented both with >>> Hibernate and JCR for example. And the relationship with the proposed > new >>> Model defined. >>> I don't like to propose an interface until I have tried to implement it. >>> Also I do not like to >>> propose an implementation until I have tried to use it. At this point > it's >>> far enough off that I >>> would rather wait than propose APIs blind. >>> >>> My experience with attachment store has shown that what we want is a set > of >>> functions which provide >>> TransactionRunnables to do various things: >>> aka: >>> TransactionRunnable<T> getDocumentSaveTransactionRunnable(XWikiDocument >>> toSave); >>> In a hibernate implementation it would return >>> TransactionRunnable<HibernateTransaction> and in a JCR >>> it would return TransactionRunnable<JCRTransaction>. >>> >>> We cannot have APIs like this until TransactionRunnable is agreed upon >>> these will return instances >>> of it. >>> >>>> * I was also expecting a strategy defined to migrate users from the >>> current implementation of the storage to the new one >>> IMO we should change the persistence engine and implement the same > schema, >>> once the persistence >>> engine is rebuilt, then we can consider modifying the schema. The schema > is >>> a specification, it may >>> not be perfect but it is something to comply with. It is important to me >>> that a we prove that a new >>> persistence engine is able comply with existing specifications before we >>> start designing new ones >>> around it. >>> >>>> >>>> I noticed some discussions between Denis and you on IRC about all this. >>> Does you latest findings change the proposal below? >>> Everything proposed still holds true but I did add 2 new features. >>> >>> 1. There is a way for a TransactionRunnable<DBTransaction> to be passed > an >>> instance of DBTransaction >>> using a new method called getContext(). >>> >>> 2. There is a new class which serves what I believe is an edge use case. >>> Suppose you want to define a TransactionRunnable (we will call it >>> YourTransactionRunnable) which >>> must run inside of a DBTransaction but it must _also_ run after an > instance >>> of MyTransactionRunnable. >>> You can make MyTransactionRunnable a >>> "ProvidingTransactionRunnable<DBTransaction, MyInterface>" and >>> then MyTransactionRunnable must run inside of a DBTransaction and we > define >>> YourTransactionRunnable >>> as a TransactionRunnable<MyInterface>. This also allows >>> MyTransactionRunnable to share information >>> since YourTransactionRunnable.getContext() will provide an > implementation >>> of MyInterface. Of course >>> this feature must be used with care as it provides the tools to write >>> horrible constructs but IMO it >>> is the type of feature which when you need it, there is no other way >>> around. >>> >> >> I just need to add that we have not been able to apply TR and PTR in >> particular to a simple store with a JDBC connected database. From our >> experience, the PTR cause more issue than it solve and we have to find a >> better way to convey datas between transactions of a given transaction >> chain. Data dependencies between transaction could not be easily solved > at >> compile time since there is many combinatory situation in real life. >> >> My suggestion would be to provide access to previous transaction using a >> hash of Interfaces exposed by previous transaction. Checking availability > of >> needed interfaces could be done by transactions à preRun time which could >> avoid the need of uselessly running the whole chain, if there is a >> dependency problem. > > What you can do is make DBTransaction (the context) extend Map and then put > things in that map. I > really don't like that path though since is means that TR3 uses things > provided by TR1 and doesn't > declare that need. It breaks atomicity by sharing data between > transactions, TR3 can't just be > plugged in somewhere else where TR1 is not used. It also breaks breaks > compile time safety, plugging > TR3 in somewhere else (or reordering them) will compile but fail to run. > That pattern doesn't really afford you any of the benefits of TR other than > the try, catch safety > so, to me, it makes sense in that case for the code for TR1, TR2, and TR3 > to all be inside of the > same TransactionRunnable. > >> >> I also doubt that the way of multiple transactions are executed in a same >> single one is correct or useful. Would really prefer to see all > transactions >> run at a given level before going down the chain. > > I don't see this as making sense. My goal is to make sure when something > goes wrong it fails as > early as possible so there is the least amount to rollback. > > SaveDocumentRunnable > | > +--SaveObjectRunnable > | | > | +--SavePropertyRunnable EXCEPTION > | +--SavePropertyRunnable <-- Doesn't run. > | > +--SaveObjectRunnable <-- doesn't run > | > etc. Doesn't run. > > I'm interested to know what is the rationale for wanting them to run one > level then the next? > IE: save document, then all objects then all properties. > >> This would provide the way >> to bundle transaction (and even transaction chain) together by running > them >> in a single transaction. This would also helps not mixing dependencies > since >> currently there is an implicit availability of earlier sibling chain of >> transactions that does not fit the idea of a context in evolution checked > at >> compile time and could provide unchecked implicit dependencies. > > I still don't quite follow, can you give a real world example? > Well, let me base my real world example on your sample above. Imagine that documents are linked to their objects using their primary key, PK which is generated by the underlying database when you records the document for the first time, and that we needs creating a new documents with Objects in a single transaction. Your SaveObjectRunnable will need that PK as a FK, how does it get it? Or said in another way, it needs a context that provide the PK of the document for which it is saving the objects.

Well, lets insert a CreateDocumentRunnable before it, that may have some lower TR to complete, like saving document and retrieving the generated PK. How does this CreateDocumentRunnable is maybe a sibling of SaveObjectRunnable provide the PK to it ?

How do you check that SaveObjectRunnable requirement to have that PK is fulfilled ? Have you an idea of the structure needed for this example ?

Denis