It's probably possible to do it completely in Clojure, but you have to subclass Atom. There's no need for any transaction boundary; you just have to make sure that compareAndSet does a durable swap.
My plan was to get durable refs done and then extend the other mutable identities, including atom. I'd love to work with you on it! On Sep 23, 8:27 am, Per Vognsen <per.vogn...@gmail.com> wrote: > Cool! > > I'm getting back to Clojure after an extended absence. Just today I > was pondering the design of a solution to a similar problem, though I > suspect our requirements diverge on several points. My tentative > conclusion was that it could be done entirely in Clojure and without > modifying existing code. Maybe you can poke holes in my fledging plan > since you've obviously been thinking about this sort of problem longer > than me: > > There's a new pref reference type. It consists of a key and an atom > containing nil for unloaded objects and an STM reference for loaded > objects. > > When a pref is dereferenced, it checks its atom. If nil, it first > loads the object from disk into a fresh STM reference (which has a > metadata field pointing back to the pref) and mutates the atom so it > points to it. In either case it finishes by dereferencing the STM > reference. > > When a pref is mutated, it first goes through the same motions as for > dereferencing. Then it simply forwards the mutation to the underlying > STM reference. > > Watchers are installed on STM references backed by prefs. Thus we are > notified when something is mutated. > > There is a pref-specific transaction boundary form called atomic, > analogous to dosync. The watchers are used to determine which prefs > were mutated during the transaction so as to flag them dirty for > write-back or write-through caching; this is why we need the pref back > reference in the metadata. > > Anyway, even assuming this all works, it will obviously be less > computationally efficient than extending LockingTransaction.java with > special support. > > -Per > > > > On Thu, Sep 23, 2010 at 10:21 AM, Alyssa Kwan <alyssa.c.k...@gmail.com> wrote: > > This is in reference > > tohttp://groups.google.com/group/clojure/browse_frm/thread/d6deac0c34d0.... > > > Hi everyone! > > > I have hacked the Clojure core to add durability to refs. The syntax > > to create these is (dref <val> <key> <path>), where <key> and <path> > > are strings. Then you use them just like refs. Creating a dref > > creates a global identity, such that subsequent dref calls to the same > > key and path will get the same dref. On subsequent dref calls, the > > <val> will be ignored and the persisted value used. This includes in > > subsequent VM instances. > > > Get it here: git://github.com/kwanalyssa/clojure.git > > > 1. <path> refers to BDB JE databases which get created in the "data" > > directory of your project. > > 2. BDB JE is used. I'm ignorant of IP and licensing issues. Making > > BDB JE core to Clojure is probably an issue. > > 3. Currently only a subset of Clojure primitives types is supported. > > No BigDecimal or Ratio yet. See the comprehensive list (and > > serialization mappings) at the bottom of src/jvm/clojure/lang/ > > DRef.java. BDB JE TupleBindings are used. Submissions welcome, > > especially for the persistent data structures. > > 4. How do we approach the problem of storing objects with lexical > > environments? > > 5. Unit tests welcome! I didn't do TDD since the work is in Java and > > there's no Java-level tests in the project. Please add your own in > > test/clojure/test_clojure/drefs.clj. I'm new to concurrency, so tests > > along those lines would be awesome! > > 6. The ACID part is not really guaranteed!!! STM is currently one- > > phase-commit. I inserted two-phase-commits to the data stores in the > > middle of the one-phase-commit. There's the remote possibility that > > STM in-memory changes fail AFTER writing to disk. It's REALLY remote, > > but it is possible. STM would have to be made 2PC to make this > > airtight. That's way beyond my current grasp of both concurrency and > > Clojure implementation. > > 7. I was aiming for an API where <path> is optional. However, I > > didn't want to stray from the ref API, which has variable arity. > > Suggestions on how to reconcile the two are welcome! > > 8. To maintain global identity, I use a static cache, which requires > > non-hard references to avoid OOM issues. This is my first time doing > > this, so please check my code to make sure that I'm doing it right. > > I'm using SoftReferences, though WeakReferences may be better for real- > > life usage patterns. Let me know! > > > Please dig in! Feedback appreciated! > > Alyssa Kwan > > > -- > > You received this message because you are subscribed to the Google > > Groups "Clojure" group. > > To post to this group, send email to clojure@googlegroups.com > > Note that posts from new members are moderated - please be patient with > > your first post. > > To unsubscribe from this group, send email to > > clojure+unsubscr...@googlegroups.com > > For more options, visit this group at > >http://groups.google.com/group/clojure?hl=en -- You received this message because you are subscribed to the Google Groups "Clojure" group. To post to this group, send email to clojure@googlegroups.com Note that posts from new members are moderated - please be patient with your first post. To unsubscribe from this group, send email to clojure+unsubscr...@googlegroups.com For more options, visit this group at http://groups.google.com/group/clojure?hl=en
