Hi Jim, On 08/08/2010, Jim Fulton <jim@zope.com> wrote:
On Thu, Aug 5, 2010 at 2:36 AM, Martin Aspeli <optilude+lists@gmail.com> wrote: ...
I have a package (plone.registry) that currently has a persistent structure like this:
Registry(Persistent) | +--> Records(Persistent) | +--> BTree of Record(Persistent) | +--> PersistentField(Persistent)
That is, a Registry is a persistent object containing a persistent Records object that in turn contains a BTree of persistent Record
Since BTrees are mapping, I assume that you mean the values are records and that the keys are something boring like strings or integers.
Yes. The keys are strings.
I like to use mathematical notation when talking about BTrees and sets, as in:
Registry BTree {? -> Record}
objects that contain a persistent PersistentField and a primitive value.
This is quite inefficient, though, because it results in a lot of object loads. On any given request, some of our projects load a dozen or more values from the registry. Each is just a simple primitive, but we need to load the full shebang to make it work.
Not sure what you mean by "full shebang".
The Registry, Records object, the relevant Record in the relevant BTree, and possibly the PersistentField object. In the "new" branch it just looks up the value in the relevant BTree.
Now, I'd like to move to this structure:
Registry(Persistent) | +--> Records | +--> BTree of Field | +--> BTree of values
I'm foggy on what "field" and "value" are here or what your queries are doing. Maybe this is just a distraction.
Somewhat, unless you've worked with plone.registry. The point is to allow the "get a value" API to just look at self.values[key], which is a fast lookup and doesn't load anything except the relevant BTree bucket + the registry itself.
Here, there's only one Persistent object, plus the two BTrees: one holding all the fields and one holding all the values. Records no longer needs to be persistent (its attributes become part of the parent Registry's _p_jar).
I wonder what role "Records" plays independent of the "Registry".
None, really. The main reason to have it is to be able to have an API like registry.records with dict-like notation (there's also __getitem__ on the registry, which returns the value of a given key, not the Record). I made ``records`` an attribute of type Records, and Records derives from Persistent. I wish I hadn't, since it can just live in its parent's _p_jar.
I also wonder why it matters whether it is persistent or not.
It's better if it isn't (one fewer object to load/fill up the cache), though the real culprits are the many Record objects each being persistent and loaded separately. On a given request, we can end up loading a dozen or more values from the registry, which means a dozen or more objects in the cache and associated overhead.
Fields no longer need to be persistent either, since they are in effect immutable objects. Values are primitives anyway.
I've done this (in a branch) and it works for new sites. However, I'm having a nightmare trying to migrate old sites. As soon as I access anything that uses the registry, I get ZODB errors, because the persistent structure is now different. In particular, it's trying to read a value into e.g. a Records object that used to derive from Persistent, but now no longer does.
What savings do you get by making Records non-persistent?
One fewer persistent object. I think the real saving is in making the Record object non-persistent, especially since the "read" use case can just read from the ``values`` BTree with the structure above.
What is the best way to manage this type of migration?
Today, it probably makes the most sense to make new classes for the non-persistemnt objects. You'll then need to write a script to rebuild the data structures.
Okay. So there's no way to get at the data if I take Persistent out of the base classes for Records / Record. Martin