- From: Ray Whitmer <ray@imall.com>
- Date: Tue, 20 Oct 1998 13:26:24 -0600
- To: Miles Sabin <msabin@cromwellmedia.co.uk>
- CC: "'DOM list'" <www-dom@w3.org>
Miles Sabin wrote: > Ray Whitmer wrote, > > I must have missed something here. Even with unpredictable > > Java GC, we are still much better off with weak references > > than without. I would interpret this as greatly helping > > address the problem. > > I guess a lot turns on what you count as 'much better off'. > As far as I can see all you gain is not having to implement > the handle-body pattern ... but I think for anyone able to > implement the DOM reasonably well that ought to be a trivial > exercise. For this case, handle-body is equal to weak references. With either one of the two, you are much better off than without either, which was what I intended to communicate. > As it happens, handle-body+finalize is the guts of the > underlying implementation of weak refs in 1.2 as a peek > at the JDK sources will reveal. When I looked at it in the past, it did not work that way. I believe it solves problems that cannot be solved merely by handle-body. I will look at it again, and get back to you privately. > OK, that's speculation ... I'd be interested to see some > stats and analysis from someone (Ray?) who's put that sort > of implementation together. My DOM runs on the server, but I don't have good stats available yet for reclaiming NodeLists for a variety of reasons: 1. Our implementation directly implements children as lists, so this is not an issue when processing child lists. At most, we only ever need one NodeList per node, which never get fixed up artificially. So they never pile up in the first place. 2. getElementsByTagName is not something we call much on the server for long-lived documents, because for documents which stay around for a long time, we build advanced indexes for them so we can find things in them quickly. 3. Our stress tests have been quite artificial. We will be going live within the next month or so. 4. If we ever did forsee a use case that might stress test it, I would quickly build indexes. If you create the tests, I will be happy to run it and give you back any stats you would like. Ray Whitmer ray@imall.com
Received on Tuesday, 20 October 1998 15:27:59 UTC