Thanks for your commments Ulrich. I just posted a message asking if anyone had attempted this approach! Sounds like you have, and it works :-) Thanks for information, this sounds pretty close to what my preferred approach would be.
You say you get 2000 docs/minute. I've done some benchmarking and managed to get our data indexing at ~1000/minute on an Athlon 1800+ (and most of that speed was acheived by bumping the IndexWriter.mergeFactor up to 100 or so). Our data is coming from a database table, each record contains about 40 fields, and I'm indexing 8 of those fields (an ID, 4 number fields, 3 text fields including one that has ~2k text). Does this sound reasonable to you, or do you have any tips that might improve that performance? "Ulrich Mayring" <[EMAIL PROTECTED]> wrote in message news:[EMAIL PROTECTED] > Chris Miller wrote: > > > > The main thing I'm interested in is how you handle updates to Lucene's > > index. I'd imagine you have a fairly high turnover of CVs and jobs, so index > > updates must place a reasonable load on the CPU/disk. Do you keep CVs and > > jobs in the same index or two different ones? And what is the process you > > use to update the index(es) - do you batch-process updates or do you handle > > them in real-time as changes are made? > > The way we do it: we re-index everything periodically in a temporary > directory and then rename the temporary directory. That way the index > remains accessible at all times and its currency is simply determined by > the interval I run the re-indexing in. > > > We need to be able to handle indexing about 60,000 documents/day, > > while allowing (many) searches to continue operating alongside. > > On an entry-level Sun I can index about 23 documents per second and > these are real-life HTML pages. Thus in less than one hour you would be > finished with a complete index run and save yourself all kinds of > trouble with crashes during indexing etc. > > On my 2 GHz Linux workstation it's even faster: more than 2000 documents > per minute, so you'd be done in half an hour. > > BTW, we're not using the supplied JavaCC-based HTML parser, instead we > got htmlparser.sourceforge.net, which is a joy to use and pretty fast. > > Ulrich --------------------------------------------------------------------- To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]