On Fri, Jul 23, 2010 at 2:23 PM, MitchK <mitc...@web.de> wrote: > why do we do not send the output of TermsComponent of every node in the > cluster to a Hadoop instance? > Since TermsComponent does the map-part of the map-reduce concept, Hadoop > only needs to reduce the stuff. Maybe we even do not need Hadoop for this. > After reducing, every node in the cluster gets the current values to compute > the idf. > We can store this information in a HashMap-based SolrCache (or something > like that) to provide constant-time access. To keep the values up to date, > we can repeat that after every x minutes.
There's already a patch in JIRA that does distributed IDF. Hadoop wouldn't be the right tool for that anyway... it's for batch oriented systems, not low-latency queries. > If we got that, it does not care whereas we use doc_X from shard_A or > shard_B, since they will all have got the same scores. That only works if the docs are exactly the same - they may not be. -Yonik http://www.lucidimagination.com