That page seems to be saying that the 'distributed' APIs take place on the leader, and the ordinary prepare/process APIs out at the leaves. I'll set out to prove or disprove that tomorrow.
On Mon, Apr 9, 2012 at 8:17 PM, Mark Miller <markrmil...@gmail.com> wrote: > > On Apr 9, 2012, at 7:34 PM, Benson Margulies wrote: > >> Those of you insomniacs who have read my messages here over the last >> few weeks might recall that I've been working on a request handler >> that wraps the SearchHandler to rewrite queries and then reorder >> results. >> >> (I haven't quite worked out how to apply Grant's alternative >> suggestions without losing the performance advantages I was looking >> for in the first place.) >> >> Today, I realized that the RequestHandler approach, as opposed to >> search components, wasn't going to be viable. I was growing too much >> dependency on internal Solr quirks. >> >> So I refactored it into a pair of SearchComponents -- one to go first >> and rewrite the query, and one to go after "query" and rescore. >> >> And it works just fine ----- until I configure it into a SolrCloud >> cluster. At which point it started coming up with very wrong answers. >> >> I think that the reason is that I don't have an implementation of the >> distributedProcess method, or, more generally, that I don't understand >> the protocol on a SearchComponent when distributed processing is >> happening. Has anyone written anything yet about these considerations? >> I can put multiple processes in the debugging and see who gets called >> with what, but I was hoping for some sort of short cut. > > > > Grant started something on this once: > http://wiki.apache.org/solr/WritingDistributedSearchComponents > It's only a start though. > > Unfortunately, to this point, adventurous souls have had to debug and study > there way to understanding the distrib process solo mostly. > > Perhaps we can encourage anyone that has written a distributed component to > help jump in on that wiki page. Any takers? > > - Mark Miller > lucidimagination.com > > > > > > > > > > >
