The speed of ingest via HTTP improves greatly once you do two things: 1. Batch multiple documents into a single request. 2. Index with multiple threads at once.
Michael Della Bitta Applications Developer o: +1 646 532 3062 appinions inc. "The Science of Influence Marketing" 18 East 41st Street New York, NY 10017 t: @appinions <https://twitter.com/Appinions> | g+: plus.google.com/appinions<https://plus.google.com/u/0/b/112002776285509593336/112002776285509593336/posts> w: appinions.com <http://www.appinions.com/> On Mon, Apr 7, 2014 at 12:40 PM, Daniel Collins <danwcoll...@gmail.com>wrote: > I have to agree with Shawn. We have a SolrCloud setup with 256 shards, > ~400M documents in total, with 4-way replication (so its quite a big > setup!) I had thought that HTTP would slow things down, so we recently > trialed a JNI approach (clients are C++) so we could call SolrJ and get the > benefits of JavaBin encoding for our indexing.... > > Once we had done benchmarks with both solutions, I think we saved about 1ms > per document (on average) with JNI, so it wasn't as big a gain as we were > expecting. There are other benefits of SolrJ (zookeeper integration, > better routing, etc) and we were doing local HTTP (so it was literally just > a TCP port to localhost, no actual net traffic) but that just goes to prove > what other posters have said here. Check whether HTTP really *is* the > bottleneck before you try to replace it! > > > On 7 April 2014 17:05, Shawn Heisey <s...@elyograg.org> wrote: > > > On 4/7/2014 5:52 AM, Jonathan Varsanik wrote: > > > >> Do you mean to tell me that the people on this list that are indexing > >> 100s of millions of documents are doing this over http? I have been > using > >> custom Lucene code to index files, as I thought this would be faster for > >> many documents and I wanted some non-standard OCR and index fields. Is > >> there a better way? > >> > >> To the OP: You can also use Lucene to locally index files for Solr. > >> > > > > My sharded index has 94 million docs in it. All normal indexing and > > maintenance is done with SolrJ, over http.Currently full rebuilds are > done > > with the dataimport handler loading from MySQL, but that is legacy. This > > is NOT a SolrCloud installation. It is also not a replicated setup -- my > > indexing program keeps both copies up to date independently, similar to > > what happens behind the scenes with SolrCloud. > > > > The single-thread DIH is very well optimized, and is faster than what I > > have written myself -- also single-threaded. > > > > The real reason that we still use DIH for rebuilds is that I can run the > > DIH simultaenously on all shards. A full rebuild that way takes about 5 > > hours. A SolrJ process feeding all shards with a single thread would > take > > a lot longer. Once I have time to work on it, I can make the SolrJ > rebuild > > multi-threaded, and I expect it will be similar to DIH in rebuild speed. > > Hopefully I can make it faster. > > > > There is always overhead with HTTP. On a gigabit LAN, I don't think it's > > high enough to matter. > > > > Using Lucene to index files for Solr is an option -- but that requires > > writing a custom Lucene application, and knowledge about how to turn the > > Solr schema into Lucene code. A lot of users on this list (me included) > do > > not have the skills required. I know SolrJ reasonably well, but Lucene > is > > a nut that I haven't cracked. > > > > Thanks, > > Shawn > > > > >