> redistributing shards from oversubscribed nodes to other nodes Redistributing shards on a live system is not possible however because the updates in-flight will likely be lost. Also it is not simple technology to build from the ground-up.
As is today, one would need to schedule downtime, for multi-terabyte live realtime systems, that not acceptable and will cause the system to not meet SLAs. Solr Cloud seems limited to a simple hashing algorithm for sending updates to the appropriate shard. This is precisely what Dynamo (and Cassandra) solves, eg, elastically and dynamically rearranging the hash 'ring' both logically and physically. In addition, there is the potential for data loss which Cassandra has the technology for. On Tue, Apr 17, 2012 at 1:33 PM, Otis Gospodnetic <otis_gospodne...@yahoo.com> wrote: > I think Jason is right - there is no index splitting in ES and SolrCloud, so > one has to think ahead, "overshard", and then count on redistributing shards > from oversubscribed nodes to other nodes. No resharding on demand and no > index/shard splitting yet. > > Otis > ---- > Performance Monitoring SaaS for Solr - > http://sematext.com/spm/solr-performance-monitoring/index.html > > > >>________________________________ >> From: Jason Rutherglen <jason.rutherg...@gmail.com> >>To: solr-user@lucene.apache.org >>Sent: Monday, April 16, 2012 8:42 PM >>Subject: Re: Options for automagically Scaling Solr (without needing >>distributed index/replication) in a Hadoop environment >> >>One of big weaknesses of Solr Cloud (and ES?) is the lack of the >>ability to redistribute shards across servers. Meaning, as a single >>shard grows too large, splitting the shard, while live updates. >> >>How do you plan on elastically adding more servers without this feature? >> >>Cassandra and HBase handle elasticity in their own ways. Cassandra >>has successfully implemented the Dynamo model and HBase uses the >>traditional BigTable 'split'. Both systems are complex though are at >>a singular level of maturity. >> >>Also Cassandra [successfully] implements multiple data center support, >>is that available in SC or ES? >> >>On Thu, Apr 12, 2012 at 7:23 PM, Otis Gospodnetic >><otis_gospodne...@yahoo.com> wrote: >>> Hello Ali, >>> >>>> I'm trying to setup a large scale *Crawl + Index + Search *infrastructure >>> >>>> using Nutch and Solr/Lucene. The targeted scale is *5 Billion web pages*, >>>> crawled + indexed every *4 weeks, *with a search latency of less than 0.5 >>>> seconds. >>> >>> >>> That's fine. Whether it's doable with any tech will depend on how much >>> hardware you give it, among other things. >>> >>>> Needless to mention, the search index needs to scale to 5Billion pages. It >>>> is also possible that I might need to store multiple indexes -- one for >>>> crawled content, and one for ancillary data that is also very large. Each >>>> of these indices would likely require a logically distributed and >>>> replicated index. >>> >>> >>> Yup, OK. >>> >>>> However, I would like for such a system to be homogenous with the Hadoop >>>> infrastructure that is already installed on the cluster (for the crawl). In >>>> other words, I would much prefer if the replication and distribution of the >>>> Solr/Lucene index be done automagically on top of Hadoop/HDFS, instead of >>>> using another scalability framework (such as SolrCloud). In addition, it >>>> would be ideal if this environment was flexible enough to be dynamically >>>> scaled based on the size requirements of the index and the search traffic >>>> at the time (i.e. if it is deployed on an Amazon cluster, it should be easy >>>> enough to automatically provision additional processing power into the >>>> cluster without requiring server re-starts). >>> >>> >>> There is no such thing just yet. >>> There is no Search+Hadoop/HDFS in a box just yet. There was an attempt to >>> automatically index HBase content, but that was either not completed or not >>> committed into HBase. >>> >>>> However, I'm not sure which Solr-based tool in the Hadoop ecosystem would >>>> be ideal for this scenario. I've heard mention of Solr-on-HBase, Solandra, >>>> Lily, ElasticSearch, IndexTank etc, but I'm really unsure which of these is >>>> mature enough and would be the right architectural choice to go along with >>>> a Nutch crawler setup, and to also satisfy the dynamic/auto-scaling aspects >>>> above. >>> >>> >>> Here is a summary on all of them: >>> * Search on HBase - I assume you are referring to the same thing I >>> mentioned above. Not ready. >>> * Solandra - uses Cassandra+Solr, plus DataStax now has a different >>> (commercial) offering that combines search and Cassandra. Looks good. >>> * Lily - data stored in HBase cluster gets indexed to a separate Solr >>> instance(s) on the side. Not really integrated the way you want it to be. >>> * ElasticSearch - solid at this point, the most dynamic solution today, can >>> scale well (we are working on a maaaany-B documents index and hundreds of >>> nodes with ElasticSearch right now), etc. But again, not integrated with >>> Hadoop the way you want it. >>> * IndexTank - has some technical weaknesses, not integrated with Hadoop, >>> not sure about its future considering LinkedIn uses Zoie and Sensei already. >>> * And there is SolrCloud, which is coming soon and will be solid, but is >>> again not integrated. >>> >>> If I were you and I had to pick today - I'd pick ElasticSearch if I were >>> completely open. If I had Solr bias I'd give SolrCloud a try first. >>> >>>> Lastly, how much hardware (assuming a medium sized EC2 instance) would you >>>> estimate my needing with this setup, for regular web-data (HTML text) at >>>> this scale? >>> >>> I don't know off the topic of my head, but I'm guessing several hundred for >>> serving search requests. >>> >>> HTH, >>> >>> Otis >>> -- >>> Search Analytics - http://sematext.com/search-analytics/index.html >>> >>> Scalable Performance Monitoring - http://sematext.com/spm/index.html >>> >>> >>>> Any architectural guidance would be greatly appreciated. The more details >>>> provided, the wider my grin :). >>>> >>>> Many many thanks in advance. >>>> >>>> Thanks, >>>> Safdar >>>> >> >> >>
