On Wed, 2013-10-30 at 14:24 +0100, Shawn Heisey wrote: > On 10/30/2013 4:00 AM, Toke Eskildsen wrote: > > Why would TRIM have any influence on whether or not a driver failure > > also means server failure? > > I left out a step in my description. > > Lack of TRIM support in RAID means that I would avoid RAID with SSD. No > RAID means that when the SSD fails, that Solr is out of commission until > its SSD can be replaced.
That makes sense, thanks. > I don't know how Solr would behave if you put each core on its own SSD > and one of them fails. Hopefully it's smart enough to keep going with > the cores that have working filesystems. I don't know either. Seems like it would be a useful thing to test. We did some comparison on 9 shards of 420GB (against a SAN), where we tested SolrCloud with 9 independent Solr instances vs. a single instance with multiple cores. The overhead of independent instances did not seem severe for that shard size and should be resilient against single drive failure. As we're looking at a cumulative heap requirement of 100GB+ due to grouping and faceting, it might be preferable to run with independent Solrs anyway to minimize garbage collection pauses. I do not know if that logic extends in general to large Solr installations. Regards, Toke Eskildsen, State and University Library, Denmark
