Right now lets say you have one shard - everything there hashes to range X.
Now you want to split that shard with an Index Splitter. You divide range X in two - giving you two ranges - then you start splitting. This is where the current Splitter needs a little modification. You decide which doc should go into which new index by rehashing each doc id in the index you are splitting - if its hash is greater than X/2, it goes into index1 - if its less, index2. I think there are a couple current Splitter impls, but one of them does something like, give me an id - now if the id's in the index are above that id, goto index1, if below, index2. We need to instead do a quick hash rather than simple id compare. Why do you need to do this on every shard? The other part we need that we dont have is to store hash range assignments in zookeeper - we don't do that yet because it's not needed yet. Instead we currently just simply calculate that on the fly (too often at the moment - on every request :) I intend to fix that of course). At the start, zk would say, for range X, goto this shard. After the split, it would say, for range less than X/2 goto the old node, for range greater than X/2 goto the new node. - Mark On Dec 1, 2011, at 7:44 PM, Jamie Johnson wrote: > hmmm.....This doesn't sound like the hashing algorithm that's on the > branch, right? The algorithm you're mentioning sounds like there is > some logic which is able to tell that a particular range should be > distributed between 2 shards instead of 1. So seems like a trade off > between repartitioning the entire index (on every shard) and having a > custom hashing algorithm which is able to handle the situation where 2 > or more shards map to a particular range. > > On Thu, Dec 1, 2011 at 7:34 PM, Mark Miller <markrmil...@gmail.com> wrote: >> >> On Dec 1, 2011, at 7:20 PM, Jamie Johnson wrote: >> >>> I am not familiar with the index splitter that is in contrib, but I'll >>> take a look at it soon. So the process sounds like it would be to run >>> this on all of the current shards indexes based on the hash algorithm. >> >> Not something I've thought deeply about myself yet, but I think the idea >> would be to split as many as you felt you needed to. >> >> If you wanted to keep the full balance always, this would mean splitting >> every shard at once, yes. But this depends on how many boxes (partitions) >> you are willing/able to add at a time. >> >> You might just split one index to start - now it's hash range would be >> handled by two shards instead of one (if you have 3 replicas per shard, this >> would mean adding 3 more boxes). When you needed to expand again, you would >> split another index that was still handling its full starting range. As you >> grow, once you split every original index, you'd start again, splitting one >> of the now half ranges. >> >>> Is there also an index merger in contrib which could be used to merge >>> indexes? I'm assuming this would be the process? >> >> You can merge with IndexWriter.addIndexes (Solr also has an admin command >> that can do this). But I'm not sure where this fits in? >> >> - Mark >> >>> >>> On Thu, Dec 1, 2011 at 7:18 PM, Mark Miller <markrmil...@gmail.com> wrote: >>>> Not yet - we don't plan on working on this until a lot of other stuff is >>>> working solid at this point. But someone else could jump in! >>>> >>>> There are a couple ways to go about it that I know of: >>>> >>>> A more long term solution may be to start using micro shards - each index >>>> starts as multiple indexes. This makes it pretty fast to move mirco shards >>>> around as you decide to change partitions. It's also less flexible as you >>>> are limited by the number of micro shards you start with. >>>> >>>> A more simple and likely first step is to use an index splitter . We >>>> already have one in lucene contrib - we would just need to modify it so >>>> that it splits based on the hash of the document id. This is super >>>> flexible, but splitting will obviously take a little while on a huge index. >>>> The current index splitter is a multi pass splitter - good enough to start >>>> with, but most files under codec control these days, we may be able to make >>>> a single pass splitter soon as well. >>>> >>>> Eventually you could imagine using both options - micro shards that could >>>> also be split as needed. Though I still wonder if micro shards will be >>>> worth the extra complications myself... >>>> >>>> Right now though, the idea is that you should pick a good number of >>>> partitions to start given your expected data ;) Adding more replicas is >>>> trivial though. >>>> >>>> - Mark >>>> >>>> On Thu, Dec 1, 2011 at 6:35 PM, Jamie Johnson <jej2...@gmail.com> wrote: >>>> >>>>> Another question, is there any support for repartitioning of the index >>>>> if a new shard is added? What is the recommended approach for >>>>> handling this? It seemed that the hashing algorithm (and probably >>>>> any) would require the index to be repartitioned should a new shard be >>>>> added. >>>>> >>>>> On Thu, Dec 1, 2011 at 6:32 PM, Jamie Johnson <jej2...@gmail.com> wrote: >>>>>> Thanks I will try this first thing in the morning. >>>>>> >>>>>> On Thu, Dec 1, 2011 at 3:39 PM, Mark Miller <markrmil...@gmail.com> >>>>> wrote: >>>>>>> On Thu, Dec 1, 2011 at 10:08 AM, Jamie Johnson <jej2...@gmail.com> >>>>> wrote: >>>>>>> >>>>>>>> I am currently looking at the latest solrcloud branch and was >>>>>>>> wondering if there was any documentation on configuring the >>>>>>>> DistributedUpdateProcessor? What specifically in solrconfig.xml needs >>>>>>>> to be added/modified to make distributed indexing work? >>>>>>>> >>>>>>> >>>>>>> >>>>>>> Hi Jaime - take a look at solrconfig-distrib-update.xml in >>>>>>> solr/core/src/test-files >>>>>>> >>>>>>> You need to enable the update log, add an empty replication handler def, >>>>>>> and an update chain with solr.DistributedUpdateProcessFactory in it. >>>>>>> >>>>>>> -- >>>>>>> - Mark >>>>>>> >>>>>>> http://www.lucidimagination.com >>>>>>> >>>>>> >>>>> >>>> >>>> >>>> >>>> -- >>>> - Mark >>>> >>>> http://www.lucidimagination.com >>>> >> >> - Mark Miller >> lucidimagination.com >> >> >> >> >> >> >> >> >> >> >> >> - Mark Miller lucidimagination.com
