Re: Solr cloud performance degradation with billions of documents
bq: I am interested in knowing, when you have multiple collections like this case (60), and you just query one collection, Yes.. and no. You're correct about OS memory swapping in and out, but the JVM memory is a different matter. There'll be some low-level caches filled up. Each collection may have filterCache entries. Or sort entries. Or... There are lots of Java memory-resident structures that are _not_ swapped out. Furthermore, each collection may have 1-n warming queries fired when it's loaded. And the top-level caches configured in solrconfig.xml may have autowarm counts. And The long and short of it is that each collection will consume memory when it's loaded in varying amounts. Memory that MUST live in the JVM. Other memory will be paged in and out, see Uwe's excellent blog here: http://blog.thetaphi.de/2012/07/use-lucenes-mmapdirectory-on-64bit.html So as you add more and more collections, you'll hit this kind of problem. Now, there is the LotsOfCores code, see: http://wiki.apache.org/solr/LotsOfCores WARNING: This is NOT supported (yet) for SolrCloud. That code does load/unload cores as necessary based on the configuration parameters that determine 1 whether a core can be unloaded/loaded 2 how many transient cores can be in memory at once Best, Erick On Sat, Aug 16, 2014 at 6:43 PM, shushuai zhu ss...@yahoo.com.invalid wrote: Erik, --- I fear the problem will be this: you won't even be able to do basic searches as the number of shards on a particular machine increase. To test, fire off a simple search for each of your 60 days. I expect it'll blow you out of the water. This assumes that all your shards are hosted in the same JVM on each of your 32 machines. But that's totally a guess. --- In this case, assuming there are 60 collections, and only one collection is queried each time, should the memory requirements be those for that collection only? My understanding is, when a new collection is queried, the indexes (cores) of the old collection in OS cache are to be swapped out and the indexes of new collection are brought in, but the memory requirements should be roughly the same as long as two collections have similar sizes. I am interested in knowing, when you have multiple collections like this case (60), and you just query one collection, should other collections matter from performance perspective? Since different collections contain different cores, if querying one collection involves cores in other collections, is it a bug? Thanks. Shushuai From: Erick Erickson erickerick...@gmail.com To: solr-user@lucene.apache.org Sent: Friday, August 15, 2014 7:30 PM Subject: Re: Solr cloud performance degradation with billions of documents Toke: bq: I would have agreed with you fully an hour ago. Well, I now disagree with myself too :) I don't mind talking to myself. I don't even mind arguing with myself. I really _do_ mind losing the arguments I have with myself though. Scott: OK, that has a much better chance of working, I obviously misunderstood. So you'll have 60 different collections and each collection will have one shard on each machine. When the time comes to roll some of the collections off the end due to age, collection aliasing may be helpful. I still think you're significantly undersized, but you know your problem space better than I do. I fear the problem will be this: you won't even be able to do basic searches as the number of shards on a particular machine increase. To test, fire off a simple search for each of your 60 days. I expect it'll blow you out of the water. This assumes that all your shards are hosted in the same JVM on each of your 32 machines. But that's totally a guess. Keep us posted! On Fri, Aug 15, 2014 at 2:40 PM, Toke Eskildsen t...@statsbiblioteket.dk wrote: Erick Erickson [erickerick...@gmail.com] wrote: I guess that my main issue is that from everything I've seen so far, this project is doomed. You simply cannot put 7B documents in a single shard, period. Lucene has a 2B hard limit. I would have agreed with you fully an hour ago and actually planned to ask Wilbur to check if he had corrupted his indexes. However, his latest post suggests that the scenario is more about having a larger amount of more resonably sized shards in play than building gigantic shards. For instance, Wilburn is talking about only using 6G of memory. Even at 2B docs/shard, I'd be surprised to see it function at all. Don't try sorting on a timestamp for instance. I haven't understood Wilburns setup completely, as it seems to me that he will quickly run out of memory for starting new shards. But if we are looking at shards of 30GB and 160M documents, 6GB sounds a lot better. Regards, Toke Eskildsen
Re: Solr cloud performance degradation with billions of documents
Erik, --- I fear the problem will be this: you won't even be able to do basic searches as the number of shards on a particular machine increase. To test, fire off a simple search for each of your 60 days. I expect it'll blow you out of the water. This assumes that all your shards are hosted in the same JVM on each of your 32 machines. But that's totally a guess. --- In this case, assuming there are 60 collections, and only one collection is queried each time, should the memory requirements be those for that collection only? My understanding is, when a new collection is queried, the indexes (cores) of the old collection in OS cache are to be swapped out and the indexes of new collection are brought in, but the memory requirements should be roughly the same as long as two collections have similar sizes. I am interested in knowing, when you have multiple collections like this case (60), and you just query one collection, should other collections matter from performance perspective? Since different collections contain different cores, if querying one collection involves cores in other collections, is it a bug? Thanks. Shushuai From: Erick Erickson erickerick...@gmail.com To: solr-user@lucene.apache.org Sent: Friday, August 15, 2014 7:30 PM Subject: Re: Solr cloud performance degradation with billions of documents Toke: bq: I would have agreed with you fully an hour ago. Well, I now disagree with myself too :) I don't mind talking to myself. I don't even mind arguing with myself. I really _do_ mind losing the arguments I have with myself though. Scott: OK, that has a much better chance of working, I obviously misunderstood. So you'll have 60 different collections and each collection will have one shard on each machine. When the time comes to roll some of the collections off the end due to age, collection aliasing may be helpful. I still think you're significantly undersized, but you know your problem space better than I do. I fear the problem will be this: you won't even be able to do basic searches as the number of shards on a particular machine increase. To test, fire off a simple search for each of your 60 days. I expect it'll blow you out of the water. This assumes that all your shards are hosted in the same JVM on each of your 32 machines. But that's totally a guess. Keep us posted! On Fri, Aug 15, 2014 at 2:40 PM, Toke Eskildsen t...@statsbiblioteket.dk wrote: Erick Erickson [erickerick...@gmail.com] wrote: I guess that my main issue is that from everything I've seen so far, this project is doomed. You simply cannot put 7B documents in a single shard, period. Lucene has a 2B hard limit. I would have agreed with you fully an hour ago and actually planned to ask Wilbur to check if he had corrupted his indexes. However, his latest post suggests that the scenario is more about having a larger amount of more resonably sized shards in play than building gigantic shards. For instance, Wilburn is talking about only using 6G of memory. Even at 2B docs/shard, I'd be surprised to see it function at all. Don't try sorting on a timestamp for instance. I haven't understood Wilburns setup completely, as it seems to me that he will quickly run out of memory for starting new shards. But if we are looking at shards of 30GB and 160M documents, 6GB sounds a lot better. Regards, Toke Eskildsen
Re: Solr cloud performance degradation with billions of documents
Toke: You make valid points. You're completely right that my reflexes are for sub-second responses so I tend to think of lots and lots of memory being a requirement. I agree that depending on the problem space the percentage of the index that has to be in memory varies widely, I've seen a large variance in projects. And I know you've done some _very_ interesting things tuning-wise! I guess that my main issue is that from everything I've seen so far, this project is doomed. You simply cannot put 7B documents in a single shard, period. Lucene has a 2B hard limit. Wilburn is making assumptions here that are simply wrong. Or my math is off, that's been known to happen too. For instance, Wilburn is talking about only using 6G of memory. Even at 2B docs/shard, I'd be surprised to see it function at all. Don't try sorting on a timestamp for instance. I've never seen 2B docs fit on a shard and be OK performance-wise. Or, for that matter, perform at all. If there are situations like that I'd _love_ to know the details... For any chance of success, Wilburn has to go back and do some reassessment IMO. There's no magic knob to turn to overcome the fundamental limitations that are going to creeping out of the woodwork. Indexing throughput is the least of his problems. I further suspect (but don't know for sure) that the first time realistic queries start hitting the system it'll OOM. All that said, I don't know for sure of course. On Thu, Aug 14, 2014 at 11:57 AM, Toke Eskildsen t...@statsbiblioteket.dk wrote: Erick Erickson [erickerick...@gmail.com] wrote: Solr requires holding large parts of the index in memory. For the entire corpus. At once. That requirement is under the assumption that one must have the lowest possible latency at each individual box. You might as well argue for the fastest possible memory or the fastest possible CPU being a requirement. The advice is good in some contexts and a waste of money in other. I not-so-humbly point to http://sbdevel.wordpress.com/2014/08/13/whale-hunting-with-solr/ where we (for simple searches) handily achieve our goal of sub-second response times for a 10TB index with just 1.4% of the index cached in RAM. Had our goal been sub-50ms, it would be another matter, but it is not. Just as Wilburn's problem is not to minimize latency for each individual box, but to achieve a certain throughput for indexing, while performing searches. Wilburn's hardware is currently able to keep up, although barely, with 300B documents. He needs to handle 900B. Tripling (or quadrupling) the amount of machines should do the trick. Increasing the amount of RAM on each current machine might also work (qua the well known effect of RAM with Lucene/Solr). Using local SSDs, if he is not doing so already, might also work (qua the article above). - Toke Eskildsen
RE: Solr cloud performance degradation with billions of documents
Erick, You make some very good valid points. Let me clear a few things up, though. We are not trying to put 7B docs into one single shard, because we are using collections, created daily, which spread the index across the 32 shards that make up the cloud/collection. Last I counted, we are putting about 160M docs per collection per shard. You are very correct about the memory issues. In fact, we cannot do any complicated searches or faceting without Solr returning memory errors. Basic searching still works fine, fortunately. This limit on search is acceptable in our case, though not ideal, to ensure the project succeeds and comes in under budget. Thanks, Scott -Original Message- From: Erick Erickson [mailto:erickerick...@gmail.com] Sent: Friday, August 15, 2014 7:52 AM To: solr-user@lucene.apache.org Subject: Re: Solr cloud performance degradation with billions of documents Toke: You make valid points. You're completely right that my reflexes are for sub-second responses so I tend to think of lots and lots of memory being a requirement. I agree that depending on the problem space the percentage of the index that has to be in memory varies widely, I've seen a large variance in projects. And I know you've done some _very_ interesting things tuning-wise! I guess that my main issue is that from everything I've seen so far, this project is doomed. You simply cannot put 7B documents in a single shard, period. Lucene has a 2B hard limit. Wilburn is making assumptions here that are simply wrong. Or my math is off, that's been known to happen too. For instance, Wilburn is talking about only using 6G of memory. Even at 2B docs/shard, I'd be surprised to see it function at all. Don't try sorting on a timestamp for instance. I've never seen 2B docs fit on a shard and be OK performance-wise. Or, for that matter, perform at all. If there are situations like that I'd _love_ to know the details... For any chance of success, Wilburn has to go back and do some reassessment IMO. There's no magic knob to turn to overcome the fundamental limitations that are going to creeping out of the woodwork. Indexing throughput is the least of his problems. I further suspect (but don't know for sure) that the first time realistic queries start hitting the system it'll OOM. All that said, I don't know for sure of course. On Thu, Aug 14, 2014 at 11:57 AM, Toke Eskildsen t...@statsbiblioteket.dk wrote: Erick Erickson [erickerick...@gmail.com] wrote: Solr requires holding large parts of the index in memory. For the entire corpus. At once. That requirement is under the assumption that one must have the lowest possible latency at each individual box. You might as well argue for the fastest possible memory or the fastest possible CPU being a requirement. The advice is good in some contexts and a waste of money in other. I not-so-humbly point to http://sbdevel.wordpress.com/2014/08/13/whale-hunting-with-solr/ where we (for simple searches) handily achieve our goal of sub-second response times for a 10TB index with just 1.4% of the index cached in RAM. Had our goal been sub-50ms, it would be another matter, but it is not. Just as Wilburn's problem is not to minimize latency for each individual box, but to achieve a certain throughput for indexing, while performing searches. Wilburn's hardware is currently able to keep up, although barely, with 300B documents. He needs to handle 900B. Tripling (or quadrupling) the amount of machines should do the trick. Increasing the amount of RAM on each current machine might also work (qua the well known effect of RAM with Lucene/Solr). Using local SSDs, if he is not doing so already, might also work (qua the article above). - Toke Eskildsen
RE: Solr cloud performance degradation with billions of documents
Wilburn, Scott [scott.wilb...@verizonwireless.com.INVALID] wrote: You make some very good valid points. Let me clear a few things up, though. We are not trying to put 7B docs into one single shard, because we are using collections, created daily, which spread the index across the 32 shards that make up the cloud/collection. Just to be sure I understand: You make a new collection, consisting of 32 shards, each day? And when you do, the old collection is not updated anymore? As your primary problem is indexing speed degradation, dividing your machines into a dedicated search pool and a dedicated index (plus search in the collection being build) pool might work. This would require you to move finished collections from the indexers to the searchers, but it would make it possible for you to have quite fine-grained control over how much power should be given to each of the two jobs, by adjusting the pool sizes. Furthermore having shards that are no longer updated allows for optimization down to a single segment, which might also help with performance. You are very correct about the memory issues. In fact, we cannot do any complicated searches or faceting without Solr returning memory errors. Could you describe the field(s) you would like to facet on? Number/string? Single-/multi-value? Have you tried with DocValues? Under the right circumstances, faceting can be done surprisingly cheap. - Toke Eskildsen
RE: Solr cloud performance degradation with billions of documents
Erick Erickson [erickerick...@gmail.com] wrote: I guess that my main issue is that from everything I've seen so far, this project is doomed. You simply cannot put 7B documents in a single shard, period. Lucene has a 2B hard limit. I would have agreed with you fully an hour ago and actually planned to ask Wilbur to check if he had corrupted his indexes. However, his latest post suggests that the scenario is more about having a larger amount of more resonably sized shards in play than building gigantic shards. For instance, Wilburn is talking about only using 6G of memory. Even at 2B docs/shard, I'd be surprised to see it function at all. Don't try sorting on a timestamp for instance. I haven't understood Wilburns setup completely, as it seems to me that he will quickly run out of memory for starting new shards. But if we are looking at shards of 30GB and 160M documents, 6GB sounds a lot better. Regards, Toke Eskildsen
Re: Solr cloud performance degradation with billions of documents
Toke: bq: I would have agreed with you fully an hour ago. Well, I now disagree with myself too :) I don't mind talking to myself. I don't even mind arguing with myself. I really _do_ mind losing the arguments I have with myself though. Scott: OK, that has a much better chance of working, I obviously misunderstood. So you'll have 60 different collections and each collection will have one shard on each machine. When the time comes to roll some of the collections off the end due to age, collection aliasing may be helpful. I still think you're significantly undersized, but you know your problem space better than I do. I fear the problem will be this: you won't even be able to do basic searches as the number of shards on a particular machine increase. To test, fire off a simple search for each of your 60 days. I expect it'll blow you out of the water. This assumes that all your shards are hosted in the same JVM on each of your 32 machines. But that's totally a guess. Keep us posted! On Fri, Aug 15, 2014 at 2:40 PM, Toke Eskildsen t...@statsbiblioteket.dk wrote: Erick Erickson [erickerick...@gmail.com] wrote: I guess that my main issue is that from everything I've seen so far, this project is doomed. You simply cannot put 7B documents in a single shard, period. Lucene has a 2B hard limit. I would have agreed with you fully an hour ago and actually planned to ask Wilbur to check if he had corrupted his indexes. However, his latest post suggests that the scenario is more about having a larger amount of more resonably sized shards in play than building gigantic shards. For instance, Wilburn is talking about only using 6G of memory. Even at 2B docs/shard, I'd be surprised to see it function at all. Don't try sorting on a timestamp for instance. I haven't understood Wilburns setup completely, as it seems to me that he will quickly run out of memory for starting new shards. But if we are looking at shards of 30GB and 160M documents, 6GB sounds a lot better. Regards, Toke Eskildsen
RE: Solr cloud performance degradation with billions of documents
Erick, Thanks for your suggestion to look into MapReduceIndexerTool, I'm looking into that now. I agree what I am trying to do is a tall order, and the more I hear from all of your comments, the more I am convinced that lack of memory is my biggest problem. I'm going to work on increasing the memory now, but was wondering if there are any configuration or other techniques that could also increase ingest performance? Does anyone know if a cloud of this size( hundreds of billions ) with an ingest rate of 5 billion new each day, has ever been attempted before? Thanks, Scott -Original Message- From: Erick Erickson [mailto:erickerick...@gmail.com] Sent: Wednesday, August 13, 2014 4:48 PM To: solr-user@lucene.apache.org Subject: Re: Solr cloud performance degradation with billions of documents Several points: 1 Have you considered using the MapReduceIndexerTool for your ingestion? Assuming you don't have duplicate IDs, i.e. each doc is new, you can spread your indexing across as many nodes as you have in your cluster. That said, it's not entirely clear that you'll gain throughput since you have as many nodes as you do. 2 Um, fitting this many documents into 6G of memory is ambitious. 2 Very ambitious. Actually it's impossible. By my calculations: bq: 4 separate and individual clouds of 32 shards each so 128 shards in aggregate bq: inserting into these clouds per day is 5 Billion each in two clouds, 3 Billion into the third, and 2 Billion into the fourth so we're talking 15B docs/day bq: the plan is to keep up to 60 days... So were talking 900B documents. It just won't work. 900B/128 docs/shard is over 7B documents/shard on average. Your two larger collections will have more than that, the two smaller ones less. But it doesn't matter because: 1: Lucene has a limit of 2B docs per core(shard), positive signed int. 2: It ain't gonna fit in 6G of memory even without this limit I'm pretty sure. 3: I've rarely heard of a single shard coping with over 300M docs without performance issues. I usually start getting nervous around 100M and insist on stress testing. Of course it depends lots on your query profile. So you're going to need a LOT more shards. You might be able to squeeze some more from your hardware by hosting multiple shards on for each collection on each machine, but I'm pretty sure your present setup is inadequate for your projected load. Of course I may be misinterpreting what you're saying hugely, but from what I understand this system just won't work. Best, Erick On Wed, Aug 13, 2014 at 2:39 PM, Markus Jelsma markus.jel...@openindex.io wrote: Hi - You are running mapred jobs on the same nodes as Solr runs right? The first thing i would think of is that your OS file buffer cache is abused. The mappers read all data, presumably residing on the same node. The mapper output and shuffling part would take place on the same node, only the reducer output is sent to your nodes, which i assume are on the same machines. Those same machines have a large Lucene index. All this data, written to and read from the same disk, competes for a nice spot in the OS buffer cache. Forget it if i misread anything, but when you're using serious figures of size, then do not abuse your caches. Have a separate mapred and Solr cluster, because they both eat cache space. I assume you can see serious IO WAIT times. Split the stuff and maybe even use smaller hardware, but more. M -Original message- From:Wilburn, Scott scott.wilb...@verizonwireless.com.INVALID Sent: Wednesday 13th August 2014 23:09 To: solr-user@lucene.apache.org Subject: Solr cloud performance degradation with billions of documents Hello everyone, I am trying to use SolrCloud to index a very large number of simple documents and have run into some performance and scalability limitations and was wondering what can be done about it. Hardware wise, I have a 32-node Hadoop cluster that I use to run all of the Solr shards and each node has 128GB of memory. The current SolrCloud setup is split into 4 separate and individual clouds of 32 shards each thereby giving four running shards per cloud or one cloud per eight nodes. Each shard is currently assigned a 6GB heap size. I’d prefer to avoid increasing heap memory for Solr shards to have enough to run other MapReduce jobs on the cluster. The rate of documents that I am currently inserting into these clouds per day is 5 Billion each in two clouds, 3 Billion into the third, and 2 Billion into the fourth ; however to account for capacity, the aim is to scale the solution to support double that amount of documents. To index these documents, there are MapReduce jobs that run that generate the Solr XML documents and will then submit these documents via SolrJ's CloudSolrServer interface. In testing, I have found that limiting the number of active parallel inserts to 80 per cloud gave
Re: Solr cloud performance degradation with billions of documents
You're using the term cloud again. Maybe that's the cause of your misunderstanding - SolrCloud probably should have been named SolrCluster since that's what it really is, a cluster rather than a cloud. The term cloud conjures up images of vast, unlimited numbers of nodes, thousands, tens of thousands of machines, but SolrCloud is much more modest than that. Again, start with a model of 100 million documents on a fairly commodity box (say, 32GB as opposed to expensive 16-core 256GB machines). So, 1 billion docs means 10 servers, times replication - I assume you want to serve a healthy query load. So, 5 billion docs needs 50 servers, times replication. 100 billion docs would require 1,000 servers. 500 billion documents would require 5,000 servers, times replication. Not quite Google class, but not a typical SolrCloud cluster either. You will have to test for yourself whether that 100 million number is achievable for your particular hardware and data. Maybe you can double it... or maybe only half of that. And, once again, make sure your index for each node fits in the OS system memory available for file caching. I haven't heard of any specific experiences of SolrCloud beyond dozens of nodes, but 64 nodes is probably a reasonable expectation for a SolrCloud cluster. How much bigger than that a SolrCloud cluster could grow is unknown. Whatever the actual practical limit, based on your own hardware, I/O, and network, and your own data schema and data patterns, which you will have to test for yourself, you will probably need to use an application layer to shard your 100s of billions to specific SolrCloud clusters. -- Jack Krupansky -Original Message- From: Wilburn, Scott Sent: Thursday, August 14, 2014 11:05 AM To: solr-user@lucene.apache.org Subject: RE: Solr cloud performance degradation with billions of documents Erick, Thanks for your suggestion to look into MapReduceIndexerTool, I'm looking into that now. I agree what I am trying to do is a tall order, and the more I hear from all of your comments, the more I am convinced that lack of memory is my biggest problem. I'm going to work on increasing the memory now, but was wondering if there are any configuration or other techniques that could also increase ingest performance? Does anyone know if a cloud of this size( hundreds of billions ) with an ingest rate of 5 billion new each day, has ever been attempted before? Thanks, Scott -Original Message- From: Erick Erickson [mailto:erickerick...@gmail.com] Sent: Wednesday, August 13, 2014 4:48 PM To: solr-user@lucene.apache.org Subject: Re: Solr cloud performance degradation with billions of documents Several points: 1 Have you considered using the MapReduceIndexerTool for your ingestion? Assuming you don't have duplicate IDs, i.e. each doc is new, you can spread your indexing across as many nodes as you have in your cluster. That said, it's not entirely clear that you'll gain throughput since you have as many nodes as you do. 2 Um, fitting this many documents into 6G of memory is ambitious. 2 Very ambitious. Actually it's impossible. By my calculations: bq: 4 separate and individual clouds of 32 shards each so 128 shards in aggregate bq: inserting into these clouds per day is 5 Billion each in two clouds, 3 Billion into the third, and 2 Billion into the fourth so we're talking 15B docs/day bq: the plan is to keep up to 60 days... So were talking 900B documents. It just won't work. 900B/128 docs/shard is over 7B documents/shard on average. Your two larger collections will have more than that, the two smaller ones less. But it doesn't matter because: 1: Lucene has a limit of 2B docs per core(shard), positive signed int. 2: It ain't gonna fit in 6G of memory even without this limit I'm pretty sure. 3: I've rarely heard of a single shard coping with over 300M docs without performance issues. I usually start getting nervous around 100M and insist on stress testing. Of course it depends lots on your query profile. So you're going to need a LOT more shards. You might be able to squeeze some more from your hardware by hosting multiple shards on for each collection on each machine, but I'm pretty sure your present setup is inadequate for your projected load. Of course I may be misinterpreting what you're saying hugely, but from what I understand this system just won't work. Best, Erick On Wed, Aug 13, 2014 at 2:39 PM, Markus Jelsma markus.jel...@openindex.io wrote: Hi - You are running mapred jobs on the same nodes as Solr runs right? The first thing i would think of is that your OS file buffer cache is abused. The mappers read all data, presumably residing on the same node. The mapper output and shuffling part would take place on the same node, only the reducer output is sent to your nodes, which i assume are on the same machines. Those same machines have a large Lucene index. All this data, written to and read from
RE: Solr cloud performance degradation with billions of documents
Wilburn, Scott [scott.wilb...@verizonwireless.com.INVALID] wrote: Thanks for your suggestion to look into MapReduceIndexerTool, I'm looking into that now. I agree what I am trying to do is a tall order, and the more I hear from all of your comments, the more I am convinced that lack of memory is my biggest problem. I'm going to work on increasing the memory now, but was wondering if there are any configuration or other techniques that could also increase ingest performance? More RAM basically compensates for slow storage, so the obvious trick is to increase your I/O performance. If your index is placed on network storage, then put it on local storage. If you are using spinning drives, then change to SSDs. If you are using SSDs then RAID them. Way cheaper than trying to match your RAM with your projected index size. Does anyone know if a cloud of this size( hundreds of billions ) with an ingest rate of 5 billion new each day, has ever been attempted before? Sorry, my experience is primarily with maximizing search performance. - Toke Eskildsen
Re: Solr cloud performance degradation with billions of documents
You are absolutely on the bleeding edge. I know of a couple of projects that are at that scale, but 1 they aren't being done on just a few nodes. As Jack says, this scale for SolrCloud is not common and there are no OOB templates to follow. 2 AFAIK, the projects I'm talking about aren't in production yet. And they're significant RD efforts on the parts of the companies involved. 3 You are _not_ going to do this on a shoestring budget. Nor is it going to be something you have up and running in 3 months. And you're talking a lot of machines here. Jack and I are both coming up with thousands of Solr servers, _that's_ the scale we're talking here! You're not going to get around this by just adding more memory either. Much as I love Solr, I have to ask whether it's the right tool for your situation. Unlike some other technologies, Solr requires holding large parts of the index in memory. For the entire corpus. At once. At the scale you're talking, you need compelling reasons to invest in all that. So I'd carefully look at what your problem is and whether Solr/search is the right tool for the job or not. On Thu, Aug 14, 2014 at 9:51 AM, Toke Eskildsen t...@statsbiblioteket.dk wrote: Wilburn, Scott [scott.wilb...@verizonwireless.com.INVALID] wrote: Thanks for your suggestion to look into MapReduceIndexerTool, I'm looking into that now. I agree what I am trying to do is a tall order, and the more I hear from all of your comments, the more I am convinced that lack of memory is my biggest problem. I'm going to work on increasing the memory now, but was wondering if there are any configuration or other techniques that could also increase ingest performance? More RAM basically compensates for slow storage, so the obvious trick is to increase your I/O performance. If your index is placed on network storage, then put it on local storage. If you are using spinning drives, then change to SSDs. If you are using SSDs then RAID them. Way cheaper than trying to match your RAM with your projected index size. Does anyone know if a cloud of this size( hundreds of billions ) with an ingest rate of 5 billion new each day, has ever been attempted before? Sorry, my experience is primarily with maximizing search performance. - Toke Eskildsen
RE: Solr cloud performance degradation with billions of documents
Thanks, Jack. I'd like to stay away from a terminology debate, since it is clear you know what I am talking about. But just to give my opinion, I prefer the term 'cloud' because it differentiates it from the term 'cluster', which refers to the Hadoop environment which I am running it on. I would also refrain from using the term 'node' when talking about Solr for the same reason. I already have this setup and running as described in my original email, with over 300 billion total records so far and counting, so changing my hardware configuration is really not an option( except adding more memory ). My issue is specific to keeping up with the volume of new documents, as my ingest rate is barely able to keep up, and I fear will eventually be perpetually latent as the amount of documents in the cloud/cluster continues to grow. I now have a few things to try, thanks to all of your comments. I am very appreciative. Thanks, Scott -Original Message- From: Jack Krupansky [mailto:j...@basetechnology.com] Sent: Thursday, August 14, 2014 8:31 AM To: solr-user@lucene.apache.org Subject: Re: Solr cloud performance degradation with billions of documents You're using the term cloud again. Maybe that's the cause of your misunderstanding - SolrCloud probably should have been named SolrCluster since that's what it really is, a cluster rather than a cloud. The term cloud conjures up images of vast, unlimited numbers of nodes, thousands, tens of thousands of machines, but SolrCloud is much more modest than that. Again, start with a model of 100 million documents on a fairly commodity box (say, 32GB as opposed to expensive 16-core 256GB machines). So, 1 billion docs means 10 servers, times replication - I assume you want to serve a healthy query load. So, 5 billion docs needs 50 servers, times replication. 100 billion docs would require 1,000 servers. 500 billion documents would require 5,000 servers, times replication. Not quite Google class, but not a typical SolrCloud cluster either. You will have to test for yourself whether that 100 million number is achievable for your particular hardware and data. Maybe you can double it... or maybe only half of that. And, once again, make sure your index for each node fits in the OS system memory available for file caching. I haven't heard of any specific experiences of SolrCloud beyond dozens of nodes, but 64 nodes is probably a reasonable expectation for a SolrCloud cluster. How much bigger than that a SolrCloud cluster could grow is unknown. Whatever the actual practical limit, based on your own hardware, I/O, and network, and your own data schema and data patterns, which you will have to test for yourself, you will probably need to use an application layer to shard your 100s of billions to specific SolrCloud clusters. -- Jack Krupansky -Original Message- From: Wilburn, Scott Sent: Thursday, August 14, 2014 11:05 AM To: solr-user@lucene.apache.org Subject: RE: Solr cloud performance degradation with billions of documents Erick, Thanks for your suggestion to look into MapReduceIndexerTool, I'm looking into that now. I agree what I am trying to do is a tall order, and the more I hear from all of your comments, the more I am convinced that lack of memory is my biggest problem. I'm going to work on increasing the memory now, but was wondering if there are any configuration or other techniques that could also increase ingest performance? Does anyone know if a cloud of this size( hundreds of billions ) with an ingest rate of 5 billion new each day, has ever been attempted before? Thanks, Scott -Original Message- From: Erick Erickson [mailto:erickerick...@gmail.com] Sent: Wednesday, August 13, 2014 4:48 PM To: solr-user@lucene.apache.org Subject: Re: Solr cloud performance degradation with billions of documents Several points: 1 Have you considered using the MapReduceIndexerTool for your ingestion? Assuming you don't have duplicate IDs, i.e. each doc is new, you can spread your indexing across as many nodes as you have in your cluster. That said, it's not entirely clear that you'll gain throughput since you have as many nodes as you do. 2 Um, fitting this many documents into 6G of memory is ambitious. 2 Very ambitious. Actually it's impossible. By my calculations: bq: 4 separate and individual clouds of 32 shards each so 128 shards in aggregate bq: inserting into these clouds per day is 5 Billion each in two clouds, 3 Billion into the third, and 2 Billion into the fourth so we're talking 15B docs/day bq: the plan is to keep up to 60 days... So were talking 900B documents. It just won't work. 900B/128 docs/shard is over 7B documents/shard on average. Your two larger collections will have more than that, the two smaller ones less. But it doesn't matter because: 1: Lucene has a limit of 2B docs per core(shard), positive signed int. 2: It ain't gonna fit in 6G of memory even without this limit I'm
RE: Solr cloud performance degradation with billions of documents
Erick Erickson [erickerick...@gmail.com] wrote: Solr requires holding large parts of the index in memory. For the entire corpus. At once. That requirement is under the assumption that one must have the lowest possible latency at each individual box. You might as well argue for the fastest possible memory or the fastest possible CPU being a requirement. The advice is good in some contexts and a waste of money in other. I not-so-humbly point to http://sbdevel.wordpress.com/2014/08/13/whale-hunting-with-solr/ where we (for simple searches) handily achieve our goal of sub-second response times for a 10TB index with just 1.4% of the index cached in RAM. Had our goal been sub-50ms, it would be another matter, but it is not. Just as Wilburn's problem is not to minimize latency for each individual box, but to achieve a certain throughput for indexing, while performing searches. Wilburn's hardware is currently able to keep up, although barely, with 300B documents. He needs to handle 900B. Tripling (or quadrupling) the amount of machines should do the trick. Increasing the amount of RAM on each current machine might also work (qua the well known effect of RAM with Lucene/Solr). Using local SSDs, if he is not doing so already, might also work (qua the article above). - Toke Eskildsen
Re: Solr cloud performance degradation with billions of documents
Could you clarify what you mean with the term cloud, as in per cloud and individual clouds? That's not a proper Solr or SolrCloud concept per se. SolrCloud works with a single cluster of nodes. And there is no interaction between separate SolrCloud clusters. -- Jack Krupansky -Original Message- From: Wilburn, Scott Sent: Wednesday, August 13, 2014 5:08 PM To: solr-user@lucene.apache.org Subject: Solr cloud performance degradation with billions of documents Hello everyone, I am trying to use SolrCloud to index a very large number of simple documents and have run into some performance and scalability limitations and was wondering what can be done about it. Hardware wise, I have a 32-node Hadoop cluster that I use to run all of the Solr shards and each node has 128GB of memory. The current SolrCloud setup is split into 4 separate and individual clouds of 32 shards each thereby giving four running shards per cloud or one cloud per eight nodes. Each shard is currently assigned a 6GB heap size. I’d prefer to avoid increasing heap memory for Solr shards to have enough to run other MapReduce jobs on the cluster. The rate of documents that I am currently inserting into these clouds per day is 5 Billion each in two clouds, 3 Billion into the third, and 2 Billion into the fourth ; however to account for capacity, the aim is to scale the solution to support double that amount of documents. To index these documents, there are MapReduce jobs that run that generate the Solr XML documents and will then submit these documents via SolrJ's CloudSolrServer interface. In testing, I have found that limiting the number of active parallel inserts to 80 per cloud gave the best performance as anything higher gave diminishing returns, most likely due to the constant shuffling of documents internally to SolrCloud. From an index perspective, dated collections are being created to hold an entire day's of documents and generally the inserting happens primarily on the current day (the previous days are only to allow for searching) and the plan is to keep up to 60 days (or collections) in each cloud. A single shard index in one collection in the busiest cloud currently takes up 30G disk space or 960G for the entire collection. The documents are being auto committed with a hard commit time of 4 minutes (opensearcher = false) and soft commit time of 8 minutes. From a search perspective, the use case is fairly generic and simple searches of the type :, so there is no need to tune the system to use any of the more advanced querying features. Therefore, the most important thing for me is to have the indexing performance be able to keep up with the rate of input. In the initial load testing, I was able to achieve a projected indexing rate of 10 Billion documents per cloud per day for a grand total of 40 Billion per day. However, the initial load testing was done on fairly empty clouds with just a few small collections. Now that there have been several days of documents being indexed, I am starting to see a fairly steep drop-off in indexing performance once the clouds reached about 15 full collections (or about 80-100 Billion documents per cloud) in the two biggest clouds. Based on current application logging I’m seeing a 40% drop off in indexing performance. Because of this, I have concerns on how performance will hold as more collections are added. My question to the community is if anyone else has had any experience in using Solr at this scale (hundreds of Billions) and if anyone has observed such a decline in indexing performance as the number of collections increases. My understanding is that each collection is a separate index and therefore the inserting rate should remain constant. Aside from that, what other tweaks or changes can be done in the SolrCloud configuration to increase the rate of indexing performance? Am I hitting a hard limitation of what Solr can handle? Thanks, Scott
RE: Solr cloud performance degradation with billions of documents
Thanks for replying Jack. I have 4 SolrCloud instances( or clusters ), each consisting of 32 shards. The clusters do not have any interaction with each other. Thanks, Scott -Original Message- From: Jack Krupansky [mailto:j...@basetechnology.com] Sent: Wednesday, August 13, 2014 2:17 PM To: solr-user@lucene.apache.org Subject: Re: Solr cloud performance degradation with billions of documents Could you clarify what you mean with the term cloud, as in per cloud and individual clouds? That's not a proper Solr or SolrCloud concept per se. SolrCloud works with a single cluster of nodes. And there is no interaction between separate SolrCloud clusters. -- Jack Krupansky -Original Message- From: Wilburn, Scott Sent: Wednesday, August 13, 2014 5:08 PM To: solr-user@lucene.apache.org Subject: Solr cloud performance degradation with billions of documents Hello everyone, I am trying to use SolrCloud to index a very large number of simple documents and have run into some performance and scalability limitations and was wondering what can be done about it. Hardware wise, I have a 32-node Hadoop cluster that I use to run all of the Solr shards and each node has 128GB of memory. The current SolrCloud setup is split into 4 separate and individual clouds of 32 shards each thereby giving four running shards per cloud or one cloud per eight nodes. Each shard is currently assigned a 6GB heap size. I’d prefer to avoid increasing heap memory for Solr shards to have enough to run other MapReduce jobs on the cluster. The rate of documents that I am currently inserting into these clouds per day is 5 Billion each in two clouds, 3 Billion into the third, and 2 Billion into the fourth ; however to account for capacity, the aim is to scale the solution to support double that amount of documents. To index these documents, there are MapReduce jobs that run that generate the Solr XML documents and will then submit these documents via SolrJ's CloudSolrServer interface. In testing, I have found that limiting the number of active parallel inserts to 80 per cloud gave the best performance as anything higher gave diminishing returns, most likely due to the constant shuffling of documents internally to SolrCloud. From an index perspective, dated collections are being created to hold an entire day's of documents and generally the inserting happens primarily on the current day (the previous days are only to allow for searching) and the plan is to keep up to 60 days (or collections) in each cloud. A single shard index in one collection in the busiest cloud currently takes up 30G disk space or 960G for the entire collection. The documents are being auto committed with a hard commit time of 4 minutes (opensearcher = false) and soft commit time of 8 minutes. From a search perspective, the use case is fairly generic and simple searches of the type :, so there is no need to tune the system to use any of the more advanced querying features. Therefore, the most important thing for me is to have the indexing performance be able to keep up with the rate of input. In the initial load testing, I was able to achieve a projected indexing rate of 10 Billion documents per cloud per day for a grand total of 40 Billion per day. However, the initial load testing was done on fairly empty clouds with just a few small collections. Now that there have been several days of documents being indexed, I am starting to see a fairly steep drop-off in indexing performance once the clouds reached about 15 full collections (or about 80-100 Billion documents per cloud) in the two biggest clouds. Based on current application logging I’m seeing a 40% drop off in indexing performance. Because of this, I have concerns on how performance will hold as more collections are added. My question to the community is if anyone else has had any experience in using Solr at this scale (hundreds of Billions) and if anyone has observed such a decline in indexing performance as the number of collections increases. My understanding is that each collection is a separate index and therefore the inserting rate should remain constant. Aside from that, what other tweaks or changes can be done in the SolrCloud configuration to increase the rate of indexing performance? Am I hitting a hard limitation of what Solr can handle? Thanks, Scott
RE: Solr cloud performance degradation with billions of documents
Wilburn, Scott [scott.wilb...@verizonwireless.com.INVALID] wrote: Hardware wise, I have a 32-node Hadoop cluster that I use to run all of the Solr shards and each node has 128GB of memory. The current SolrCloud setup is split into 4 separate and individual clouds of 32 shards each thereby giving four running shards per cloud or one cloud per eight nodes. You mean 4 running shards per node, right? With 6GB/shard that leaves about 100GB RAM for everything else on each node. [Snip: 10 billion insertions/day] That is nearly 4000 insertions/second per node. Quite a lot. A single shard index in one collection in the busiest cloud currently takes up 30G disk space or 960G for the entire collection. The documents are being auto committed with a hard commit time of 4 minutes (opensearcher = false) and soft commit time of 8 minutes. And you have 4 of these collections, so each node holds about 120GB of index with heavy updating? In the initial load testing, I was able to achieve a projected indexing rate of 10 Billion documents per cloud per day for a grand total of 40 Billion per day. However, the initial load testing was done on fairly empty clouds with just a few small collections. Now that there have been several days of documents being indexed, I am starting to see a fairly steep drop-off in indexing performance once the clouds reached about 15 full collections [...] If a single collection is 30GB and you have 15 now. That means your indexes takes up about 450GB on each node, which has less than 100GB free memory. Everything is not disk cached and since you are doing searches while you index, your indexer must compete for the disk cache. It seems natural that this would slow down indexing, with the slow down getting progressively worse as you fill the storage with active indexes. If you could isolate the old collections from the ones being updated, you could avoid this cache competition. Or you could of course throw more hardware at the problem: Are you stuck on spinning drives or are you using SSDs? - Toke Eskildsen
RE: Solr cloud performance degradation with billions of documents
Hi - You are running mapred jobs on the same nodes as Solr runs right? The first thing i would think of is that your OS file buffer cache is abused. The mappers read all data, presumably residing on the same node. The mapper output and shuffling part would take place on the same node, only the reducer output is sent to your nodes, which i assume are on the same machines. Those same machines have a large Lucene index. All this data, written to and read from the same disk, competes for a nice spot in the OS buffer cache. Forget it if i misread anything, but when you're using serious figures of size, then do not abuse your caches. Have a separate mapred and Solr cluster, because they both eat cache space. I assume you can see serious IO WAIT times. Split the stuff and maybe even use smaller hardware, but more. M -Original message- From:Wilburn, Scott scott.wilb...@verizonwireless.com.INVALID Sent: Wednesday 13th August 2014 23:09 To: solr-user@lucene.apache.org Subject: Solr cloud performance degradation with billions of documents Hello everyone, I am trying to use SolrCloud to index a very large number of simple documents and have run into some performance and scalability limitations and was wondering what can be done about it. Hardware wise, I have a 32-node Hadoop cluster that I use to run all of the Solr shards and each node has 128GB of memory. The current SolrCloud setup is split into 4 separate and individual clouds of 32 shards each thereby giving four running shards per cloud or one cloud per eight nodes. Each shard is currently assigned a 6GB heap size. I’d prefer to avoid increasing heap memory for Solr shards to have enough to run other MapReduce jobs on the cluster. The rate of documents that I am currently inserting into these clouds per day is 5 Billion each in two clouds, 3 Billion into the third, and 2 Billion into the fourth ; however to account for capacity, the aim is to scale the solution to support double that amount of documents. To index these documents, there are MapReduce jobs that run that generate the Solr XML documents and will then submit these documents via SolrJ's CloudSolrServer interface. In testing, I have found that limiting the number of active parallel inserts to 80 per cloud gave the best performance as anything higher gave diminishing returns, most likely due to the constant shuffling of documents internally to SolrCloud. From an index perspective, dated collections are being created to hold an entire day's of documents and generally the inserting happens primarily on the current day (the previous days are only to allow for searching) and the plan is to keep up to 60 days (or collections) in each cloud. A single shar d index in one collection in the busiest cloud currently takes up 30G disk space or 960G for the entire collection. The documents are being auto committed with a hard commit time of 4 minutes (opensearcher = false) and soft commit time of 8 minutes. From a search perspective, the use case is fairly generic and simple searches of the type :, so there is no need to tune the system to use any of the more advanced querying features. Therefore, the most important thing for me is to have the indexing performance be able to keep up with the rate of input. In the initial load testing, I was able to achieve a projected indexing rate of 10 Billion documents per cloud per day for a grand total of 40 Billion per day. However, the initial load testing was done on fairly empty clouds with just a few small collections. Now that there have been several days of documents being indexed, I am starting to see a fairly steep drop-off in indexing performance once the clouds reached about 15 full collections (or about 80-100 Billion documents per cloud) in the two biggest clouds. Based on current application logging I’m seeing a 40% drop off in indexing performance. Because of this, I have concerns on how performance will hold as more collections are added. My question to the community is if anyone else has had any experience in using Solr at this scale (hundreds of Billions) and if anyone has observed such a decline in indexing performance as the number of collections increases. My understanding is that each collection is a separate index and therefore the inserting rate should remain constant. Aside from that, what other tweaks or changes can be done in the SolrCloud configuration to increase the rate of indexing performance? Am I hitting a hard limitation of what Solr can handle? Thanks, Scott
Re: Solr cloud performance degradation with billions of documents
Be careful when you say instance - that usually refers to a single Solr node. Anyway... 32 shards - with a replication factor of 1? So, given your worst case here, 5 billion documents in a 32-node cluster, that's 156 million documents per node. What is the index size on a typical node? And how much system memory is available for caching of file reads? Generally, you want to have enough system memory to cache the full index. Or do you have SSD? But please clarify what you mean by about 80-100 Billion documents per cloud. Is it really 5 billion total, refreshed every day, or 5 billion added per day and lots of days stored? If you start seeing indexing rate drop off, that could be caused by not having enough RAM system memory to cache the full index. In particular, Lucene will occasionally be performing index merges, which would otherwise be I/O-intensive. I would start with a rule of thumb of 100 million documents per node (and that is million, not billion.) That could be a lot higher - or a lot lower - based on your actual schema and data value distribution. -- Jack Krupansky -Original Message- From: Wilburn, Scott Sent: Wednesday, August 13, 2014 5:42 PM To: solr-user@lucene.apache.org Subject: RE: Solr cloud performance degradation with billions of documents Thanks for replying Jack. I have 4 SolrCloud instances( or clusters ), each consisting of 32 shards. The clusters do not have any interaction with each other. Thanks, Scott -Original Message- From: Jack Krupansky [mailto:j...@basetechnology.com] Sent: Wednesday, August 13, 2014 2:17 PM To: solr-user@lucene.apache.org Subject: Re: Solr cloud performance degradation with billions of documents Could you clarify what you mean with the term cloud, as in per cloud and individual clouds? That's not a proper Solr or SolrCloud concept per se. SolrCloud works with a single cluster of nodes. And there is no interaction between separate SolrCloud clusters. -- Jack Krupansky -Original Message- From: Wilburn, Scott Sent: Wednesday, August 13, 2014 5:08 PM To: solr-user@lucene.apache.org Subject: Solr cloud performance degradation with billions of documents Hello everyone, I am trying to use SolrCloud to index a very large number of simple documents and have run into some performance and scalability limitations and was wondering what can be done about it. Hardware wise, I have a 32-node Hadoop cluster that I use to run all of the Solr shards and each node has 128GB of memory. The current SolrCloud setup is split into 4 separate and individual clouds of 32 shards each thereby giving four running shards per cloud or one cloud per eight nodes. Each shard is currently assigned a 6GB heap size. I’d prefer to avoid increasing heap memory for Solr shards to have enough to run other MapReduce jobs on the cluster. The rate of documents that I am currently inserting into these clouds per day is 5 Billion each in two clouds, 3 Billion into the third, and 2 Billion into the fourth ; however to account for capacity, the aim is to scale the solution to support double that amount of documents. To index these documents, there are MapReduce jobs that run that generate the Solr XML documents and will then submit these documents via SolrJ's CloudSolrServer interface. In testing, I have found that limiting the number of active parallel inserts to 80 per cloud gave the best performance as anything higher gave diminishing returns, most likely due to the constant shuffling of documents internally to SolrCloud. From an index perspective, dated collections are being created to hold an entire day's of documents and generally the inserting happens primarily on the current day (the previous days are only to allow for searching) and the plan is to keep up to 60 days (or collections) in each cloud. A single shard index in one collection in the busiest cloud currently takes up 30G disk space or 960G for the entire collection. The documents are being auto committed with a hard commit time of 4 minutes (opensearcher = false) and soft commit time of 8 minutes. From a search perspective, the use case is fairly generic and simple searches of the type :, so there is no need to tune the system to use any of the more advanced querying features. Therefore, the most important thing for me is to have the indexing performance be able to keep up with the rate of input. In the initial load testing, I was able to achieve a projected indexing rate of 10 Billion documents per cloud per day for a grand total of 40 Billion per day. However, the initial load testing was done on fairly empty clouds with just a few small collections. Now that there have been several days of documents being indexed, I am starting to see a fairly steep drop-off in indexing performance once the clouds reached about 15 full collections (or about 80-100 Billion documents per cloud) in the two biggest clouds. Based on current
Re: Solr cloud performance degradation with billions of documents
Several points: 1 Have you considered using the MapReduceIndexerTool for your ingestion? Assuming you don't have duplicate IDs, i.e. each doc is new, you can spread your indexing across as many nodes as you have in your cluster. That said, it's not entirely clear that you'll gain throughput since you have as many nodes as you do. 2 Um, fitting this many documents into 6G of memory is ambitious. Very ambitious. Actually it's impossible. By my calculations: bq: 4 separate and individual clouds of 32 shards each so 128 shards in aggregate bq: inserting into these clouds per day is 5 Billion each in two clouds, 3 Billion into the third, and 2 Billion into the fourth so we're talking 15B docs/day bq: the plan is to keep up to 60 days... So were talking 900B documents. It just won't work. 900B/128 docs/shard is over 7B documents/shard on average. Your two larger collections will have more than that, the two smaller ones less. But it doesn't matter because: 1: Lucene has a limit of 2B docs per core(shard), positive signed int. 2: It ain't gonna fit in 6G of memory even without this limit I'm pretty sure. 3: I've rarely heard of a single shard coping with over 300M docs without performance issues. I usually start getting nervous around 100M and insist on stress testing. Of course it depends lots on your query profile. So you're going to need a LOT more shards. You might be able to squeeze some more from your hardware by hosting multiple shards on for each collection on each machine, but I'm pretty sure your present setup is inadequate for your projected load. Of course I may be misinterpreting what you're saying hugely, but from what I understand this system just won't work. Best, Erick On Wed, Aug 13, 2014 at 2:39 PM, Markus Jelsma markus.jel...@openindex.io wrote: Hi - You are running mapred jobs on the same nodes as Solr runs right? The first thing i would think of is that your OS file buffer cache is abused. The mappers read all data, presumably residing on the same node. The mapper output and shuffling part would take place on the same node, only the reducer output is sent to your nodes, which i assume are on the same machines. Those same machines have a large Lucene index. All this data, written to and read from the same disk, competes for a nice spot in the OS buffer cache. Forget it if i misread anything, but when you're using serious figures of size, then do not abuse your caches. Have a separate mapred and Solr cluster, because they both eat cache space. I assume you can see serious IO WAIT times. Split the stuff and maybe even use smaller hardware, but more. M -Original message- From:Wilburn, Scott scott.wilb...@verizonwireless.com.INVALID Sent: Wednesday 13th August 2014 23:09 To: solr-user@lucene.apache.org Subject: Solr cloud performance degradation with billions of documents Hello everyone, I am trying to use SolrCloud to index a very large number of simple documents and have run into some performance and scalability limitations and was wondering what can be done about it. Hardware wise, I have a 32-node Hadoop cluster that I use to run all of the Solr shards and each node has 128GB of memory. The current SolrCloud setup is split into 4 separate and individual clouds of 32 shards each thereby giving four running shards per cloud or one cloud per eight nodes. Each shard is currently assigned a 6GB heap size. I’d prefer to avoid increasing heap memory for Solr shards to have enough to run other MapReduce jobs on the cluster. The rate of documents that I am currently inserting into these clouds per day is 5 Billion each in two clouds, 3 Billion into the third, and 2 Billion into the fourth ; however to account for capacity, the aim is to scale the solution to support double that amount of documents. To index these documents, there are MapReduce jobs that run that generate the Solr XML documents and will then submit these documents via SolrJ's CloudSolrServer interface. In testing, I have found that limiting the number of active parallel inserts to 80 per cloud gave the best performance as anything higher gave diminishing returns, most likely due to the constant shuffling of documents internally to SolrCloud. From an index perspective, dated collections are being created to hold an entire day's of documents and generally the inserting happens primarily on the current day (the previous days are only to allow for searching) and the plan is to keep up to 60 days (or collections) in each cloud. A single shar d index in one collection in the busiest cloud currently takes up 30G disk space or 960G for the entire collection. The documents are being auto committed with a hard commit time of 4 minutes (opensearcher = false) and soft commit time of 8 minutes. From a search perspective, the use case is fairly generic and simple searches of the type :, so there is no need to tune the system to use any of the more