Re: Best approach to handle large volume of documents with constantly high incoming rate?
Jack, thanks. Actually the 20K events/sec is some low-end rate we estimated. It is not necessarily related to sensor; when you want to centralize data from many sources, regardless multi-tenancy, even for a single tenant, many events per second have to be handled. I have a question regarding to the size of nodes used in Solr Cloud, what are the general pros/cons between using big or small nodes to setup Solr Clouds for similar cases as I described? For example, mainly considering memory: 256 (GB) x 4 vs. 32 (GB) x 32 or a little extreme: 256 (GB) x 4 vs. 8 (GB) x 128 Is it better to use fewer bigger nodes to setup a Solr Cloud or better to use more small nodes to setup a Solr Cloud? In the latter (a little extreme example), multiple Solr Clouds could be considered as Erick mentioned. Regards. Shushuai From: Jack Krupansky j...@basetechnology.com To: solr-user@lucene.apache.org Sent: Sunday, March 23, 2014 1:03 AM Subject: Re: Best approach to handle large volume of documents with constantly high incoming rate? I defer to Erick on on this level of detail and experience. Let's continue the discussion - some of it will be a matter of how to configure and tune Solr, how to select, configure, and tune hardware, the need for further Lucene/Solr improvements, and how much further we have to go to get to the next level with Big Data. I mean, 20K events/sec is not necessarily beyond the realm of reality these days with sensor data (20K/sec = 1 event every 50 microseconds) -- Jack Krupansky -Original Message- From: Erick Erickson Sent: Saturday, March 22, 2014 11:02 PM To: solr-user@lucene.apache.org ; shushuai zhu Subject: Re: Best approach to handle large volume of documents with constantly high incoming rate? Well, the commonsense limits Jack is referring to in that post are more (IMO) scales you should count on having to do some _serious_ prototyping/configuring/etc. As you scale out, you'll run into edge cases that aren't the common variety, aren't reliably tested every night, etc. I mean how would you set up a test bed that had 1,000 nodes? Sure, it can be done, but nobody's volunteered yet to provide the Apache Solr project that much hardware. I suspect that it would make Uwe's week if someone did though. In the practical limit vein, one example: You'll run up against the laggard problem. Let's assume that you successfully put up 2,000 nodes, for simplicity's sake, no replicas, just leaders and they all stay up all the time. To successfully do a search, you need to send out a request to all 2,000 nodes. The chance that one of them is slow for _any_ reason (GC, high CPU load, it's just tired) increases the more nodes you have. And since you have to wait until the slowest node responds, your query rate will suffer correspondingly. I've seen 4 node clusters handle 5,000 docs/sec update rate FWIW. YMMV of course. However, you say ...dedicated indexing servers There's no such thing in SolrCloud. Every document gets sent to every member of the slice it belongs to. How else could NRT be supported? When I saw that comment I wondered how well you understand SolrCloud. I flat guarantee you'll understand SolrCloud really, really well if yo try to scale as you indicate :). There'll be a whole bunch of learning experiences along the way, some will be painful. I guarantee that too. Responding to your points 1) Yes, no, and maybe. For relatively small docs on relatively modern hardware, it's a good place to start. Then you have to push it until it falls over to determine your _real_ rates. See: http://searchhub.org/dev/2012/07/23/sizing-hardware-in-the-abstract-why-we-dont-have-a-definitive-answer/ 2) Nobody knows. There's no theoretical reason why SolrCloud shouldn't; no a-priori hard limits. I _strongly_ suspect you'll be on the bleeding edge of size, though. Expect some things to be learning experiences. 3) No, it doesn't mean that at all. 64 is an arbitrary number that means, IMO, here there be dragons. As you start to scale out beyond this you'll run into pesky issues I expect. Your network won't be as reliable as you think. You'll find one of your VMs (which I expect you'll be running on) has some glitches. Someone loaded a very CPU intensive program on three of your machines and your Solrs on those machines is being starved. Etc. 4) I've personally seen 1,000 node clusters. You ought to see the very cool. SolrCloud admin graph I recently saw... But I expect you'll actually be in for some kind of divide-and-conquer strategy whereby you have a bunch of clusters that are significantly smaller. You could, for instance, determine that the use-case you support is searching across small ranges, say a week at a time and have 52 clusters of 128 machines or so. You could have 365 clusters of 20 machines. It all depends on how the index will be used. 5) Not at all. See above, I've seen 5K/sec on 4 nodes, also supporting simultaneous
Re: Best approach to handle large volume of documents with constantly high incoming rate?
Any thoughts? Can Solr Cloud support such use case with acceptable performance? On Thursday, March 20, 2014 7:51 PM, shushuai zhu ss...@yahoo.com wrote: Hi, I am looking for some advice to handle large volume of documents with a very high incoming rate. The size of each document is about 0.5 KB and the incoming rate could be more than 20K per second and we want to store about one year's documents in Solr for near real=time searching. The goal is to achieve acceptable indexing and querying performance. We will use techniques like soft commit, dedicated indexing servers, etc. My main question is about how to structure the collection/shard/core to achieve the goals. Since the incoming rate is very high, we do not want the incoming documents to affect the existing older indexes. Some thoughts are to create a latest index to hold the incoming documents (say latest half hour's data, about 36M docs) so queries on older data could be faster since the old indexes are not affected. There seem three ways to grow the time dimension by adding/splitting/creating a new object listed below every half hour: collection shard core Which is the best way to grow the time dimension? Any limitation in that direction? Or there is some better approach? As an example, I am thinking about having 4 nodes with the following configuration to setup a Solr Cloud: Memory: 128 GB Storage: 4 TB How to set the collection/shard/core to deal with the use case? Thanks in advance. Shushuai
Re: Best approach to handle large volume of documents with constantly high incoming rate?
20K docs/sec = 20,000 * 60 * 60 * 24 = 1,728,000,000 = 1.7 billion docs/day * 365 = 630,720,000,000 = 631 billion docs/yr At 100 million docs/node = 6,308 nodes! And you think you can do it with 4 nodes? Oh, and that's before replication! 0.5K/doc * 631 billion docs = 322 TB. -- Jack Krupansky -Original Message- From: shushuai zhu Sent: Saturday, March 22, 2014 11:32 AM To: solr-user@lucene.apache.org Subject: Re: Best approach to handle large volume of documents with constantly high incoming rate? Any thoughts? Can Solr Cloud support such use case with acceptable performance? On Thursday, March 20, 2014 7:51 PM, shushuai zhu ss...@yahoo.com wrote: Hi, I am looking for some advice to handle large volume of documents with a very high incoming rate. The size of each document is about 0.5 KB and the incoming rate could be more than 20K per second and we want to store about one year's documents in Solr for near real=time searching. The goal is to achieve acceptable indexing and querying performance. We will use techniques like soft commit, dedicated indexing servers, etc. My main question is about how to structure the collection/shard/core to achieve the goals. Since the incoming rate is very high, we do not want the incoming documents to affect the existing older indexes. Some thoughts are to create a latest index to hold the incoming documents (say latest half hour's data, about 36M docs) so queries on older data could be faster since the old indexes are not affected. There seem three ways to grow the time dimension by adding/splitting/creating a new object listed below every half hour: collection shard core Which is the best way to grow the time dimension? Any limitation in that direction? Or there is some better approach? As an example, I am thinking about having 4 nodes with the following configuration to setup a Solr Cloud: Memory: 128 GB Storage: 4 TB How to set the collection/shard/core to deal with the use case? Thanks in advance. Shushuai
Re: Best approach to handle large volume of documents with constantly high incoming rate?
Jack, thanks for your reply. Sorry for the confusion about 4 nodes. What I meant was to use 4 nodes to do some POC, mainly focusing on handling the high incoming rate in a few days instead of storing data over one year. You estimated the required nodes (6,308) and storage (322TB) based on the incoming rate and doc size. I have a few questions regarding to them: 1) Is 100 million docs/node some general capacity guideline for a Solr node? 2) Assuming we can provide 6,308 nodes, can Solr Cloud really scale to that level? I found you indicated some common sense limits of Solr Cluster size of 64 nodes in the following mail thread http://find.searchhub.org/document/d823643e65fe2015#84f0c89df2426990 3) If 64 nodes are something we know Solr Cloud can scale up to, then does it mean I can only be sure that 1% of the mentioned workload can be handle by Solr Cloud? (64 is about 1% of 6,308 nodes) 4) The above mentioned Solr Limitations mail thread did mention some cluster with 512 nodes but not really verified whether it worked or not; assuming it worked, it just means we may be able to handle a little less than 10% of the desired workload. 5) Given above simple deduction, it seems 2K docs/sec (10% of the mentioned incoming rate) is the practical limitation of Solr Cloud we can guess for our use case? 6) If the incoming rate is controlled to be around 1k or 2k docs/sec and we want to use Solr Cluster with 64 nodes (or more if it still works), what kind of collection/shard/core structure should be? I am more looking for architectural advice regarding to Solr Cloud structure to handle high incoming rate of relatively small docs. Regards. Shushuai On Saturday, March 22, 2014 2:17 PM, Jack Krupansky j...@basetechnology.com wrote: 20K docs/sec = 20,000 * 60 * 60 * 24 = 1,728,000,000 = 1.7 billion docs/day * 365 = 630,720,000,000 = 631 billion docs/yr At 100 million docs/node = 6,308 nodes! And you think you can do it with 4 nodes? Oh, and that's before replication! 0.5K/doc * 631 billion docs = 322 TB. -- Jack Krupansky -Original Message- From: shushuai zhu Sent: Saturday, March 22, 2014 11:32 AM To: solr-user@lucene.apache.org Subject: Re: Best approach to handle large volume of documents with constantly high incoming rate? Any thoughts? Can Solr Cloud support such use case with acceptable performance? On Thursday, March 20, 2014 7:51 PM, shushuai zhu ss...@yahoo.com wrote: Hi, I am looking for some advice to handle large volume of documents with a very high incoming rate. The size of each document is about 0.5 KB and the incoming rate could be more than 20K per second and we want to store about one year's documents in Solr for near real=time searching. The goal is to achieve acceptable indexing and querying performance. We will use techniques like soft commit, dedicated indexing servers, etc. My main question is about how to structure the collection/shard/core to achieve the goals. Since the incoming rate is very high, we do not want the incoming documents to affect the existing older indexes. Some thoughts are to create a latest index to hold the incoming documents (say latest half hour's data, about 36M docs) so queries on older data could be faster since the old indexes are not affected. There seem three ways to grow the time dimension by adding/splitting/creating a new object listed below every half hour: collection shard core Which is the best way to grow the time dimension? Any limitation in that direction? Or there is some better approach? As an example, I am thinking about having 4 nodes with the following configuration to setup a Solr Cloud: Memory: 128 GB Storage: 4 TB How to set the collection/shard/core to deal with the use case? Thanks in advance. Shushuai
Re: Best approach to handle large volume of documents with constantly high incoming rate?
Well, the commonsense limits Jack is referring to in that post are more (IMO) scales you should count on having to do some _serious_ prototyping/configuring/etc. As you scale out, you'll run into edge cases that aren't the common variety, aren't reliably tested every night, etc. I mean how would you set up a test bed that had 1,000 nodes? Sure, it can be done, but nobody's volunteered yet to provide the Apache Solr project that much hardware. I suspect that it would make Uwe's week if someone did though. In the practical limit vein, one example: You'll run up against the laggard problem. Let's assume that you successfully put up 2,000 nodes, for simplicity's sake, no replicas, just leaders and they all stay up all the time. To successfully do a search, you need to send out a request to all 2,000 nodes. The chance that one of them is slow for _any_ reason (GC, high CPU load, it's just tired) increases the more nodes you have. And since you have to wait until the slowest node responds, your query rate will suffer correspondingly. I've seen 4 node clusters handle 5,000 docs/sec update rate FWIW. YMMV of course. However, you say ...dedicated indexing servers There's no such thing in SolrCloud. Every document gets sent to every member of the slice it belongs to. How else could NRT be supported? When I saw that comment I wondered how well you understand SolrCloud. I flat guarantee you'll understand SolrCloud really, really well if yo try to scale as you indicate :). There'll be a whole bunch of learning experiences along the way, some will be painful. I guarantee that too. Responding to your points 1) Yes, no, and maybe. For relatively small docs on relatively modern hardware, it's a good place to start. Then you have to push it until it falls over to determine your _real_ rates. See: http://searchhub.org/dev/2012/07/23/sizing-hardware-in-the-abstract-why-we-dont-have-a-definitive-answer/ 2) Nobody knows. There's no theoretical reason why SolrCloud shouldn't; no a-priori hard limits. I _strongly_ suspect you'll be on the bleeding edge of size, though. Expect some things to be learning experiences. 3) No, it doesn't mean that at all. 64 is an arbitrary number that means, IMO, here there be dragons. As you start to scale out beyond this you'll run into pesky issues I expect. Your network won't be as reliable as you think. You'll find one of your VMs (which I expect you'll be running on) has some glitches. Someone loaded a very CPU intensive program on three of your machines and your Solrs on those machines is being starved. Etc. 4) I've personally seen 1,000 node clusters. You ought to see the very cool. SolrCloud admin graph I recently saw... But I expect you'll actually be in for some kind of divide-and-conquer strategy whereby you have a bunch of clusters that are significantly smaller. You could, for instance, determine that the use-case you support is searching across small ranges, say a week at a time and have 52 clusters of 128 machines or so. You could have 365 clusters of 20 machines. It all depends on how the index will be used. 5) Not at all. See above, I've seen 5K/sec on 4 nodes, also supporting simultaneous searching. 6) N/A I really can't give you advice. You haven't, for instance, said anything about searches. What kind of SLA are you aiming for? What kind of queries? Faceting? Grouping? I can change the memory footprint of Solr by firing off really ugly queries. In essence, you absolutely must prototype, see the link above. And do a lot of homework defining how you will search the corpus. Otherwise you're guessing. But at this kind of scale, expect to do something other than throw all the docs at a 64 node cluster and expect it to just work. It'll be a lot of work. On Sat, Mar 22, 2014 at 6:48 PM, shushuai zhu ss...@yahoo.com wrote: Jack, thanks for your reply. Sorry for the confusion about 4 nodes. What I meant was to use 4 nodes to do some POC, mainly focusing on handling the high incoming rate in a few days instead of storing data over one year. You estimated the required nodes (6,308) and storage (322TB) based on the incoming rate and doc size. I have a few questions regarding to them: 1) Is 100 million docs/node some general capacity guideline for a Solr node? 2) Assuming we can provide 6,308 nodes, can Solr Cloud really scale to that level? I found you indicated some common sense limits of Solr Cluster size of 64 nodes in the following mail thread http://find.searchhub.org/document/d823643e65fe2015#84f0c89df2426990 3) If 64 nodes are something we know Solr Cloud can scale up to, then does it mean I can only be sure that 1% of the mentioned workload can be handle by Solr Cloud? (64 is about 1% of 6,308 nodes) 4) The above mentioned Solr Limitations mail thread did mention some cluster with 512 nodes but not really verified whether it worked or not; assuming it worked, it just means we may be able to handle a little less than 10% of the desired
Re: Best approach to handle large volume of documents with constantly high incoming rate?
Erick, Thanks a lot for the detailed answers. They are very helpful and I do get some idea from them. As per our searches, we will mainly do term and field (AND/OR) searches, histogram, and faceting. Generally the queries are bound by time (e.g, last hour, last day, last week, or even last month). The queries are not complicated but the challenge is to support near real-time queries with high incoming rate. I have to admit that I just started looking into Solr although I used ElasticSearch for a little while. By default, ElasticSearch creates daily indexes to scale in time dimension and that is the reason I asked if I could customize Solr to do some similar thing to scale out in time dimension. I did notice the following slides talking about scaling time via multiple collections: http://www.slideshare.net/sematext/solr-for-indexing-and-searching-logs but I found more discussions about the limited number of collections Solr can support (not more than 1000 mainly due to znode 1 MB limit from zookeeper?). So, I feel it might be better to scale time via multiple shards or cores (Solr has lotsOfCores feature). Will appreciate very much if more architectural advice could be given for this use case. Regards. Shushuai On Saturday, March 22, 2014 11:10 PM, Erick Erickson erickerick...@gmail.com wrote: Well, the commonsense limits Jack is referring to in that post are more (IMO) scales you should count on having to do some _serious_ prototyping/configuring/etc. As you scale out, you'll run into edge cases that aren't the common variety, aren't reliably tested every night, etc. I mean how would you set up a test bed that had 1,000 nodes? Sure, it can be done, but nobody's volunteered yet to provide the Apache Solr project that much hardware. I suspect that it would make Uwe's week if someone did though. In the practical limit vein, one example: You'll run up against the laggard problem. Let's assume that you successfully put up 2,000 nodes, for simplicity's sake, no replicas, just leaders and they all stay up all the time. To successfully do a search, you need to send out a request to all 2,000 nodes. The chance that one of them is slow for _any_ reason (GC, high CPU load, it's just tired) increases the more nodes you have. And since you have to wait until the slowest node responds, your query rate will suffer correspondingly. I've seen 4 node clusters handle 5,000 docs/sec update rate FWIW. YMMV of course. However, you say ...dedicated indexing servers There's no such thing in SolrCloud. Every document gets sent to every member of the slice it belongs to. How else could NRT be supported? When I saw that comment I wondered how well you understand SolrCloud. I flat guarantee you'll understand SolrCloud really, really well if yo try to scale as you indicate :). There'll be a whole bunch of learning experiences along the way, some will be painful. I guarantee that too. Responding to your points 1) Yes, no, and maybe. For relatively small docs on relatively modern hardware, it's a good place to start. Then you have to push it until it falls over to determine your _real_ rates. See: http://searchhub.org/dev/2012/07/23/sizing-hardware-in-the-abstract-why-we-dont-have-a-definitive-answer/ 2) Nobody knows. There's no theoretical reason why SolrCloud shouldn't; no a-priori hard limits. I _strongly_ suspect you'll be on the bleeding edge of size, though. Expect some things to be learning experiences. 3) No, it doesn't mean that at all. 64 is an arbitrary number that means, IMO, here there be dragons. As you start to scale out beyond this you'll run into pesky issues I expect. Your network won't be as reliable as you think. You'll find one of your VMs (which I expect you'll be running on) has some glitches. Someone loaded a very CPU intensive program on three of your machines and your Solrs on those machines is being starved. Etc. 4) I've personally seen 1,000 node clusters. You ought to see the very cool. SolrCloud admin graph I recently saw... But I expect you'll actually be in for some kind of divide-and-conquer strategy whereby you have a bunch of clusters that are significantly smaller. You could, for instance, determine that the use-case you support is searching across small ranges, say a week at a time and have 52 clusters of 128 machines or so. You could have 365 clusters of 20 machines. It all depends on how the index will be used. 5) Not at all. See above, I've seen 5K/sec on 4 nodes, also supporting simultaneous searching. 6) N/A I really can't give you advice. You haven't, for instance, said anything about searches. What kind of SLA are you aiming for? What kind of queries? Faceting? Grouping? I can change the memory footprint of Solr by firing off really ugly queries. In essence, you absolutely must prototype, see the link above. And do a lot of homework defining how you will search the corpus. Otherwise you're guessing. But at this kind of scale, expect to do
Re: Best approach to handle large volume of documents with constantly high incoming rate?
I defer to Erick on on this level of detail and experience. Let's continue the discussion - some of it will be a matter of how to configure and tune Solr, how to select, configure, and tune hardware, the need for further Lucene/Solr improvements, and how much further we have to go to get to the next level with Big Data. I mean, 20K events/sec is not necessarily beyond the realm of reality these days with sensor data (20K/sec = 1 event every 50 microseconds) -- Jack Krupansky -Original Message- From: Erick Erickson Sent: Saturday, March 22, 2014 11:02 PM To: solr-user@lucene.apache.org ; shushuai zhu Subject: Re: Best approach to handle large volume of documents with constantly high incoming rate? Well, the commonsense limits Jack is referring to in that post are more (IMO) scales you should count on having to do some _serious_ prototyping/configuring/etc. As you scale out, you'll run into edge cases that aren't the common variety, aren't reliably tested every night, etc. I mean how would you set up a test bed that had 1,000 nodes? Sure, it can be done, but nobody's volunteered yet to provide the Apache Solr project that much hardware. I suspect that it would make Uwe's week if someone did though. In the practical limit vein, one example: You'll run up against the laggard problem. Let's assume that you successfully put up 2,000 nodes, for simplicity's sake, no replicas, just leaders and they all stay up all the time. To successfully do a search, you need to send out a request to all 2,000 nodes. The chance that one of them is slow for _any_ reason (GC, high CPU load, it's just tired) increases the more nodes you have. And since you have to wait until the slowest node responds, your query rate will suffer correspondingly. I've seen 4 node clusters handle 5,000 docs/sec update rate FWIW. YMMV of course. However, you say ...dedicated indexing servers There's no such thing in SolrCloud. Every document gets sent to every member of the slice it belongs to. How else could NRT be supported? When I saw that comment I wondered how well you understand SolrCloud. I flat guarantee you'll understand SolrCloud really, really well if yo try to scale as you indicate :). There'll be a whole bunch of learning experiences along the way, some will be painful. I guarantee that too. Responding to your points 1) Yes, no, and maybe. For relatively small docs on relatively modern hardware, it's a good place to start. Then you have to push it until it falls over to determine your _real_ rates. See: http://searchhub.org/dev/2012/07/23/sizing-hardware-in-the-abstract-why-we-dont-have-a-definitive-answer/ 2) Nobody knows. There's no theoretical reason why SolrCloud shouldn't; no a-priori hard limits. I _strongly_ suspect you'll be on the bleeding edge of size, though. Expect some things to be learning experiences. 3) No, it doesn't mean that at all. 64 is an arbitrary number that means, IMO, here there be dragons. As you start to scale out beyond this you'll run into pesky issues I expect. Your network won't be as reliable as you think. You'll find one of your VMs (which I expect you'll be running on) has some glitches. Someone loaded a very CPU intensive program on three of your machines and your Solrs on those machines is being starved. Etc. 4) I've personally seen 1,000 node clusters. You ought to see the very cool. SolrCloud admin graph I recently saw... But I expect you'll actually be in for some kind of divide-and-conquer strategy whereby you have a bunch of clusters that are significantly smaller. You could, for instance, determine that the use-case you support is searching across small ranges, say a week at a time and have 52 clusters of 128 machines or so. You could have 365 clusters of 20 machines. It all depends on how the index will be used. 5) Not at all. See above, I've seen 5K/sec on 4 nodes, also supporting simultaneous searching. 6) N/A I really can't give you advice. You haven't, for instance, said anything about searches. What kind of SLA are you aiming for? What kind of queries? Faceting? Grouping? I can change the memory footprint of Solr by firing off really ugly queries. In essence, you absolutely must prototype, see the link above. And do a lot of homework defining how you will search the corpus. Otherwise you're guessing. But at this kind of scale, expect to do something other than throw all the docs at a 64 node cluster and expect it to just work. It'll be a lot of work. On Sat, Mar 22, 2014 at 6:48 PM, shushuai zhu ss...@yahoo.com wrote: Jack, thanks for your reply. Sorry for the confusion about 4 nodes. What I meant was to use 4 nodes to do some POC, mainly focusing on handling the high incoming rate in a few days instead of storing data over one year. You estimated the required nodes (6,308) and storage (322TB) based on the incoming rate and doc size. I have a few questions regarding to them: 1) Is 100 million docs/node some general capacity guideline for a
Best approach to handle large volume of documents with constantly high incoming rate?
Hi, I am looking for some advice to handle large volume of documents with a very high incoming rate. The size of each document is about 0.5 KB and the incoming rate could be more than 20K per second and we want to store about one year's documents in Solr for near real=time searching. The goal is to achieve acceptable indexing and querying performance. We will use techniques like soft commit, dedicated indexing servers, etc. My main question is about how to structure the collection/shard/core to achieve the goals. Since the incoming rate is very high, we do not want the incoming documents to affect the existing older indexes. Some thoughts are to create a latest index to hold the incoming documents (say latest half hour's data, about 36M docs) so queries on older data could be faster since the old indexes are not affected. There seem three ways to grow the time dimension by adding/splitting/creating a new object listed below every half hour: collection shard core Which is the best way to grow the time dimension? Any limitation in that direction? Or there is some better approach? As an example, I am thinking about having 4 nodes with the following configuration to setup a Solr Cloud: Memory: 128 GB Storage: 4 TB How to set the collection/shard/core to deal with the use case? Thanks in advance. Shushuai