You are viewing a plain text version of this content. The canonical link for it is here.
Posted to solr-user@lucene.apache.org by "Kaminski, Adi" <Ad...@verint.com> on 2019/08/01 06:11:49 UTC
RE: SolrCloud recommended I/O RAID level
Hi Shawn,
Thanks for your reply, fully agree with your comments, it clarifies more the need of RAID10 in this case.
One additional follow-up question - in case we follow this guidelines and having RAID10 (which leaves us with effective capacity of 50%), why would I need
replication factor of 2 in our SolrCloud core/collection ? Won't it be double protection layer, while the IO layer mirroring of RAID10 actually brings the value, and no need to copy anything
when we have IO failures ?
Thanks,
Adi
-----Original Message-----
From: Shawn Heisey <ap...@elyograg.org>
Sent: Tuesday, July 30, 2019 9:44 PM
To: solr-user@lucene.apache.org
Subject: Re: SolrCloud recommended I/O RAID level
On 7/30/2019 12:12 PM, Kaminski, Adi wrote:
> Indeed RAID10 with both mirroring and striping should satisfy the
> need, but per some benchmarks in the network there is still an impact
> on write performance on it compared to RAID0 which is considered as
> much better (attaching a table that summarizes different RAID levels
> and their pros/cons and capacity ratio).
RAID10 offers the best combination of performance and reliability.
RAID0 might beat it *slightly* on performance, but if ANY drive fails on RAID0, the entire volume is lost.
> If we have ~200-320 shards spread by our 7 Solr node servers (part of
> SolrCloud cluster) on single core/collection configured with
> replication factor 2, shouldn't it supply applicative level redundancy
> of indexed data ?
Yes, you could rely on Solr alone for data redundancy. But if there's a drive failure, do you REALLY want to be single-stranded for the time it takes to rebuild the entire server and copy data? That's what you would end up doing if you choose RAID0.
It is true that RAID1 or RAID10 means you have to buy double your usable capacity. I would argue that drives are cheap and will cost less than either downtime or sysadmin effort.
Thanks,
Shawn
This electronic message may contain proprietary and confidential information of Verint Systems Inc., its affiliates and/or subsidiaries. The information is intended to be for the use of the individual(s) or entity(ies) named above. If you are not the intended recipient (or authorized to receive this e-mail for the intended recipient), you may not use, copy, disclose or distribute to anyone this message or any information contained in this message. If you have received this electronic message in error, please notify us by replying to this e-mail.
RE: SolrCloud recommended I/O RAID level
Posted by "Kaminski, Adi" <Ad...@verint.com>.
Erick and Shawn - thanks for the detailed explanation.
We will try the number per your guidance that represents the best both ingestion flow (indexing documents rate) and searching flow and good QPS.
We just need to start from some number in order to size the lab accordingly, since it take quite of time to index 40M parent documents with 50-80 nested child documents, which eventually results 2-3.2B of SOLR documents.
The indexing rate on production is expected to have 5M parent documents per month, with results 250-400M Solr documents per months.
The thought about shard per 10M Solr document, in order to eventually utilize the ~400 vCPUs we'll have better, and having replication factor of 2 (to allow high availability of our data in Solr search engine layer).
Per you answers, it seems we should consider larger documents amount per shard, and then tune accordingly.
We'll take all these into consideration in order to size and simulate both indexing and search flow in load.
Thanks again,
Adi
-----Original Message-----
From: Erick Erickson <er...@gmail.com>
Sent: Sunday, August 4, 2019 6:51 PM
To: solr-user@lucene.apache.org
Subject: Re: SolrCloud recommended I/O RAID level
Shards are really not about maximizing QPS, they’re more about fitting all your data on your machines. Yes, there’s a tradeoff where if you have more shards, each is working on a smaller data set so potentially will work faster, but there’s a corresponding offset of having to do the bookkeeping. By that I mean:
- A request comes into some node, call it the “aggregator”. It’s just one of the Solr nodes.
- A sub-request is sent to one replica of every shard that returns the top N docs from each (i.e. the rows parameter).
- The aggregator sorts all the responses to get the real top N
- For each of the final top N, the aggregator sends a request to the replica it came from to get the return fields.
That’s compounded by the “laggard” problem. In the above, the aggregator can’t know the true top N until it gets a response from a replica from _every_ shard. If even one of those replicas hits a GC pause, network glitch, or just happens to be very, vey busy, it’ll delay the entire response. The more replicas the aggregator has to query, the more likely one of them (and it only takes one) will be slow.
So there’s no easy answer. The first question you must answer is how many docs you can fit on a shard. Find that number. Is the QPS fast enough with that number? If so, that’ll give you the number of shards you need (leave some room for growth of course!). If a shard can host lots and lots of doc but the response time is too slow, shard some more. etc.
Best,
Erick
> On Aug 4, 2019, at 10:53 AM, Kaminski, Adi <Ad...@verint.com> wrote:
>
> Erick - thanks a lot for answering and sharing the below article, it's very helpful !
>
> I have another follow-up question - assuming we have 400 vCPUs across
> our SolrCloud cluster nodes, will it be better to have 400 shards with replication factor 2 or 200 shards with replication factor 4 ? What utilizes better the CPUs - shards/or their replicas ?
>
> Thanks,
> Adi
>
> -----Original Message-----
> From: Erick Erickson <er...@gmail.com>
> Sent: Thursday, August 1, 2019 6:48 PM
> To: solr-user@lucene.apache.org
> Subject: Re: SolrCloud recommended I/O RAID level
>
> Yes, I’m exactly talking about the idea of someone pulling the plug on one of your machines. RAID doesn’t help you at all with that.
>
> 10M docs/shard is pretty low actually, depending on the
> characteristics of docs of course. I’ve ween 10M docs strain things.
> I’ve seen 300M docs fit in 16G heap. My personal straw-man starting
> point is 50M docs/shard. You have to test. Old but still valid:
> https://lucidworks.com/post/sizing-hardware-in-the-abstract-why-we-don
> t-have-a-definitive-answer/
>
> bq. Will additional replicas will assist with QPS
>
> It Depends (tm). The article I linked above will help you answer the question, the only way I know of to answer definitively is to stress-test on a small (2 nodes will do) cluster. Obviously if you’ve maxed out all your CPUs with no replicas, adding more replicas won’t help. But first you have to convince me that you’re maxing out all your CPUs.
>
> Test, test, test.
>
> Best,
> Erick
>
>> On Aug 1, 2019, at 7:35 AM, Kaminski, Adi <Ad...@verint.com> wrote:
>>
>> Hi Erick,
>>
>> Thanks for the detailed explanation, much appreciated.
>>
>>
>>
>> Regarding #1 reason you have mentioned - but if RAID10 is chosen per
>> Shawn's suggestion, then we should be protected, no ? Unless you mean the whole server is not available and not specific disk of configured RAID10.
>>
>>
>>
>> Regarding #2 reason you have mentioned -
>>
>> · If we’re having 7 Solr node/servers as part of SolrCloud cluster, with 48 vCPU in each server (so in total ~336 vCPU in total)
>>
>> · And we are talking about 2-3.2B of Solr docs distributed on this cluster
>>
>> · And assuming that each shard will handled 10M Solr docs in average (that what we saw is some forums as basic “rule of thumb”)
>>
>> · Questions:
>>
>> o Will additional replicas will assist with QPS ? As we will basically are having ratio of 1 vCPU per 1 shard (the mster replica)
>>
>> o Wouldn’t additional replicas will overclock per total amout of vCPUs, and actually cause delays in queries and lower QPS ?
>>
>>
>>
>> Thanks,
>>
>> Adi
>>
>>
>>
>> -----Original Message-----
>> From: Erick Erickson <er...@gmail.com>
>> Sent: Thursday, August 1, 2019 2:03 PM
>> To: solr-user@lucene.apache.org
>> Subject: Re: SolrCloud recommended I/O RAID level
>>
>>
>>
>> “why would I need a replication factor of 2….”
>>
>>
>>
>> Two reasons:
>>
>>
>>
>> 1> assuming your replicas are located on different physical machines, you can lose one and still keep on running. RAID won’t help if the machine dies.
>>
>>
>>
>> 2> capacity. Depending on your query load, having more than one replica distributes query loads across multiple Solr instances, increasing the QPS you can handle.
>>
>>
>>
>> If you don’t care about either of those issues, then no reason.
>>
>>
>>
>> Best,
>>
>> Erick
>>
>>
>>
>>> On Aug 1, 2019, at 2:11 AM, Kaminski, Adi <Ad...@verint.com>> wrote:
>>
>>>
>>
>>> Hi Shawn,
>>
>>> Thanks for your reply, fully agree with your comments, it clarifies more the need of RAID10 in this case.
>>
>>>
>>
>>> One additional follow-up question - in case we follow this
>>> guidelines
>>
>>> and having RAID10 (which leaves us with effective capacity of 50%),
>>
>>> why would I need replication factor of 2 in our SolrCloud core/collection ? Won't it be double protection layer, while the IO layer mirroring of RAID10 actually brings the value, and no need to copy anything when we have IO failures ?
>>
>>>
>>
>>> Thanks,
>>
>>> Adi
>>
>>>
>>
>>> -----Original Message-----
>>
>>> From: Shawn Heisey <ap...@elyograg.org>>
>>
>>> Sent: Tuesday, July 30, 2019 9:44 PM
>>
>>> To: solr-user@lucene.apache.org<ma...@lucene.apache.org>
>>
>>> Subject: Re: SolrCloud recommended I/O RAID level
>>
>>>
>>
>>> On 7/30/2019 12:12 PM, Kaminski, Adi wrote:
>>
>>>> Indeed RAID10 with both mirroring and striping should satisfy the
>>
>>>> need, but per some benchmarks in the network there is still an
>>>> impact
>>
>>>> on write performance on it compared to RAID0 which is considered as
>>
>>>> much better (attaching a table that summarizes different RAID
>>>> levels
>>
>>>> and their pros/cons and capacity ratio).
>>
>>>
>>
>>> RAID10 offers the best combination of performance and reliability.
>>
>>> RAID0 might beat it *slightly* on performance, but if ANY drive fails on RAID0, the entire volume is lost.
>>
>>>
>>
>>>> If we have ~200-320 shards spread by our 7 Solr node servers (part
>>>> of
>>
>>>> SolrCloud cluster) on single core/collection configured with
>>
>>>> replication factor 2, shouldn't it supply applicative level
>>
>>>> redundancy of indexed data ?
>>
>>>
>>
>>> Yes, you could rely on Solr alone for data redundancy. But if there's a drive failure, do you REALLY want to be single-stranded for the time it takes to rebuild the entire server and copy data? That's what you would end up doing if you choose RAID0.
>>
>>>
>>
>>> It is true that RAID1 or RAID10 means you have to buy double your usable capacity. I would argue that drives are cheap and will cost less than either downtime or sysadmin effort.
>>
>>>
>>
>>> Thanks,
>>
>>> Shawn
>>
>>>
>>
>>>
>>
>>> This electronic message may contain proprietary and confidential information of Verint Systems Inc., its affiliates and/or subsidiaries. The information is intended to be for the use of the individual(s) or entity(ies) named above. If you are not the intended recipient (or authorized to receive this e-mail for the intended recipient), you may not use, copy, disclose or distribute to anyone this message or any information contained in this message. If you have received this electronic message in error, please notify us by replying to this e-mail.
>>
>>
>>
>>
>> This electronic message may contain proprietary and confidential information of Verint Systems Inc., its affiliates and/or subsidiaries. The information is intended to be for the use of the individual(s) or entity(ies) named above. If you are not the intended recipient (or authorized to receive this e-mail for the intended recipient), you may not use, copy, disclose or distribute to anyone this message or any information contained in this message. If you have received this electronic message in error, please notify us by replying to this e-mail.
>
>
>
> This electronic message may contain proprietary and confidential information of Verint Systems Inc., its affiliates and/or subsidiaries. The information is intended to be for the use of the individual(s) or entity(ies) named above. If you are not the intended recipient (or authorized to receive this e-mail for the intended recipient), you may not use, copy, disclose or distribute to anyone this message or any information contained in this message. If you have received this electronic message in error, please notify us by replying to this e-mail.
This electronic message may contain proprietary and confidential information of Verint Systems Inc., its affiliates and/or subsidiaries. The information is intended to be for the use of the individual(s) or entity(ies) named above. If you are not the intended recipient (or authorized to receive this e-mail for the intended recipient), you may not use, copy, disclose or distribute to anyone this message or any information contained in this message. If you have received this electronic message in error, please notify us by replying to this e-mail.
Re: SolrCloud recommended I/O RAID level
Posted by Erick Erickson <er...@gmail.com>.
Shards are really not about maximizing QPS, they’re more about fitting all your data on your machines. Yes, there’s a tradeoff where if you have more shards, each is working on a smaller data set so potentially will work faster, but there’s a corresponding offset of having to do the bookkeeping. By that I mean:
- A request comes into some node, call it the “aggregator”. It’s just one of the Solr nodes.
- A sub-request is sent to one replica of every shard that returns the top N docs from each (i.e. the rows parameter).
- The aggregator sorts all the responses to get the real top N
- For each of the final top N, the aggregator sends a request to the replica it came from to get the return fields.
That’s compounded by the “laggard” problem. In the above, the aggregator can’t know the true top N until it gets a response from a replica from _every_ shard. If even one of those replicas hits a GC pause, network glitch, or just happens to be very, vey busy, it’ll delay the entire response. The more replicas the aggregator has to query, the more likely one of them (and it only takes one) will be slow.
So there’s no easy answer. The first question you must answer is how many docs you can fit on a shard. Find that number. Is the QPS fast enough with that number? If so, that’ll give you the number of shards you need (leave some room for growth of course!). If a shard can host lots and lots of doc but the response time is too slow, shard some more. etc.
Best,
Erick
> On Aug 4, 2019, at 10:53 AM, Kaminski, Adi <Ad...@verint.com> wrote:
>
> Erick - thanks a lot for answering and sharing the below article, it's very helpful !
>
> I have another follow-up question - assuming we have 400 vCPUs across our SolrCloud cluster nodes, will it be better to have 400 shards with replication factor 2
> or 200 shards with replication factor 4 ? What utilizes better the CPUs - shards/or their replicas ?
>
> Thanks,
> Adi
>
> -----Original Message-----
> From: Erick Erickson <er...@gmail.com>
> Sent: Thursday, August 1, 2019 6:48 PM
> To: solr-user@lucene.apache.org
> Subject: Re: SolrCloud recommended I/O RAID level
>
> Yes, I’m exactly talking about the idea of someone pulling the plug on one of your machines. RAID doesn’t help you at all with that.
>
> 10M docs/shard is pretty low actually, depending on the characteristics of docs of course. I’ve ween 10M docs strain things. I’ve seen 300M docs fit in 16G heap. My personal straw-man starting point is 50M docs/shard. You have to test. Old but still valid: https://lucidworks.com/post/sizing-hardware-in-the-abstract-why-we-dont-have-a-definitive-answer/
>
> bq. Will additional replicas will assist with QPS
>
> It Depends (tm). The article I linked above will help you answer the question, the only way I know of to answer definitively is to stress-test on a small (2 nodes will do) cluster. Obviously if you’ve maxed out all your CPUs with no replicas, adding more replicas won’t help. But first you have to convince me that you’re maxing out all your CPUs.
>
> Test, test, test.
>
> Best,
> Erick
>
>> On Aug 1, 2019, at 7:35 AM, Kaminski, Adi <Ad...@verint.com> wrote:
>>
>> Hi Erick,
>>
>> Thanks for the detailed explanation, much appreciated.
>>
>>
>>
>> Regarding #1 reason you have mentioned - but if RAID10 is chosen per
>> Shawn's suggestion, then we should be protected, no ? Unless you mean the whole server is not available and not specific disk of configured RAID10.
>>
>>
>>
>> Regarding #2 reason you have mentioned -
>>
>> · If we’re having 7 Solr node/servers as part of SolrCloud cluster, with 48 vCPU in each server (so in total ~336 vCPU in total)
>>
>> · And we are talking about 2-3.2B of Solr docs distributed on this cluster
>>
>> · And assuming that each shard will handled 10M Solr docs in average (that what we saw is some forums as basic “rule of thumb”)
>>
>> · Questions:
>>
>> o Will additional replicas will assist with QPS ? As we will basically are having ratio of 1 vCPU per 1 shard (the mster replica)
>>
>> o Wouldn’t additional replicas will overclock per total amout of vCPUs, and actually cause delays in queries and lower QPS ?
>>
>>
>>
>> Thanks,
>>
>> Adi
>>
>>
>>
>> -----Original Message-----
>> From: Erick Erickson <er...@gmail.com>
>> Sent: Thursday, August 1, 2019 2:03 PM
>> To: solr-user@lucene.apache.org
>> Subject: Re: SolrCloud recommended I/O RAID level
>>
>>
>>
>> “why would I need a replication factor of 2….”
>>
>>
>>
>> Two reasons:
>>
>>
>>
>> 1> assuming your replicas are located on different physical machines, you can lose one and still keep on running. RAID won’t help if the machine dies.
>>
>>
>>
>> 2> capacity. Depending on your query load, having more than one replica distributes query loads across multiple Solr instances, increasing the QPS you can handle.
>>
>>
>>
>> If you don’t care about either of those issues, then no reason.
>>
>>
>>
>> Best,
>>
>> Erick
>>
>>
>>
>>> On Aug 1, 2019, at 2:11 AM, Kaminski, Adi <Ad...@verint.com>> wrote:
>>
>>>
>>
>>> Hi Shawn,
>>
>>> Thanks for your reply, fully agree with your comments, it clarifies more the need of RAID10 in this case.
>>
>>>
>>
>>> One additional follow-up question - in case we follow this guidelines
>>
>>> and having RAID10 (which leaves us with effective capacity of 50%),
>>
>>> why would I need replication factor of 2 in our SolrCloud core/collection ? Won't it be double protection layer, while the IO layer mirroring of RAID10 actually brings the value, and no need to copy anything when we have IO failures ?
>>
>>>
>>
>>> Thanks,
>>
>>> Adi
>>
>>>
>>
>>> -----Original Message-----
>>
>>> From: Shawn Heisey <ap...@elyograg.org>>
>>
>>> Sent: Tuesday, July 30, 2019 9:44 PM
>>
>>> To: solr-user@lucene.apache.org<ma...@lucene.apache.org>
>>
>>> Subject: Re: SolrCloud recommended I/O RAID level
>>
>>>
>>
>>> On 7/30/2019 12:12 PM, Kaminski, Adi wrote:
>>
>>>> Indeed RAID10 with both mirroring and striping should satisfy the
>>
>>>> need, but per some benchmarks in the network there is still an
>>>> impact
>>
>>>> on write performance on it compared to RAID0 which is considered as
>>
>>>> much better (attaching a table that summarizes different RAID levels
>>
>>>> and their pros/cons and capacity ratio).
>>
>>>
>>
>>> RAID10 offers the best combination of performance and reliability.
>>
>>> RAID0 might beat it *slightly* on performance, but if ANY drive fails on RAID0, the entire volume is lost.
>>
>>>
>>
>>>> If we have ~200-320 shards spread by our 7 Solr node servers (part
>>>> of
>>
>>>> SolrCloud cluster) on single core/collection configured with
>>
>>>> replication factor 2, shouldn't it supply applicative level
>>
>>>> redundancy of indexed data ?
>>
>>>
>>
>>> Yes, you could rely on Solr alone for data redundancy. But if there's a drive failure, do you REALLY want to be single-stranded for the time it takes to rebuild the entire server and copy data? That's what you would end up doing if you choose RAID0.
>>
>>>
>>
>>> It is true that RAID1 or RAID10 means you have to buy double your usable capacity. I would argue that drives are cheap and will cost less than either downtime or sysadmin effort.
>>
>>>
>>
>>> Thanks,
>>
>>> Shawn
>>
>>>
>>
>>>
>>
>>> This electronic message may contain proprietary and confidential information of Verint Systems Inc., its affiliates and/or subsidiaries. The information is intended to be for the use of the individual(s) or entity(ies) named above. If you are not the intended recipient (or authorized to receive this e-mail for the intended recipient), you may not use, copy, disclose or distribute to anyone this message or any information contained in this message. If you have received this electronic message in error, please notify us by replying to this e-mail.
>>
>>
>>
>>
>> This electronic message may contain proprietary and confidential information of Verint Systems Inc., its affiliates and/or subsidiaries. The information is intended to be for the use of the individual(s) or entity(ies) named above. If you are not the intended recipient (or authorized to receive this e-mail for the intended recipient), you may not use, copy, disclose or distribute to anyone this message or any information contained in this message. If you have received this electronic message in error, please notify us by replying to this e-mail.
>
>
>
> This electronic message may contain proprietary and confidential information of Verint Systems Inc., its affiliates and/or subsidiaries. The information is intended to be for the use of the individual(s) or entity(ies) named above. If you are not the intended recipient (or authorized to receive this e-mail for the intended recipient), you may not use, copy, disclose or distribute to anyone this message or any information contained in this message. If you have received this electronic message in error, please notify us by replying to this e-mail.
Re: SolrCloud recommended I/O RAID level
Posted by Shawn Heisey <ap...@elyograg.org>.
On 8/4/2019 8:53 AM, Kaminski, Adi wrote:
> Erick - thanks a lot for answering and sharing the below article, it's very helpful !
>
> I have another follow-up question - assuming we have 400 vCPUs across our SolrCloud cluster nodes, will it be better to have 400 shards with replication factor 2
> or 200 shards with replication factor 4 ? What utilizes better the CPUs - shards/or their replicas ?
Why so many shards? When I think of sharding an index, it's typically
somewhere between 2 and 10 shards. In my opinion, the only use case for
over-sharding is for situations where the index will be growing at a
significant rate. As it grows, you can add servers and spread the
shards out.
I think the best rule of thumb for the number of shards is "the smallest
number that will get the job done. One if possible."
A key piece of information for this decision is the query rate. If
there is a lot of spare CPU capacity and the query rate is REALLY low,
multiple shards on the same machine might increase performance. But as
the query rate increases, performance will decrease when multiple shards
are on each machine, and it will become important for one machine to
handle one shard replica.
Thanks,
Shawn
RE: SolrCloud recommended I/O RAID level
Posted by "Kaminski, Adi" <Ad...@verint.com>.
Erick - thanks a lot for answering and sharing the below article, it's very helpful !
I have another follow-up question - assuming we have 400 vCPUs across our SolrCloud cluster nodes, will it be better to have 400 shards with replication factor 2
or 200 shards with replication factor 4 ? What utilizes better the CPUs - shards/or their replicas ?
Thanks,
Adi
-----Original Message-----
From: Erick Erickson <er...@gmail.com>
Sent: Thursday, August 1, 2019 6:48 PM
To: solr-user@lucene.apache.org
Subject: Re: SolrCloud recommended I/O RAID level
Yes, I’m exactly talking about the idea of someone pulling the plug on one of your machines. RAID doesn’t help you at all with that.
10M docs/shard is pretty low actually, depending on the characteristics of docs of course. I’ve ween 10M docs strain things. I’ve seen 300M docs fit in 16G heap. My personal straw-man starting point is 50M docs/shard. You have to test. Old but still valid: https://lucidworks.com/post/sizing-hardware-in-the-abstract-why-we-dont-have-a-definitive-answer/
bq. Will additional replicas will assist with QPS
It Depends (tm). The article I linked above will help you answer the question, the only way I know of to answer definitively is to stress-test on a small (2 nodes will do) cluster. Obviously if you’ve maxed out all your CPUs with no replicas, adding more replicas won’t help. But first you have to convince me that you’re maxing out all your CPUs.
Test, test, test.
Best,
Erick
> On Aug 1, 2019, at 7:35 AM, Kaminski, Adi <Ad...@verint.com> wrote:
>
> Hi Erick,
>
> Thanks for the detailed explanation, much appreciated.
>
>
>
> Regarding #1 reason you have mentioned - but if RAID10 is chosen per
> Shawn's suggestion, then we should be protected, no ? Unless you mean the whole server is not available and not specific disk of configured RAID10.
>
>
>
> Regarding #2 reason you have mentioned -
>
> · If we’re having 7 Solr node/servers as part of SolrCloud cluster, with 48 vCPU in each server (so in total ~336 vCPU in total)
>
> · And we are talking about 2-3.2B of Solr docs distributed on this cluster
>
> · And assuming that each shard will handled 10M Solr docs in average (that what we saw is some forums as basic “rule of thumb”)
>
> · Questions:
>
> o Will additional replicas will assist with QPS ? As we will basically are having ratio of 1 vCPU per 1 shard (the mster replica)
>
> o Wouldn’t additional replicas will overclock per total amout of vCPUs, and actually cause delays in queries and lower QPS ?
>
>
>
> Thanks,
>
> Adi
>
>
>
> -----Original Message-----
> From: Erick Erickson <er...@gmail.com>
> Sent: Thursday, August 1, 2019 2:03 PM
> To: solr-user@lucene.apache.org
> Subject: Re: SolrCloud recommended I/O RAID level
>
>
>
> “why would I need a replication factor of 2….”
>
>
>
> Two reasons:
>
>
>
> 1> assuming your replicas are located on different physical machines, you can lose one and still keep on running. RAID won’t help if the machine dies.
>
>
>
> 2> capacity. Depending on your query load, having more than one replica distributes query loads across multiple Solr instances, increasing the QPS you can handle.
>
>
>
> If you don’t care about either of those issues, then no reason.
>
>
>
> Best,
>
> Erick
>
>
>
>> On Aug 1, 2019, at 2:11 AM, Kaminski, Adi <Ad...@verint.com>> wrote:
>
>>
>
>> Hi Shawn,
>
>> Thanks for your reply, fully agree with your comments, it clarifies more the need of RAID10 in this case.
>
>>
>
>> One additional follow-up question - in case we follow this guidelines
>
>> and having RAID10 (which leaves us with effective capacity of 50%),
>
>> why would I need replication factor of 2 in our SolrCloud core/collection ? Won't it be double protection layer, while the IO layer mirroring of RAID10 actually brings the value, and no need to copy anything when we have IO failures ?
>
>>
>
>> Thanks,
>
>> Adi
>
>>
>
>> -----Original Message-----
>
>> From: Shawn Heisey <ap...@elyograg.org>>
>
>> Sent: Tuesday, July 30, 2019 9:44 PM
>
>> To: solr-user@lucene.apache.org<ma...@lucene.apache.org>
>
>> Subject: Re: SolrCloud recommended I/O RAID level
>
>>
>
>> On 7/30/2019 12:12 PM, Kaminski, Adi wrote:
>
>>> Indeed RAID10 with both mirroring and striping should satisfy the
>
>>> need, but per some benchmarks in the network there is still an
>>> impact
>
>>> on write performance on it compared to RAID0 which is considered as
>
>>> much better (attaching a table that summarizes different RAID levels
>
>>> and their pros/cons and capacity ratio).
>
>>
>
>> RAID10 offers the best combination of performance and reliability.
>
>> RAID0 might beat it *slightly* on performance, but if ANY drive fails on RAID0, the entire volume is lost.
>
>>
>
>>> If we have ~200-320 shards spread by our 7 Solr node servers (part
>>> of
>
>>> SolrCloud cluster) on single core/collection configured with
>
>>> replication factor 2, shouldn't it supply applicative level
>
>>> redundancy of indexed data ?
>
>>
>
>> Yes, you could rely on Solr alone for data redundancy. But if there's a drive failure, do you REALLY want to be single-stranded for the time it takes to rebuild the entire server and copy data? That's what you would end up doing if you choose RAID0.
>
>>
>
>> It is true that RAID1 or RAID10 means you have to buy double your usable capacity. I would argue that drives are cheap and will cost less than either downtime or sysadmin effort.
>
>>
>
>> Thanks,
>
>> Shawn
>
>>
>
>>
>
>> This electronic message may contain proprietary and confidential information of Verint Systems Inc., its affiliates and/or subsidiaries. The information is intended to be for the use of the individual(s) or entity(ies) named above. If you are not the intended recipient (or authorized to receive this e-mail for the intended recipient), you may not use, copy, disclose or distribute to anyone this message or any information contained in this message. If you have received this electronic message in error, please notify us by replying to this e-mail.
>
>
>
>
> This electronic message may contain proprietary and confidential information of Verint Systems Inc., its affiliates and/or subsidiaries. The information is intended to be for the use of the individual(s) or entity(ies) named above. If you are not the intended recipient (or authorized to receive this e-mail for the intended recipient), you may not use, copy, disclose or distribute to anyone this message or any information contained in this message. If you have received this electronic message in error, please notify us by replying to this e-mail.
This electronic message may contain proprietary and confidential information of Verint Systems Inc., its affiliates and/or subsidiaries. The information is intended to be for the use of the individual(s) or entity(ies) named above. If you are not the intended recipient (or authorized to receive this e-mail for the intended recipient), you may not use, copy, disclose or distribute to anyone this message or any information contained in this message. If you have received this electronic message in error, please notify us by replying to this e-mail.
Re: SolrCloud recommended I/O RAID level
Posted by Erick Erickson <er...@gmail.com>.
Yes, I’m exactly talking about the idea of someone pulling the plug on one of your machines. RAID doesn’t help you at all with that.
10M docs/shard is pretty low actually, depending on the characteristics of docs of course. I’ve ween 10M docs strain things. I’ve seen 300M docs fit in 16G heap. My personal straw-man starting point is 50M docs/shard. You have to test. Old but still valid: https://lucidworks.com/post/sizing-hardware-in-the-abstract-why-we-dont-have-a-definitive-answer/
bq. Will additional replicas will assist with QPS
It Depends (tm). The article I linked above will help you answer the question, the only way I know of to answer definitively is to stress-test on a small (2 nodes will do) cluster. Obviously if you’ve maxed out all your CPUs with no replicas, adding more replicas won’t help. But first you have to convince me that you’re maxing out all your CPUs.
Test, test, test.
Best,
Erick
> On Aug 1, 2019, at 7:35 AM, Kaminski, Adi <Ad...@verint.com> wrote:
>
> Hi Erick,
>
> Thanks for the detailed explanation, much appreciated.
>
>
>
> Regarding #1 reason you have mentioned - but if RAID10 is chosen per Shawn's suggestion, then we should be protected, no ? Unless you mean the whole server is not available
> and not specific disk of configured RAID10.
>
>
>
> Regarding #2 reason you have mentioned -
>
> · If we’re having 7 Solr node/servers as part of SolrCloud cluster, with 48 vCPU in each server (so in total ~336 vCPU in total)
>
> · And we are talking about 2-3.2B of Solr docs distributed on this cluster
>
> · And assuming that each shard will handled 10M Solr docs in average (that what we saw is some forums as basic “rule of thumb”)
>
> · Questions:
>
> o Will additional replicas will assist with QPS ? As we will basically are having ratio of 1 vCPU per 1 shard (the mster replica)
>
> o Wouldn’t additional replicas will overclock per total amout of vCPUs, and actually cause delays in queries and lower QPS ?
>
>
>
> Thanks,
>
> Adi
>
>
>
> -----Original Message-----
> From: Erick Erickson <er...@gmail.com>
> Sent: Thursday, August 1, 2019 2:03 PM
> To: solr-user@lucene.apache.org
> Subject: Re: SolrCloud recommended I/O RAID level
>
>
>
> “why would I need a replication factor of 2….”
>
>
>
> Two reasons:
>
>
>
> 1> assuming your replicas are located on different physical machines, you can lose one and still keep on running. RAID won’t help if the machine dies.
>
>
>
> 2> capacity. Depending on your query load, having more than one replica distributes query loads across multiple Solr instances, increasing the QPS you can handle.
>
>
>
> If you don’t care about either of those issues, then no reason.
>
>
>
> Best,
>
> Erick
>
>
>
>> On Aug 1, 2019, at 2:11 AM, Kaminski, Adi <Ad...@verint.com>> wrote:
>
>>
>
>> Hi Shawn,
>
>> Thanks for your reply, fully agree with your comments, it clarifies more the need of RAID10 in this case.
>
>>
>
>> One additional follow-up question - in case we follow this guidelines
>
>> and having RAID10 (which leaves us with effective capacity of 50%),
>
>> why would I need replication factor of 2 in our SolrCloud core/collection ? Won't it be double protection layer, while the IO layer mirroring of RAID10 actually brings the value, and no need to copy anything when we have IO failures ?
>
>>
>
>> Thanks,
>
>> Adi
>
>>
>
>> -----Original Message-----
>
>> From: Shawn Heisey <ap...@elyograg.org>>
>
>> Sent: Tuesday, July 30, 2019 9:44 PM
>
>> To: solr-user@lucene.apache.org<ma...@lucene.apache.org>
>
>> Subject: Re: SolrCloud recommended I/O RAID level
>
>>
>
>> On 7/30/2019 12:12 PM, Kaminski, Adi wrote:
>
>>> Indeed RAID10 with both mirroring and striping should satisfy the
>
>>> need, but per some benchmarks in the network there is still an impact
>
>>> on write performance on it compared to RAID0 which is considered as
>
>>> much better (attaching a table that summarizes different RAID levels
>
>>> and their pros/cons and capacity ratio).
>
>>
>
>> RAID10 offers the best combination of performance and reliability.
>
>> RAID0 might beat it *slightly* on performance, but if ANY drive fails on RAID0, the entire volume is lost.
>
>>
>
>>> If we have ~200-320 shards spread by our 7 Solr node servers (part of
>
>>> SolrCloud cluster) on single core/collection configured with
>
>>> replication factor 2, shouldn't it supply applicative level
>
>>> redundancy of indexed data ?
>
>>
>
>> Yes, you could rely on Solr alone for data redundancy. But if there's a drive failure, do you REALLY want to be single-stranded for the time it takes to rebuild the entire server and copy data? That's what you would end up doing if you choose RAID0.
>
>>
>
>> It is true that RAID1 or RAID10 means you have to buy double your usable capacity. I would argue that drives are cheap and will cost less than either downtime or sysadmin effort.
>
>>
>
>> Thanks,
>
>> Shawn
>
>>
>
>>
>
>> This electronic message may contain proprietary and confidential information of Verint Systems Inc., its affiliates and/or subsidiaries. The information is intended to be for the use of the individual(s) or entity(ies) named above. If you are not the intended recipient (or authorized to receive this e-mail for the intended recipient), you may not use, copy, disclose or distribute to anyone this message or any information contained in this message. If you have received this electronic message in error, please notify us by replying to this e-mail.
>
>
>
>
> This electronic message may contain proprietary and confidential information of Verint Systems Inc., its affiliates and/or subsidiaries. The information is intended to be for the use of the individual(s) or entity(ies) named above. If you are not the intended recipient (or authorized to receive this e-mail for the intended recipient), you may not use, copy, disclose or distribute to anyone this message or any information contained in this message. If you have received this electronic message in error, please notify us by replying to this e-mail.
RE: SolrCloud recommended I/O RAID level
Posted by "Kaminski, Adi" <Ad...@verint.com>.
Hi Erick,
Thanks for the detailed explanation, much appreciated.
Regarding #1 reason you have mentioned - but if RAID10 is chosen per Shawn's suggestion, then we should be protected, no ? Unless you mean the whole server is not available
and not specific disk of configured RAID10.
Regarding #2 reason you have mentioned -
· If we’re having 7 Solr node/servers as part of SolrCloud cluster, with 48 vCPU in each server (so in total ~336 vCPU in total)
· And we are talking about 2-3.2B of Solr docs distributed on this cluster
· And assuming that each shard will handled 10M Solr docs in average (that what we saw is some forums as basic “rule of thumb”)
· Questions:
o Will additional replicas will assist with QPS ? As we will basically are having ratio of 1 vCPU per 1 shard (the mster replica)
o Wouldn’t additional replicas will overclock per total amout of vCPUs, and actually cause delays in queries and lower QPS ?
Thanks,
Adi
-----Original Message-----
From: Erick Erickson <er...@gmail.com>
Sent: Thursday, August 1, 2019 2:03 PM
To: solr-user@lucene.apache.org
Subject: Re: SolrCloud recommended I/O RAID level
“why would I need a replication factor of 2….”
Two reasons:
1> assuming your replicas are located on different physical machines, you can lose one and still keep on running. RAID won’t help if the machine dies.
2> capacity. Depending on your query load, having more than one replica distributes query loads across multiple Solr instances, increasing the QPS you can handle.
If you don’t care about either of those issues, then no reason.
Best,
Erick
> On Aug 1, 2019, at 2:11 AM, Kaminski, Adi <Ad...@verint.com>> wrote:
>
> Hi Shawn,
> Thanks for your reply, fully agree with your comments, it clarifies more the need of RAID10 in this case.
>
> One additional follow-up question - in case we follow this guidelines
> and having RAID10 (which leaves us with effective capacity of 50%),
> why would I need replication factor of 2 in our SolrCloud core/collection ? Won't it be double protection layer, while the IO layer mirroring of RAID10 actually brings the value, and no need to copy anything when we have IO failures ?
>
> Thanks,
> Adi
>
> -----Original Message-----
> From: Shawn Heisey <ap...@elyograg.org>>
> Sent: Tuesday, July 30, 2019 9:44 PM
> To: solr-user@lucene.apache.org<ma...@lucene.apache.org>
> Subject: Re: SolrCloud recommended I/O RAID level
>
> On 7/30/2019 12:12 PM, Kaminski, Adi wrote:
>> Indeed RAID10 with both mirroring and striping should satisfy the
>> need, but per some benchmarks in the network there is still an impact
>> on write performance on it compared to RAID0 which is considered as
>> much better (attaching a table that summarizes different RAID levels
>> and their pros/cons and capacity ratio).
>
> RAID10 offers the best combination of performance and reliability.
> RAID0 might beat it *slightly* on performance, but if ANY drive fails on RAID0, the entire volume is lost.
>
>> If we have ~200-320 shards spread by our 7 Solr node servers (part of
>> SolrCloud cluster) on single core/collection configured with
>> replication factor 2, shouldn't it supply applicative level
>> redundancy of indexed data ?
>
> Yes, you could rely on Solr alone for data redundancy. But if there's a drive failure, do you REALLY want to be single-stranded for the time it takes to rebuild the entire server and copy data? That's what you would end up doing if you choose RAID0.
>
> It is true that RAID1 or RAID10 means you have to buy double your usable capacity. I would argue that drives are cheap and will cost less than either downtime or sysadmin effort.
>
> Thanks,
> Shawn
>
>
> This electronic message may contain proprietary and confidential information of Verint Systems Inc., its affiliates and/or subsidiaries. The information is intended to be for the use of the individual(s) or entity(ies) named above. If you are not the intended recipient (or authorized to receive this e-mail for the intended recipient), you may not use, copy, disclose or distribute to anyone this message or any information contained in this message. If you have received this electronic message in error, please notify us by replying to this e-mail.
This electronic message may contain proprietary and confidential information of Verint Systems Inc., its affiliates and/or subsidiaries. The information is intended to be for the use of the individual(s) or entity(ies) named above. If you are not the intended recipient (or authorized to receive this e-mail for the intended recipient), you may not use, copy, disclose or distribute to anyone this message or any information contained in this message. If you have received this electronic message in error, please notify us by replying to this e-mail.
Re: SolrCloud recommended I/O RAID level
Posted by Erick Erickson <er...@gmail.com>.
“why would I need a replication factor of 2….”
Two reasons:
1> assuming your replicas are located on different physical machines, you can lose one and still keep on running. RAID won’t help if the machine dies.
2> capacity. Depending on your query load, having more than one replica distributes query loads across multiple Solr instances, increasing the QPS you can handle.
If you don’t care about either of those issues, then no reason.
Best,
Erick
> On Aug 1, 2019, at 2:11 AM, Kaminski, Adi <Ad...@verint.com> wrote:
>
> Hi Shawn,
> Thanks for your reply, fully agree with your comments, it clarifies more the need of RAID10 in this case.
>
> One additional follow-up question - in case we follow this guidelines and having RAID10 (which leaves us with effective capacity of 50%), why would I need
> replication factor of 2 in our SolrCloud core/collection ? Won't it be double protection layer, while the IO layer mirroring of RAID10 actually brings the value, and no need to copy anything
> when we have IO failures ?
>
> Thanks,
> Adi
>
> -----Original Message-----
> From: Shawn Heisey <ap...@elyograg.org>
> Sent: Tuesday, July 30, 2019 9:44 PM
> To: solr-user@lucene.apache.org
> Subject: Re: SolrCloud recommended I/O RAID level
>
> On 7/30/2019 12:12 PM, Kaminski, Adi wrote:
>> Indeed RAID10 with both mirroring and striping should satisfy the
>> need, but per some benchmarks in the network there is still an impact
>> on write performance on it compared to RAID0 which is considered as
>> much better (attaching a table that summarizes different RAID levels
>> and their pros/cons and capacity ratio).
>
> RAID10 offers the best combination of performance and reliability.
> RAID0 might beat it *slightly* on performance, but if ANY drive fails on RAID0, the entire volume is lost.
>
>> If we have ~200-320 shards spread by our 7 Solr node servers (part of
>> SolrCloud cluster) on single core/collection configured with
>> replication factor 2, shouldn't it supply applicative level redundancy
>> of indexed data ?
>
> Yes, you could rely on Solr alone for data redundancy. But if there's a drive failure, do you REALLY want to be single-stranded for the time it takes to rebuild the entire server and copy data? That's what you would end up doing if you choose RAID0.
>
> It is true that RAID1 or RAID10 means you have to buy double your usable capacity. I would argue that drives are cheap and will cost less than either downtime or sysadmin effort.
>
> Thanks,
> Shawn
>
>
> This electronic message may contain proprietary and confidential information of Verint Systems Inc., its affiliates and/or subsidiaries. The information is intended to be for the use of the individual(s) or entity(ies) named above. If you are not the intended recipient (or authorized to receive this e-mail for the intended recipient), you may not use, copy, disclose or distribute to anyone this message or any information contained in this message. If you have received this electronic message in error, please notify us by replying to this e-mail.