Good points Bryan, some more color: Regular EBS is *not* okay for C*. But AWS has some nicer EBS now that has performed okay recently:
http://docs.aws.amazon.com/AWSEC2/latest/UserGuide/EBSVolumeTypes.html https://www.youtube.com/watch?v=1R-mgOcOSd4 The cloud vendors are moving toward shared storage and we can't ignore that in the long term (they will push us in that direction financially). Fortunately their shared storage offerings are also getting better. For example google's elastic storage offerring provides very reliable latencies <https://www.youtube.com/watch?v=qf-7IhCqCcI>which is what we care the most about, not iops. On the practical side, a key thing I've noticed with real deployments is that the size of the volume affects how fast it will perform and how stable it's latencies will be so make sure to get large EBS volumes > 1tb to get decent performance, even if your nodes aren't that dense. All the best, [image: datastax_logo.png] <http://www.datastax.com/> Sebastián Estévez Solutions Architect | 954 905 8615 | sebastian.este...@datastax.com [image: linkedin.png] <https://www.linkedin.com/company/datastax> [image: facebook.png] <https://www.facebook.com/datastax> [image: twitter.png] <https://twitter.com/datastax> [image: g+.png] <https://plus.google.com/+Datastax/about> <http://feeds.feedburner.com/datastax> <http://goog_410786983> <http://www.datastax.com/gartner-magic-quadrant-odbms> DataStax is the fastest, most scalable distributed database technology, delivering Apache Cassandra to the world’s most innovative enterprises. Datastax is built to be agile, always-on, and predictably scalable to any size. With more than 500 customers in 45 countries, DataStax is the database technology and transactional backbone of choice for the worlds most innovative companies such as Netflix, Adobe, Intuit, and eBay. On Wed, Feb 3, 2016 at 7:23 PM, Bryan Cheng <br...@blockcypher.com> wrote: > From my experience, EBS has transitioned from "stay the hell away" to "OK" > as the new GP2 SSD type has come out and stabilized over the last few > years, especially with the addition of EBS-optimized instances that have > dedicated EBS bandwidth. The latter has really helped to stabilize the > problematic 99.9-percentile latency spikes that use to plague EBS volumes. > > EBS (IMHO) has always had operational advantages, but inconsistent latency > and generally poor performance in the past lead many to disregard it. > > On Wed, Feb 3, 2016 at 4:09 PM, James Rothering <jrother...@codojo.me> > wrote: > >> Just curious here ... when did EBS become OK for C*? Didn't they always >> push towards using ephemeral disks? >> >> On Wed, Feb 3, 2016 at 12:17 PM, Ben Bromhead <b...@instaclustr.com> >> wrote: >> >>> For what it's worth we've tried d2 instances and they encourage terrible >>> things like super dense nodes (increases your replacement time). In terms >>> of useable storage I would go with gp2 EBS on a m4 based instance. >>> >>> On Mon, 1 Feb 2016 at 14:25 Jack Krupansky <jack.krupan...@gmail.com> >>> wrote: >>> >>>> Ah, yes, the good old days of m1.large. >>>> >>>> -- Jack Krupansky >>>> >>>> On Mon, Feb 1, 2016 at 5:12 PM, Jeff Jirsa <jeff.ji...@crowdstrike.com> >>>> wrote: >>>> >>>>> A lot of people use the old gen instances (m1 in particular) because >>>>> they came with a ton of effectively free ephemeral storage (up to 1.6TB). >>>>> Whether or not they’re viable is a decision for each user to make. They’re >>>>> very, very commonly used for C*, though. At a time when EBS was not >>>>> sufficiently robust or reliable, a cluster of m1 instances was the de >>>>> facto >>>>> standard. >>>>> >>>>> The canonical “best practice” in 2015 was i2. We believe we’ve made a >>>>> compelling argument to use m4 or c4 instead of i2. There exists a company >>>>> we know currently testing d2 at scale, though I’m not sure they have much >>>>> in terms of concrete results at this time. >>>>> >>>>> - Jeff >>>>> >>>>> From: Jack Krupansky >>>>> Reply-To: "user@cassandra.apache.org" >>>>> Date: Monday, February 1, 2016 at 1:55 PM >>>>> >>>>> To: "user@cassandra.apache.org" >>>>> Subject: Re: EC2 storage options for C* >>>>> >>>>> Thanks. My typo - I referenced "C2 Dense Storage" which is really "D2 >>>>> Dense Storage". >>>>> >>>>> The remaining question is whether any of the "Previous Generation >>>>> Instances" should be publicly recommended going forward. >>>>> >>>>> And whether non-SSD instances should be recommended going forward as >>>>> well. sure, technically, someone could use the legacy instances, but the >>>>> question is what we should be recommending as best practice going forward. >>>>> >>>>> Yeah, the i2 instances look like the sweet spot for any non-EBS >>>>> clusters. >>>>> >>>>> -- Jack Krupansky >>>>> >>>>> On Mon, Feb 1, 2016 at 4:30 PM, Steve Robenalt <sroben...@highwire.org >>>>> > wrote: >>>>> >>>>>> Hi Jack, >>>>>> >>>>>> At the bottom of the instance-types page, there is a link to the >>>>>> previous generations, which includes the older series (m1, m2, etc), many >>>>>> of which have HDD options. >>>>>> >>>>>> There are also the d2 (Dense Storage) instances in the current >>>>>> generation that include various combos of local HDDs. >>>>>> >>>>>> The i2 series has good sized SSDs available, and has the advanced >>>>>> networking option, which is also useful for Cassandra. The enhanced >>>>>> networking is available with other instance types as well, as you'll see >>>>>> on >>>>>> the feature list under each type. >>>>>> >>>>>> Steve >>>>>> >>>>>> >>>>>> >>>>>> On Mon, Feb 1, 2016 at 1:17 PM, Jack Krupansky < >>>>>> jack.krupan...@gmail.com> wrote: >>>>>> >>>>>>> Thanks. Reading a little bit on AWS, and back to my SSD vs. magnetic >>>>>>> question, it seems like magnetic (HDD) is no longer a recommended >>>>>>> storage >>>>>>> option for databases on AWS. In particular, only the C2 Dense Storage >>>>>>> instances have local magnetic storage - all the other instance types are >>>>>>> SSD or EBS-only - and EBS Magnetic is only recommended for "Infrequent >>>>>>> Data >>>>>>> Access." >>>>>>> >>>>>>> For the record, that AWS doc has Cassandra listed as a use case for >>>>>>> i2 instance types. >>>>>>> >>>>>>> Also, the AWS doc lists EBS io2 for the NoSQL database use case and >>>>>>> gp2 only for the "small to medium databases" use case. >>>>>>> >>>>>>> Do older instances with local HDD still exist on AWS (m1, m2, etc.)? >>>>>>> Is the doc simply for any newly started instances? >>>>>>> >>>>>>> See: >>>>>>> https://aws.amazon.com/ec2/instance-types/ >>>>>>> http://aws.amazon.com/ebs/details/ >>>>>>> >>>>>>> >>>>>>> -- Jack Krupansky >>>>>>> >>>>>>> On Mon, Feb 1, 2016 at 2:09 PM, Jeff Jirsa < >>>>>>> jeff.ji...@crowdstrike.com> wrote: >>>>>>> >>>>>>>> > My apologies if my questions are actually answered on the video >>>>>>>> or slides, I just did a quick scan of the slide text. >>>>>>>> >>>>>>>> Virtually all of them are covered. >>>>>>>> >>>>>>>> > I'm curious where the EBS physical devices actually reside - are >>>>>>>> they in the same rack, the same data center, same availability zone? I >>>>>>>> mean, people try to minimize network latency between nodes, so how >>>>>>>> exactly >>>>>>>> is EBS able to avoid network latency? >>>>>>>> >>>>>>>> Not published,and probably not a straight forward answer (probably >>>>>>>> have redundancy cross-az, if it matches some of their other published >>>>>>>> behaviors). The promise they give you is ‘iops’, with a certain block >>>>>>>> size. >>>>>>>> Some instance types are optimized for dedicated, ebs-only network >>>>>>>> interfaces. Like most things in cassandra / cloud, the only way to >>>>>>>> know for >>>>>>>> sure is to test it yourself and see if observed latency is acceptable >>>>>>>> (or >>>>>>>> trust our testing, if you assume we’re sufficiently smart and honest). >>>>>>>> >>>>>>>> > Did your test use Amazon EBS–Optimized Instances? >>>>>>>> >>>>>>>> We tested dozens of instance type/size combinations (literally). >>>>>>>> The best performance was clearly with ebs-optimized instances that also >>>>>>>> have enhanced networking (c4, m4, etc) - slide 43 >>>>>>>> >>>>>>>> > SSD or magnetic or does it make any difference? >>>>>>>> >>>>>>>> SSD, GP2 (slide 64) >>>>>>>> >>>>>>>> > What info is available on EBS performance at peak times, when >>>>>>>> multiple AWS customers have spikes of demand? >>>>>>>> >>>>>>>> Not published, but experiments show that we can hit 10k iops all >>>>>>>> day every day with only trivial noisy neighbor problems, not enough to >>>>>>>> impact a real cluster (slide 58) >>>>>>>> >>>>>>>> > Is RAID much of a factor or help at all using EBS? >>>>>>>> >>>>>>>> You can use RAID to get higher IOPS than you’d normally get by >>>>>>>> default (GP2 IOPS cap is 10k, which you get with a 3.333T volume – if >>>>>>>> you >>>>>>>> need more than 10k, you can stripe volumes together up to the ebs >>>>>>>> network >>>>>>>> link max) (hinted at in slide 64) >>>>>>>> >>>>>>>> > How exactly is EBS provisioned in terms of its own HA - I mean, >>>>>>>> with a properly configured Cassandra cluster RF provides HA, so what >>>>>>>> is the >>>>>>>> equivalent for EBS? If I have RF=3, what assurance is there that those >>>>>>>> three EBS volumes aren't all in the same physical rack? >>>>>>>> >>>>>>>> There is HA, I’m not sure that AWS publishes specifics. >>>>>>>> Occasionally specific volumes will have issues (hypervisor’s dedicated >>>>>>>> ethernet link to EBS network fails, for example). Occasionally >>>>>>>> instances >>>>>>>> will have issues. The volume-specific issues seem to be less common >>>>>>>> than >>>>>>>> the instance-store “instance retired” or “instance is running on >>>>>>>> degraded >>>>>>>> hardware” events. Stop/Start and you’ve recovered (possible with EBS, >>>>>>>> not >>>>>>>> possible with instance store). The assurances are in AWS’ SLA – if the >>>>>>>> SLA >>>>>>>> is insufficient (and it probably is insufficient), use more than one AZ >>>>>>>> and/or AWS region or cloud vendor. >>>>>>>> >>>>>>>> > For multi-data center operation, what configuration options >>>>>>>> assure that the EBS volumes for each DC are truly physically separated? >>>>>>>> >>>>>>>> It used to be true that EBS control plane for a given region >>>>>>>> spanned AZs. That’s no longer true. AWS asserts that failure modes for >>>>>>>> each >>>>>>>> AZ are isolated (data may replicate between AZs, but a full outage in >>>>>>>> us-east-1a shouldn’t affect running ebs volumes in us-east-1b or >>>>>>>> us-east-1c). Slide 65 >>>>>>>> >>>>>>>> > In terms of syncing data for the commit log, if the OS call to >>>>>>>> sync an EBS volume returns, is the commit log data absolutely 100% >>>>>>>> synced >>>>>>>> at the hardware level on the EBS end, such that a power failure of the >>>>>>>> systems on which the EBS volumes reside will still guarantee >>>>>>>> availability >>>>>>>> of the fsynced data. As well, is return from fsync an absolute >>>>>>>> guarantee of >>>>>>>> sstable durability when Cassandra is about to delete the commit log, >>>>>>>> including when the two are on different volumes? In practice, we would >>>>>>>> like >>>>>>>> some significant degree of pipelining of data, such as during the full >>>>>>>> processing of flushing memtables, but for the fsync at the end a solid >>>>>>>> guarantee is needed. >>>>>>>> >>>>>>>> Most of the answers in this block are “probably not 100%, you >>>>>>>> should be writing to more than one host/AZ/DC/vendor to protect your >>>>>>>> organization from failures”. AWS targets something like 0.1% annual >>>>>>>> failure >>>>>>>> rate per volume and 99.999% availability (slide 66). We believe they’re >>>>>>>> exceeding those goals (at least based with the petabytes of data we >>>>>>>> have on >>>>>>>> gp2 volumes). >>>>>>>> >>>>>>>> >>>>>>>> >>>>>>>> From: Jack Krupansky >>>>>>>> Reply-To: "user@cassandra.apache.org" >>>>>>>> Date: Monday, February 1, 2016 at 5:51 AM >>>>>>>> >>>>>>>> To: "user@cassandra.apache.org" >>>>>>>> Subject: Re: EC2 storage options for C* >>>>>>>> >>>>>>>> I'm not a fan of guy - this appears to be the slideshare >>>>>>>> corresponding to the video: >>>>>>>> >>>>>>>> http://www.slideshare.net/AmazonWebServices/bdt323-amazon-ebs-cassandra-1-million-writes-per-second >>>>>>>> >>>>>>>> My apologies if my questions are actually answered on the video or >>>>>>>> slides, I just did a quick scan of the slide text. >>>>>>>> >>>>>>>> I'm curious where the EBS physical devices actually reside - are >>>>>>>> they in the same rack, the same data center, same availability zone? I >>>>>>>> mean, people try to minimize network latency between nodes, so how >>>>>>>> exactly >>>>>>>> is EBS able to avoid network latency? >>>>>>>> >>>>>>>> Did your test use Amazon EBS–Optimized Instances? >>>>>>>> >>>>>>>> SSD or magnetic or does it make any difference? >>>>>>>> >>>>>>>> What info is available on EBS performance at peak times, when >>>>>>>> multiple AWS customers have spikes of demand? >>>>>>>> >>>>>>>> Is RAID much of a factor or help at all using EBS? >>>>>>>> >>>>>>>> How exactly is EBS provisioned in terms of its own HA - I mean, >>>>>>>> with a properly configured Cassandra cluster RF provides HA, so what >>>>>>>> is the >>>>>>>> equivalent for EBS? If I have RF=3, what assurance is there that those >>>>>>>> three EBS volumes aren't all in the same physical rack? >>>>>>>> >>>>>>>> For multi-data center operation, what configuration options assure >>>>>>>> that the EBS volumes for each DC are truly physically separated? >>>>>>>> >>>>>>>> In terms of syncing data for the commit log, if the OS call to sync >>>>>>>> an EBS volume returns, is the commit log data absolutely 100% synced >>>>>>>> at the >>>>>>>> hardware level on the EBS end, such that a power failure of the >>>>>>>> systems on >>>>>>>> which the EBS volumes reside will still guarantee availability of the >>>>>>>> fsynced data. As well, is return from fsync an absolute guarantee of >>>>>>>> sstable durability when Cassandra is about to delete the commit log, >>>>>>>> including when the two are on different volumes? In practice, we would >>>>>>>> like >>>>>>>> some significant degree of pipelining of data, such as during the full >>>>>>>> processing of flushing memtables, but for the fsync at the end a solid >>>>>>>> guarantee is needed. >>>>>>>> >>>>>>>> >>>>>>>> -- Jack Krupansky >>>>>>>> >>>>>>>> On Mon, Feb 1, 2016 at 12:56 AM, Eric Plowe <eric.pl...@gmail.com> >>>>>>>> wrote: >>>>>>>> >>>>>>>>> Jeff, >>>>>>>>> >>>>>>>>> If EBS goes down, then EBS Gp2 will go down as well, no? I'm not >>>>>>>>> discounting EBS, but prior outages are worrisome. >>>>>>>>> >>>>>>>>> >>>>>>>>> On Sunday, January 31, 2016, Jeff Jirsa < >>>>>>>>> jeff.ji...@crowdstrike.com> wrote: >>>>>>>>> >>>>>>>>>> Free to choose what you'd like, but EBS outages were also >>>>>>>>>> addressed in that video (second half, discussion by Dennis Opacki). >>>>>>>>>> 2016 >>>>>>>>>> EBS isn't the same as 2011 EBS. >>>>>>>>>> >>>>>>>>>> -- >>>>>>>>>> Jeff Jirsa >>>>>>>>>> >>>>>>>>>> >>>>>>>>>> On Jan 31, 2016, at 8:27 PM, Eric Plowe <eric.pl...@gmail.com> >>>>>>>>>> wrote: >>>>>>>>>> >>>>>>>>>> Thank you all for the suggestions. I'm torn between GP2 vs >>>>>>>>>> Ephemeral. GP2 after testing is a viable contender for our workload. >>>>>>>>>> The >>>>>>>>>> only worry I have is EBS outages, which have happened. >>>>>>>>>> >>>>>>>>>> On Sunday, January 31, 2016, Jeff Jirsa < >>>>>>>>>> jeff.ji...@crowdstrike.com> wrote: >>>>>>>>>> >>>>>>>>>>> Also in that video - it's long but worth watching >>>>>>>>>>> >>>>>>>>>>> We tested up to 1M reads/second as well, blowing out page cache >>>>>>>>>>> to ensure we weren't "just" reading from memory >>>>>>>>>>> >>>>>>>>>>> >>>>>>>>>>> >>>>>>>>>>> -- >>>>>>>>>>> Jeff Jirsa >>>>>>>>>>> >>>>>>>>>>> >>>>>>>>>>> On Jan 31, 2016, at 9:52 AM, Jack Krupansky < >>>>>>>>>>> jack.krupan...@gmail.com> wrote: >>>>>>>>>>> >>>>>>>>>>> How about reads? Any differences between read-intensive and >>>>>>>>>>> write-intensive workloads? >>>>>>>>>>> >>>>>>>>>>> -- Jack Krupansky >>>>>>>>>>> >>>>>>>>>>> On Sun, Jan 31, 2016 at 3:13 AM, Jeff Jirsa < >>>>>>>>>>> jeff.ji...@crowdstrike.com> wrote: >>>>>>>>>>> >>>>>>>>>>>> Hi John, >>>>>>>>>>>> >>>>>>>>>>>> We run using 4T GP2 volumes, which guarantee 10k iops. Even at >>>>>>>>>>>> 1M writes per second on 60 nodes, we didn’t come close to hitting >>>>>>>>>>>> even 50% >>>>>>>>>>>> utilization (10k is more than enough for most workloads). PIOPS is >>>>>>>>>>>> not >>>>>>>>>>>> necessary. >>>>>>>>>>>> >>>>>>>>>>>> >>>>>>>>>>>> >>>>>>>>>>>> From: John Wong >>>>>>>>>>>> Reply-To: "user@cassandra.apache.org" >>>>>>>>>>>> Date: Saturday, January 30, 2016 at 3:07 PM >>>>>>>>>>>> To: "user@cassandra.apache.org" >>>>>>>>>>>> Subject: Re: EC2 storage options for C* >>>>>>>>>>>> >>>>>>>>>>>> For production I'd stick with ephemeral disks (aka instance >>>>>>>>>>>> storage) if you have running a lot of transaction. >>>>>>>>>>>> However, for regular small testing/qa cluster, or something you >>>>>>>>>>>> know you want to reload often, EBS is definitely good enough and >>>>>>>>>>>> we haven't >>>>>>>>>>>> had issues 99%. The 1% is kind of anomaly where we have flush >>>>>>>>>>>> blocked. >>>>>>>>>>>> >>>>>>>>>>>> But Jeff, kudo that you are able to use EBS. I didn't go >>>>>>>>>>>> through the video, do you actually use PIOPS or just standard GP2 >>>>>>>>>>>> in your >>>>>>>>>>>> production cluster? >>>>>>>>>>>> >>>>>>>>>>>> On Sat, Jan 30, 2016 at 1:28 PM, Bryan Cheng < >>>>>>>>>>>> br...@blockcypher.com> wrote: >>>>>>>>>>>> >>>>>>>>>>>>> Yep, that motivated my question "Do you have any idea what >>>>>>>>>>>>> kind of disk performance you need?". If you need the performance, >>>>>>>>>>>>> its hard >>>>>>>>>>>>> to beat ephemeral SSD in RAID 0 on EC2, and its a solid, battle >>>>>>>>>>>>> tested >>>>>>>>>>>>> configuration. If you don't, though, EBS GP2 will save a _lot_ of >>>>>>>>>>>>> headache. >>>>>>>>>>>>> >>>>>>>>>>>>> Personally, on small clusters like ours (12 nodes), we've >>>>>>>>>>>>> found our choice of instance dictated much more by the balance of >>>>>>>>>>>>> price, >>>>>>>>>>>>> CPU, and memory. We're using GP2 SSD and we find that for our >>>>>>>>>>>>> patterns the >>>>>>>>>>>>> disk is rarely the bottleneck. YMMV, of course. >>>>>>>>>>>>> >>>>>>>>>>>>> On Fri, Jan 29, 2016 at 7:32 PM, Jeff Jirsa < >>>>>>>>>>>>> jeff.ji...@crowdstrike.com> wrote: >>>>>>>>>>>>> >>>>>>>>>>>>>> If you have to ask that question, I strongly recommend m4 or >>>>>>>>>>>>>> c4 instances with GP2 EBS. When you don’t care about replacing >>>>>>>>>>>>>> a node >>>>>>>>>>>>>> because of an instance failure, go with i2+ephemerals. Until >>>>>>>>>>>>>> then, GP2 EBS >>>>>>>>>>>>>> is capable of amazing things, and greatly simplifies life. >>>>>>>>>>>>>> >>>>>>>>>>>>>> We gave a talk on this topic at both Cassandra Summit and AWS >>>>>>>>>>>>>> re:Invent: https://www.youtube.com/watch?v=1R-mgOcOSd4 It’s >>>>>>>>>>>>>> very much a viable option, despite any old documents online that >>>>>>>>>>>>>> say >>>>>>>>>>>>>> otherwise. >>>>>>>>>>>>>> >>>>>>>>>>>>>> >>>>>>>>>>>>>> >>>>>>>>>>>>>> From: Eric Plowe >>>>>>>>>>>>>> Reply-To: "user@cassandra.apache.org" >>>>>>>>>>>>>> Date: Friday, January 29, 2016 at 4:33 PM >>>>>>>>>>>>>> To: "user@cassandra.apache.org" >>>>>>>>>>>>>> Subject: EC2 storage options for C* >>>>>>>>>>>>>> >>>>>>>>>>>>>> My company is planning on rolling out a C* cluster in EC2. We >>>>>>>>>>>>>> are thinking about going with ephemeral SSDs. The question is >>>>>>>>>>>>>> this: Should >>>>>>>>>>>>>> we put two in RAID 0 or just go with one? We currently run a >>>>>>>>>>>>>> cluster in our >>>>>>>>>>>>>> data center with 2 250gig Samsung 850 EVO's in RAID 0 and we are >>>>>>>>>>>>>> happy with >>>>>>>>>>>>>> the performance we are seeing thus far. >>>>>>>>>>>>>> >>>>>>>>>>>>>> Thanks! >>>>>>>>>>>>>> >>>>>>>>>>>>>> Eric >>>>>>>>>>>>>> >>>>>>>>>>>>> >>>>>>>>>>>>> >>>>>>>>>>>> >>>>>>>>>>> >>>>>>>> >>>>>>> >>>>>> >>>>>> >>>>>> -- >>>>>> Steve Robenalt >>>>>> Software Architect >>>>>> sroben...@highwire.org <bza...@highwire.org> >>>>>> (office/cell): 916-505-1785 >>>>>> >>>>>> HighWire Press, Inc. >>>>>> 425 Broadway St, Redwood City, CA 94063 >>>>>> www.highwire.org >>>>>> >>>>>> Technology for Scholarly Communication >>>>>> >>>>> >>>>> >>>> -- >>> Ben Bromhead >>> CTO | Instaclustr >>> +1 650 284 9692 >>> >> >> >
