Solving the Netapp Open Solution for Hadoop Solutions Guide

I have been staring at the Netapp Open Solution for Hadoop Solutions Guide. I now have a headache.

Where do I begin? It's interesting to read a datasheet about Hadoop that is targeted at enterprise filesystem people. Not management "Hadoop solves all your problems", not developer "you have to rewrite all your apps", but something designed to convince people who worry about file systems that HDFS is a wrongness on the face of the planet. That's the only reason it would make sense. And like anyone trying to position a competing product to HDFS they have to completely discredit HDFS.

This paper, which tries to rejustify splitting the storage and compute platforms -the whole thing that the HDFS/Hadoop is designed to eliminate on cost grounds- has to try pretty hard.

In fact, I will say that after reading this paper that the MapR/EMC story makes a lot more sense. As what the Netapp paper is trying to say -unlike EMC-  is "the Big Data platform that Hadoop is wonderful, provided you ignore the cost model of SATA-on-server that HDFS and equivalent filesystems offer for topology-aware applications".

This must have been a hard article to write, so I must give the marketing team some credit for the attempt, even if they got it so wrong technically.


First: did we really need new acronyms? Really? By Page 4 there is a new one, "REPCOUNT", that I have never come across before. Then the acronym "HDFS3" that I first thought meant version 3 of HDFS, but before I could switch to JIRA and see if someone had closed an issue "design, implement and release V3 of HDFS", I see it really means "A file saved to HDFS with the block.replication.factor attribute set to 3", or more simply, 3x replicated blocks. No need to add another acronym to a document that is knee deep in them.


Now, for some more serious criticisms, starting at page 5: 

"For each 1 petabyte of data, 3 terabytes of raw capacity are needed".

If that were true, this would not be a criticism. No, it would be a sign of someone getting such a fantastic compression ration that Facebooks leading edge 30 PB server would fit into 30x3TB LFF HDDs, which you could fit into about 3U's worth of rack. If only. And don't forget that RAID uses 1.6X raw storage: it replicates too, just more efficiently.

Then there's the ingest problem. A whole section of this paper is dedicated to scaring people that they won't be able to pull in data fast enough because the network traffic will overload the system. We are not -outside my home network- using Ether over Power as the backplane any more. You can now get 46 servers with 12x3TB HDDs in them onto a single rack, with a 10GbE switch that run "with all the lights on". On a rack like that - which offers 500+TB of storage- you can sustain 10 GbE between any two servers. If the ingress and egress servers are hooked up to the same switch you could in theory move a Gigabyte per second between any two nodes or in and out the cluster. Nobody has an ingress right in that range, except maybe the physicists and their "odd" experiments. Everyone else can predict their ingress rate fairly simply, it is "look around at the amount of data you discard today". That's your ingress rate. Even a terabyte per day is pretty significant -yet on a 10GbE switched you could possibly import that from a single in-centre-off-HDFS-node in under 20 minutes. For anything out of the datacentre, your bandwidth will probably be less than 10 Gigabits, unless your name is Facebook, Yahoo!, Google or Microsoft.


Summary: ingress rate is unlikely to be bottlenecked by in-rack bandwidth even with 3x replication.

Next: Intra-HDFS bandwidth issues.

The paper says today that "A typical Hadoop node has 2GbE interfaces". No. A typical node tends to have a single 1GbE interface; bonded 2x1 GbE is rarer, as it's harder to commission.
Go that way and add on separate ToR switches and all concerns about "network switch failure" go away. You'd have to lose two network switches or power to the entire rack to trigger re-replication. At least NetApp did notice that 10GbE is going on the mainboard and said "Customers building modest clusters (of less than 128 nodes) with modern servers should consider 10GbaseT network topologies."

I'd say something different, which is "128 nodes is not a modest cluster". If you are building your rack from paired six-core CPUs with 12 LFF HDDs each with 3TB of SATA (how's that for acronyms?), then your cluster has 1536 individual cores. It will have 12*3*128 = 4608TB of storage: four petabytes. That's is not modest. That is something that can sort a petabyte of data with. Fit that up with a good 10GbE switch and enough RAM -and the latest 0.23 alpha release with Todd's CRC32c-in-SSE patch- and you could probably announce you have just won the petabyte sorting record.

Summary: "a 128 node cluster built of current CPUs and SATA storage is not modest".

It's a modest size -three racks- but it will probably be the largest file system your organisation owns. Think about that. And think about the fact that when people say "big data" what they mean is "lots of low value data". If you don't store all of it (aggressively compressed), you will regret it later.

Page 6: Drive failures. This  page and its successors wound me up so much I had to stop. You see, I've been slowly doing some changes to trunk and 0.23 to tune HDFS's failure handling here, based on papers on failures by Google, Microsoft and -um- Netapp. This means I am starting to know more about the subject, and defend my statements.

The Netapp paper discusses the consequences of HDD Failure rates using the failure rate on 5K-node cluster to scare people. Yes, there will be a high failure rate there, but it's a background noise. The latest version of Hadoop 0.20.20x doesn't  require a DataNode restart when a disk fails -take it away and only that 1-3 TB of data fails. When a disk fails -and this is where whoever wrote the paper really misses the point - that 2TB of missing data is still scattered across all other clusters in the rack.

If you have 128 servers in your "modest" cluster, 2TB disks and a block size of 256MB, then there were about 8000 blocks in the disk, which are now scattered across (128-1) servers. Sixtyish blocks per server. With twelve disks per server, that's about five blocks per SATA disk (=2500 MB). Even if -as claimed- the throughput of a single SATA disk is 30MB/s, those five blocks will take under two minutes to be read off disk. I'm not going to follow this chain through to network traffic as you'd also have to factor in the fact that servers are reading network packets from blocks being replicated to it at the same time and saving them to disk too (it should expect 60 blocks incoming), but the key point is this: on a 128 node cluster, the loss of a single 2TB cluster will generate a blip of network traffic, and your cluster will carry on as normal. No need for "24-hour staffing of data center technicians". No need for the ops teams to even get texted when a disk fails. Yes, loss of a full 36TB server is a bit more dramatic, but with 10 GbE it's still manageable. This paper is just trying to scare you based on a failure to understand how block replication is implicit striping of a file across the entire cluster, and they haven't looked at Hadoop's source to see what it does on a failure. Hint: look at BlockManager.java. I have.


To summarise:


The infrastructure handles failures, the apps are designed for it. This is not your old enterprise infrastructure. We have moved out of the enterprise world of caring about every disk that fails, and into a world of statistics. 


I'm going to stop here. I can't be bothered to read the low level details if the high level stuff is so painfully wrong, either through inadequate technical review of the marketing team's world view, or deliberately through a failure to understand HDFS.

[Update 21:57 GMT]. Actually it is weirder than I first thought. This is still HDFS, just running on more expensive hardware. You get the (current) HDFS limitations: no native filesystem mounting, a namenode to care about, security on a par with NFS, without the cost savings of pure-SATA-no-licensing-fees. Instead you have to use RAID everywhere, which not only bumps up your cost of storage, puts you at risk of RAID controller failure and errors in the OS drivers for those controller (hence their strict rules about which Linux releases to trust). If you do follow their recommendations and rely on hardware for data integrity, you've cut down the probability of node-local job execution, so all FUD about replication traffic is now moot as at least 1/3 more of your tasks will be running remote -possibly even with the Fair Scheduler, which waits for a bit to see if a local slot becomes free. What they are doing then is adding some HA hardware underneath a filesystem that is designed to give strong availability out of medium availability hardware. I have seen such a design before, and thought it sucked then too.  Information week says this is a response to EMC, but it looks more like NetApp's strategy to stay relevant, and Cloudera are partnering with them as NetApp offered them money and if it sells into more "enterprise customers" then why not? With the extra hardware costs of NetApp the cloudera licenses will look better value, and clearly both NetApp and their customers are in need of the hand-holding that Cloudera can offer.

I just wish someone from Cloudera had reviewed that solutions paper for technical validity before it got published.




[Artwork: ARYZ, Nelson Street]

Lost in time

I've been following the Occupy London debacle. And it is disaster, a PR mess for the City of London and the Church of England.

One thing about the UK is there are lots of historical things about. The castles, the iron-age hill forts, the straight roads build by the romans, the town names derived from latin. Some retain their grandeur, but have (somewhat) moved on with the times, like here: Oxford, whose curricular has been upgraded for new ideas (calculus), and whose buildings are now tourist attractions as well as a place of learning.


Then there's the City of London. I've cycled through it a few times at the weekend: empty. Boring. Almost nobody lives there. What's interesting is the last fortnight has show how much of a relic of middle ages it is, how much power it has -and, therefore, how much power the business based there have.

The key points are

• It has special rights w.r.t parliament and "the crown", including lobbying powers in parliament -a parliament which it considers itself mostly exempt from.
• It's electorate is derived not just from the citizens, but from businesses and organisations based in the City. The citizens have the minority vote.
• The process for becoming leader of the city is some process that doesn't even pretend to be democratic.

What it means today is this: the part of Britain which contains the headquarters of the most powerful business in the city is effectively self-governed by those business, and independent from the rest of the country.


Which is funny, as the People's Republic of Stokes Croft has taken up that "Passport to Pimlico" theme of an independent state within a city. They just do it as joke. Well, the joke is on them.

Reading: Availability in Globally Distributed Storage Systems

I am entertaining myself in test runs by reading Ford et al's paper,  Availability in Globally Distributed Storage Systems, from Google. This is a fantastic paper and it shows how large datacentre datasets themselves give researchers a great advantage, so it's good that we can all read it now it's been published.

Key points

1. storage node failures are not independent
2. most failures last less than 15 minutes, so the liveness protocol should be set up to not worry before then
3. most correlated failures are rack-local, meaning whole racks, switches, rack UPSs fail.
4. Managed operations (OS upgrades &c) cause a big chunk of outages.

The Managed Operations point means that ops teams need to think about how they upgrade systems in a way that don't cause replication storms.  Not directly my problem.

Point #3 is. The paper argues that if you store all copies of the data on different racks, you get far better resilience than storing multiple copies of the data on a single rack -that being exactly what Hadoop does. Hadoop has some hard-coded rules that say "two copies per rack are OK", which is done to save bandwidth on the backplane.

Now that  switches that offer fantastic backplane bandwidth are available from vendors like HP at prices that Cisco wouldn't dream of, rack locality matters less. What you save in bandwidth you lose in availability. Lose a single rack and you have to make 2x copies of every block that was created in that rack and that now only has one copy, and that is where you are vulnerable to data loss.

That needs to be fixed, either in 0.23 or its successor.

Hadoop work in the new API in Groovy.

I've been doing some actual Map/Reduce work with the new API, to see how it's changed. One issue: not enough documentation. Here then is some more, and different in a very special way: the code is in Groovy.

To use them: get the groovy-all JAR on your Hadoop classpath and use the groovyc compiler to compile your groovy source (and any java source nearby) into your JAR, bring up your cluster and submit the work like anything else.

This pair of operations, part of a test to see how well Groovy MR jobs work just counts the lines in a source file; about as simple as you can get.

The mapper:

package org.apache.example.groovymr

import org.apache.hadoop.io.IntWritable
import org.apache.hadoop.io.Text
import org.apache.hadoop.io.LongWritable
import org.apache.hadoop.mapreduce.Mapper

class GroovyLineCountMapper extends Mapper {

    final static def emitKey = new Text("lines")
    final static def one = new IntWritable(1)

    void map(LongWritable key,
             Text value,
             Mapper.Context context) {
        context.write(emitKey, one)
    }
}

Nice and simple; little different from the Java version except

• Semicolons are optional.
• The line ending rules are stricter to compensate, hence lines end with a comma or other half-finished operation.
• You don't have to be so explicit about type (the def declarations) -and let the runtime sort it out. I have mixed feelings about that.

There's one other quirk in that the Context parameter for the map operation (which is a generic type of the parent class) has to be explicitly declared as Mapper.Context. I have no idea why, except that it won't compile otherwise. The same goes for the Reducer.Context

Not much to see there then. What is more interesting is the reduction side of things.

package org.apache.example.groovymr
import org.apache.hadoop.io.IntWritable
import org.apache.hadoop.io.Text
import org.apache.hadoop.mapreduce.Reducer

class GroovyValueCountReducer 
        extends Reducer {

    void reduce(Text key,
                Iterable values,
                Reducer.Context context) {
        int sum = values.collect() { it.get() } .sum()
        context.write(key, new IntWritable(sum));
    }
}

See the line in bold? It's taking the iterable of the list of values for that key, applying a closure to them (getting the values), which returns a list of type integer, which is then all summed up. That is: there is a per-element transform (it.get()) and a merging of all the results (sum()). Which is. when you think about it, almost a Map and a Reduce.

It's turtles all the way down.

[Artwork: Banksy, 2009 Bristol exhibition]

Microsoft and Hadoop: interesting

I am probably expected to say negative things about the recent announcement of Microsoft supporting Apache Hadoop(tm) on their Azure Cloud Platform, but I won't. I am impressed and think that it is a good thing.

1. It shows that the Hadoop-ecosystem is becoming more ubiquitous. Yes it has flaws, which I know as one of my IntelliJ IDEA instances has the 0.24 trunk open in a window and I am staring at the monitoring2 code thinking "why didn't they use ByteBuffer here instead of hacking together XDR records by hand" (FWIW, Todd "ganglia" Lipcon says it aint his code). Despite these flaws, it is becoming widespread,
2. That helps the layers on top, the tools that work to the APIs, the applications.
3. This gives the Hadoop ecosystem more momentum and stops alternatives getting a foothold. In particular the LexisNexis stuff -from a company that talk about "Killing Hadoop". Got some bad news there...
4. Microsoft have promised contribute stuff back. This is more than Amazon have ever done -yet AWS must have found truckloads of problems. Everyone else does: and we file bugs, then try to fix them. I could pick any mildly-complex source file in the tree, do a line-by-line code review and find something to fix, even if its just better logging or error handling. (don't dismiss error handling BTW
5. If Amazon have forked, they get to keep that fork up to date.
6. If MS do contribute stuff back, it will make Hadoop work properly under Windows. For now you have to install Cygwin because Hadoop calls out to various unix commands a lot of the time. A windows-specific library for these operations will make Hadoop not only more useful in Windows clusters, it will make it better for developers.
7. MS will test at scale on Windows, which will find new bugs, bugs that they and Hortonworks will fix. Ideally they will add more functional tests too.
8. I get to say to @savasp that my code is running in their datacentre. Savas: mine is the networking stuff to get it to degrade better on a badly configured home network.  Your ops team should not encounter this.

It's interesting that Microsoft have done this. Why?

• It could be indicative of a low takeup of Azure outside the MS enterprise community. I used to do a lot of win32 programming (in fact I once coded on windows/386 2.04); I don't miss it, even though Visual Studio 6 used to be a really good C++ IDE. It is nicer to live in the Unix land that Kernighan and Ritchie created.(*)
• Any data mining tooling encourages you to keep data, which earns money for all cloud service providers.
• The layers on top are becoming interesting. That's the extra code layers, the GUI integration, etc.
• There's no reason why enterprise customers can't also run Hadoop on windows server within their own organisations, so integrate with the rest of their world. (I'm ignoring cost of the OS here, because if you pay for RHEL6 and CDH then the OS costs become noise).
• If you are trying to run windows code as part of your MR or Pig jobs, you now can. 
• If you are trying to share the cluster with "legacy" windows code, you can.

Do I have any concerns?

• Somehow I doubt MSFT will be eating their own dogfood here; this may reduce the rate they find problems, leaving it to the end users. Unless they have a fast upgrade rate it may take a while for those changes to roll out. (Look at AWS's update rate: sluggish. Maybe because they've forked)
• To date, Hadoop is optimised for Linux; things like the way it execs() are part of this. There is a risk that changes for Windows performance will conflict with Linux performance. What happens then?
• I forsee a growth in the out-of-depth people trying to use Hadoop and asking newbie questions now related to Windows. Though as we get them already, there may be no change.
• I really wish Windows server had SSH built in rather than telnet. Telnet is dead: move on. We want SSH and SFTP filesystems, and an SFTP filesystem client in both Windows and OS/X. It's the only way to be secure.
• I hope we don't end up in an argument over which underlying OS is best. The answer is: the one you are happy with.

(*) At least for developers. I changed the sprog's password last week as were unhappy with him, and when I passwd'd it back he asked me "why do I use the terminal?". One day he'll learn. I'll know that day as I he'll have changed my password.

[Artwork: unknown, Moon Street, Stokes Croft]

Oracle and Hadoop part 2: Hardware -overkill?

[is the sequel to Part 1]

I've been looking at what Oracle say you should run Hadoop on and thinking "why?"

I don't have the full specs, since all I've seen is slideware on the register implying this is premium hardware, not just x86 boxes with as many HDDs you can fit in a 1U with 1 or 2 multicore CPUs and an imperial truckload of DRAM. In particular, there's mentioning of Infiniband in there.

InfiniBand? Why? Is it to spread the story that the classic location-aware schedulers aren't adequate on "commodity" 12-core 64GB 24TB with 10GbE interconnect? Or are there other plans afoot?

Well, one thing to consider is the lead time for new rack-scale products, some of this exascale stuff will predate the oracle takeover, and the design goals at the time "run databases fast" met Larry's needs more than the rest of the Sun product portfolio -though he still seems to dream of making a $ or two from every mobile phone on the planet.

The question for Oracle has been "how to get from hardware proto to shipping what they can say is the best Oracle server." Well, one tactic is to identify the hardware that runs Oracle better and stop supporting it. It's what they are doing against HP's servers, and will no doubt try against IBM when the opportunity arises. That avoids all debate about which hardware to run Oracle on. It's Oracle's. Next question? Support costs? Wait and see.

While all this hardware development was going on, the massive-low-cost GFS/HDFS filesystem with integrate compute was sneaking up on the sides. Yes, it's easy to say -as Stonebraker did- that MapReduce is a step backwards. But it scales, not just technically, but financially. Look at the spreadsheets here. Just as Larry and Hurd -who also seemed over-fond of the Data Warehouse story- are getting excited about Oracle on Oracle hardware, somewhere your data enters but never leaves(*), somebody has to break them the bad news that people have discovered an alternative way to store and process data. One that doesn't need ultra-high-end single server designs, one that doesn't need oracle licenses, and one that doesn't need you to pay for storage at EMC's current rates. That must have upset Larry, and kept him and the team busy on a response.

What they have done is defensive actions: Hadoop as a way of storing the low value data near Oracle RDBMS, for you to use it as the Extract-Transform-Load part of the story. Where it does fit in, as you can offload some of the grunge work to lower end machines, the storage to SATA. It's no different from keeping log data in HDFS but the high value data in HBase on HDFS, or -better yet IMO- HP Vertica.

For that story to work best, you shouldn't overspec the hardware with things like InfiniBand. So why has that been done?

• Hypothesis 1: margins are better, helps stop people going to vendors (HP, Rackable), that can sell the servers that work best in this new world.
• Hypothesis 2: Oracle's plans in the NoSQL world depend on this interconnect.
• Hypothesis 3: Hadoop MR can benefit from it.

Is Hypothesis 3 valid? Well, in Disk-Locality in Datacenter Computing Considered Irrelevant [Ananthanarayanan2011], Ganesh and colleagues argue that improvements in in-rack and backplane bandwidth will mean you won't care whether your code is running on the same server as your data, or even the same rack as your data. Instead you will worry about whether your data is in RAM or on HDD, as that is what slows you down the most. I agree that even today on 10GbE rack-local is as fast as server-local, but if we could take HDFS out the loop for server-local FS access, that difference may reappear. And while eliminating Ethernet's classic Spanning Tree forwarding algorithm for something like TRILL would be great, it's still going to cost a lot to get a few hundred Terabits/s over the backplane, so in large clusters rack-local may still have an edge. If not, well, there's still the site-local issue that everyone's scared of dealing with today.

Of more immediate interest is Can High-Performance Interconnects Benefit Hadoop Distributed File System?, [Sur2010]. This paper looks at what happens today if you hook up a Hadoop cluster over InfiniBand. They showed it could speed things up, even more if you went to SSD. But go there and -today- you massively cut back on your storage capacity. It is an interesting read, though it irritates me that they fault HDFS for not using the allocateDirect feature of Java NIO, and didn't file a bug or fix for that. See a problem, don't just write a paper saying "our code is faster than theirs". Fix the problem in both codebases and show the speedup is still there -as you've just removed one variable from the mix.


Anyway, even with that paper, 10GbE looks pretty good, it'll be built in to the new servers and if the NIO can be fixed, it's performance may get even closer to InfiniBand. You'd then have to move to alternate RPC mechanisms to get the latency improvements that InfiniBand promises.

Did the Oracle team have these papers in mind when they did the hardware? Unlikely, but they may have felt that IB offers a differentiator over 10GbE. Which it does, in cost terms, limits of scale and complexity of bringing up a rack.

They'd better show some performance benefits for that -either in Hadoop or the NoSQL DB offering they are promising.

(*) People call this the "roach motel" model, but I prefer to refer to Northwick Park Hospital in NW London. Patients enter, but they never come out alive.

Oracle and Hadoop part 1

I suppose I should add a disclaimer that I am obviously biased against anything Oracle does, but I will start by praising them for:

1. Recognising that the Hadoop platform is both a threat and an opportunity. 
2. Doing something about it.
3. Not rewriting everything from scratch under the auspices of some JCP committee.

Point #3 is very different from how Sun would work: they'd see something nice and try and suck it into the "Java Community Program", which would either get it all overweight from the weight that Standards Bodies apply to technology (EJB3 vs Hibernate), stall it until it becomes irrelevant (most things), or ruin a good idea by proposing a complete rewrite (Restlet vs JAX-RS). Overspecify the API, underspecify the failure modes and only provide access to the test suite if you agree to all of Sun's (and now Oracle's) demands.

No, they didn't go there. Which makes me worry: what is their plan. I don't see Larry taking to the idea of "putting all the data that belongs to companies and storing it a filesystem and processing platform that I don't own". I wonder if the current Hadoop+R announcement is a plan to get in to the game, but not all of it.

Or it could just be that they realised that if they invited Apache to join some committee on the topic they'd get laughed at. 

Sometime soon I will look at the H/W specs and wonder why that was so overspecified. Not today.

What I will say that irritated me about Oracle World's announcements was not Larry Ellison, it was Mark Hurd saying snide things about HP how Oracle's technologies were better. If that's the case, given the time to market of new systems, isn't a critique of Hurd himself? That's Mark "money spent on R&D is money wasted" Hurd? That's Mark whose focus on the next quarter meant that anything long term wasn't anything he believed in? That is Mark Hurd whose appointed CIO came from Walmart, and viewed any "unofficial" IT expenditure as something to clamp down on? Those of us doing agile and advanced stuff: complex software projects ended up using Tools that weren't approved: IntelliJ IDEA, JUnit tests, Linux desktops, Hudson running CI. The tooling we set up to get things done were the complete antithesis of the CIO's world view of one single locked down windows image for the entire company, a choice of two machines: "the approved laptop" and "the approved desktop". Any bit of the company that got something done had to go under the radar and do things without Hurd and his close friends noticing.

That's what annoyed me

(Artwork: something in Eastville. Not by Banksy, before anyone asks)