2 ms·

> The original essay's point was to say that people over-weigh the pros, that there are cons, and that some of those cons have considerable real-world significa

by Samus_ 12y ago

> The original essay's point was to say that people over-weigh the pros, that there are cons, and that some of those cons have considerable real-world significance. > > It wasn't to say that there are no pros. It wasn't to provide a balanced analysis of all the pros and cons. It had a specific point, it made it, and it acknowledged that the point wasn't universal. > > This new essay misinterprets the point in order to set itself up as a more realistic viewpoint, when in fact it's arguing against a straw interpretation of the original essay. I don't see the pros mentioned, at all. What I see is another "Hadoop is fast, Hadoop is slow" comparison based entirely on speed which on its own isn't either a "pro" or a "con" is just a factor among many. Without the complete picture we'll keep misinterpreting and arguing over this technology for the wrong reasons. > Yes, fault-tolerance can be really important. And sometimes not. Such a discussion is orthogonal the point of the original. A single node is quite reliable. There are many ways to get redundancy on top of that. For the same price as a Hadoop cluster, you could buy a machine with more reliable hardware, plus a hot backup, plus a load balancer. This may be all that's needed. Redundancy and fault-tolerance are not the same thing, you're also thinking in the context of web development and this is a different scenario. In web you can use redundancy to improve reliability because you're either serving the same content to many clients or performing independent computations to each of them; data processing frameworks are used to perform a single computation with several different inputs over many machines, replicating it may give you some safety regarding hardware failures but none for software issues and you'll be also wasting a lot of power thus limiting the range of applications. Big Data frameworks partition the input, they do not replicate it; they also keep track of the results and allow you to retry failed chunks, it's a different approach than those used on web. > If you need 100 machines then you are still in the realm where you "really need distributed processing", which was one of the times where the original essay said that Hadoop could be the right solution. > > If you have 100 machines, and the single node can go 200x faster, then you really only need 1 machine. Like I said the speed is just one factor, you can't determine the right solution based exclusively on it and that's what the original article missed, you're disregarding all the other advantages that a cluster and a Big Data framework provide. Also the very concept of "machine" can vary wildly, there's an infinite range of combinations of disk, ram and processor so what is "a machine" and how can you say when you need one, ten or 100? The real choice here is a single "powerful enough" machine vs. a set of "commodity" hardware, then you can compare prices, speed, scalability and reliability of each scenario in all the possible situations you expect to encounter and hopefully some unforeseen ones as well.

dalke 12y ago
The four pros I spotted were "cool", "learning and having fun", handle "a huge amount of data" and "distributed processing." There's no claim that it's trying to give the complete picture. It's trying to solve a specific task for chess board analysis. Nor did it reject that Hadoop was always the wrong solution, only that a single node can be faster than Hadoop. An existence proof is a low barrier to cross. You say "this is a different scenario" and I have to shrug. I used to do molecular dynamics distributed across a compute cluster or using a supercomputer. That's yet another different scenario, and also one where Hadoop doesn't work. Last I heard, simulation software I helped write now runs across 8,400 cores. (See http://www.ks.uiuc.edu/Research/gpu/files/gtc2013_phillips.pdf http://www.ks.uiuc.edu/Research/gpu/files/gtc2013_phillips.p... .) The visualization project I used to lead now analyzes 100+ TB of simulated trajectory data. (See http://on-demand.gputechconf.com/gtc/2014/presentations/S4410-visualization-analysis-petascale-moledular-sims.pdf http://on-demand.gputechconf.com/gtc/2014/presentations/S441... .) MD definitely falls into "Big Data", yet the examples I gave don't use Hadoop or data partitioning like you suggest, so your statement about Big Data frameworks is really restricted to only a subset of what people do in Big Data. So yes, you're right. The original essay isn't complete. Neither is this new one, which didn't include all the problems with Hadoop, like poor inter-node latency, and lack of support for MPI. The 'real choice' isn't as dichotomous as you make it out to be. A set of powerful non-commodity hardware, like Blue Waters, may be the right approach for some scenarios. "Without the complete picture" I disagree. The new essay gives a different picture. A lot of pictures fit well in the frame of Hadoop, for the reasons given in the new essay. In this case the pictured scenario is chess data analysis. In this case it looks like a single node is faster. However, it's not better for the original original task, which was to experiment with learning Hadoop. ("I was skeptical of using Hadoop for the task, but I can understand his goal of learning and having fun with mrjob and EMR.") As the original original spec no longer exists, I cannot tell if fault-tolerance was one of them. As it was a personal learning project, no doubt it wasn't. So this new essay is adding details that weren't relevant to the original task, then complaining that the details weren't addressed. This is poor form. You ask "how can you say when you need one, ten or 100" and the easy answer is to go ahead and implement it on one node. If it takes an hour to write a single node solution would take, and it's fast enough for your foreseeable needs, then you're a lot better off than spending a day to try out a Hadoop solution. Plus, without the single node solution it's hard to know if you have a good baseline.

by Samus_ 12y ago

dalke 12y ago

The four pros I spotted were "cool", "learning and having fun", handle "a huge amount of data" and "distributed processing." There's no claim that it's trying to give the complete picture. It's trying to solve a specific task for chess board analysis. Nor did it reject that Hadoop was always the wrong solution, only that a single node can be faster than Hadoop. An existence proof is a low barrier to cross. You say "this is a different scenario" and I have to shrug. I used to do molecular dynamics distributed across a compute cluster or using a supercomputer. That's yet another different scenario, and also one where Hadoop doesn't work. Last I heard, simulation software I helped write now runs across 8,400 cores. (See http://www.ks.uiuc.edu/Research/gpu/files/gtc2013_phillips.pdf http://www.ks.uiuc.edu/Research/gpu/files/gtc2013_phillips.p... .) The visualization project I used to lead now analyzes 100+ TB of simulated trajectory data. (See http://on-demand.gputechconf.com/gtc/2014/presentations/S4410-visualization-analysis-petascale-moledular-sims.pdf http://on-demand.gputechconf.com/gtc/2014/presentations/S441... .) MD definitely falls into "Big Data", yet the examples I gave don't use Hadoop or data partitioning like you suggest, so your statement about Big Data frameworks is really restricted to only a subset of what people do in Big Data. So yes, you're right. The original essay isn't complete. Neither is this new one, which didn't include all the problems with Hadoop, like poor inter-node latency, and lack of support for MPI. The 'real choice' isn't as dichotomous as you make it out to be. A set of powerful non-commodity hardware, like Blue Waters, may be the right approach for some scenarios. "Without the complete picture" I disagree. The new essay gives a different picture. A lot of pictures fit well in the frame of Hadoop, for the reasons given in the new essay. In this case the pictured scenario is chess data analysis. In this case it looks like a single node is faster. However, it's not better for the original original task, which was to experiment with learning Hadoop. ("I was skeptical of using Hadoop for the task, but I can understand his goal of learning and having fun with mrjob and EMR.") As the original original spec no longer exists, I cannot tell if fault-tolerance was one of them. As it was a personal learning project, no doubt it wasn't. So this new essay is adding details that weren't relevant to the original task, then complaining that the details weren't addressed. This is poor form. You ask "how can you say when you need one, ten or 100" and the easy answer is to go ahead and implement it on one node. If it takes an hour to write a single node solution would take, and it's fast enough for your foreseeable needs, then you're a lot better off than spending a day to try out a Hadoop solution. Plus, without the single node solution it's hard to know if you have a good baseline.