How to Analyze Billions of Records per Second on a Single Desktop PC
clemenswinter.com
clemenswinter.com
So if not memory-bandwidth, what is the constraining factor? Or specifically, what's the limiting factor that causes the observed benchmark speeds for LocustDB?
My guess would be that a well designed in-memory database should be limited by memory-bandwidth, but that real-world memory-bandwidth shouldn't be thought of as a single number. Instead, it depends on the size and pattern of requests, the latency of different cache hierarchies, and (crucially but often forgotten) on the number of requests in flight.
Is there a better terminology that distinguishes these two uses of memory-bandwidth?
theoretical peak bandwidth
effective/observed bandwidth
I'm looking for a different denominator, which is also theoretical. Maybe call it the "theoretically achievable bandwidth", which takes into account all the details of the requests. If the "observed bandwidth" equals the "theoretically achievable bandwidth", your only path to improvement is to change your data layout or increase the parallelism of your requests. If the "observed bandwidth" is less than this theoretical, there should still be room for implementation optimizations.
Falling short of the "theoretical peak bandwidth" (even by a lot) doesn't tell you which of these is the case.
Of course you are right that this is not the whole picture, it is possible to be bottlenecked on neither memory bandwidth nor CPU throughput.
With respect to memory, two additional considerations are read amplification and latency.
Read amplification: Read amplification basically means that you are reading data but not actually using it. Suppose we have a table stored in row format with four columns C1, C2, C3, C4 which store 8 byte integers. Row format means that the values for the columns are interleaved, so the memory layout would look like this, where cn refers to some value for column Cn:
0x0000 c1 c2 c3 c4 c1 c2 c3 c4
0x0040 c1 c2 c3 c4 c1 c2 c3 c4
...
Latency: Suppose we also don't have any read amplification and make use of all values for each loaded cache line. Then after we finish processing a cache line, and load the next cache line, we might have to wait a long time (100s of cycles) before the new cache line actually arrives! So then we would actually be stalled because of memory latency during some parts of program execution.
One of the big advantages of using columnar layout (which stores all values for a column sequentially) is that it all but eliminates read amplification. Similarly, by accessing data sequentially we make it possible for the CPU to predict what data will be accessed in the future, and have it loaded into the cache before we even request it.
Of course there's a number of reasons why things might not work out quite so perfectly in practice, and those may why performance in this benchmark starts to degrade before the (usable) memory bandwidth reaches the theoretical maximum memory bandwidth. But as you can see it still possible to get very close, so it's safe to assume that queries which read much less data than the maximum bandwidth are constrained on CPU.
Your other points make sense. And those would surface as constraints in a similar fashion as a CPU throughput limitation would. I just expect those well before a CPU constraint.
I confess I have not researched CPU speeds in quite a while. Nor RAM speeds. Just surprised to see that the CPU's throughput would ever really be the bottleneck.
Sort of. Random main memory access speeds aren't there, but L1 is; as long as your memory accesses are sequential, proper coding can stream sequential main memory into L1 and keep the CPU working.
But it's effectively a lost art - the vast majority of programs spend a significant part of their time stalling on memory access.
My metaphor is a tub. If the drain is faster than the faucet, filling the tub before opening the drain will still see an empty tub.
Only way you can prevent that is if the faucet is the same speed as the drain. Right?
Now, again, the cache can help you see more time where you are getting max utilization. Hard to believe it would be a full workload, though.
DDR4 can do >20GB/s but the latency is >10ns. If you prefetch enough cycles in advance, data will be in L1 by the time you need it. If you don’t, it’s won’t.
Regardless, I've learned that RAM has gotten much faster than I would expect. Cool to see.
[1] https://ark.intel.com/products/83503/Intel-Core-i7-4980HQ-Pr...
That said, any nontrivial computation is unlikely to keep up with that speed (e.g. 64-bit data will be ~3GD/s, so whatever computation you do has to fit in amortized 1 cycle on a 3Ghz clock, or the CPU is the bottleneck)
GD=GigaDatum
Again, thanks for sticking with me in this thread. I learned a bit. Hoping to come up with ways to play with this knowledge. :)
I think it's also not even by DIMM slot but by channel, so if the channels are unevenly populated (e.g. 2 slots filled on one channel but only 1 on another), interleaving is disabled.
The constraining factor in this case is CPU throughput.
When you looked at it with VTune, did you find clear evidence of this? That is, were you consistently executing 4 instructions per cycle, or were you maxed out at 100 percent utilization for a particular execution port? While it's possible to be at these limits (say if you were hashing each input) in most cases I think it's rare to get to that point when reading from RAM.
Similarly, by accessing data sequentially we make it possible for the CPU to predict what data will be accessed in the future, and have it loaded into the cache before we even request it.
One important caveat to this (at least on Intel) is that the hardware prefetcher does not cross 4KB page boundaries. So if you are doing an 8B read every cycle, and if you are able to do your other processing in the other 3 µops available that cycle, it's taking you ~500 cycles to process each page. And then you hit a ~200 cycle penalty when you finish each page, which means you are taking a ~25% performance hit because of memory latency, and probably have a CPI closer to 3 rather than 4! So if you (or your compiler) haven't already done so, you might see a useful performance boost by adding in explicit prefetch statements (one per cacheline) for the next page ahead.
Anyway, thanks for your response. I like your approach, and I wish you luck in making it even faster!
If you take any basic computer architecture class you'd know that this is not the case. Unless there are breakthroughs within the past 5 years i have not read about. DRAM is at least 10x slower than CPU. Unless your prefetcher is so accurate it can feed the i-cache/d-cache without any stalls on the CPU side.
All of these "fast" databases have a fatal flaw - it takes forever to load the data in the first place. In this case loading takes >1000x longer than the query so loading is all that matters.
Your complaint only makes sense if you only intend to perform one single query against a dataset.
The response time is quite fantastic (less than a minute), but there are drawbacks: - GC takes in the order of minutes to complete - loading the data takes 90 min from spinning rust
So yeah, it's possible, but your bottleneck becomes the storage driver.
Source: I work at SAP's internal cloud team. For our internal customers running SAP HANA, we offer baremetal machines with up to 6 TiB RAM.
Heavy fixed processing costs up front with the benefit of light variable processing moving forward.
There's a lot more to data analysis than loading data off disk.
It's far from the correct title. A good part of the optimisation of all modern analysis DBs is in-memory... this is no different and as such he's compared his work against 2 of the major other systems in the space.
I see a lot of critics here... I think the guy should be congratulated for his work - which if you've read through all pages, appears to be done single handed!
They even admit that they couldn't be bothered to modify their ingestion scripts to not partition their data.
>The pickup_ntaname column is stored as varchars by ClickHouse and kdb+, and as dictionary encoded single byte values by LocustDB.
But it would be trivial to convert to enum/sym type in kdb+. It's silly to query and group by strings.
But even then, hardware/networking performance and tuning can throw a wrench in the most seasoned/knowledgeable approaches. Users can further bring otherwise solid setups to a grinding halt with unanticipated use-cases.
The only hope when you hit these inevitable road-blocks is that you're working for someone that appreciates the difficulty of the problem.
> One note about the results for kdb+: The ingestion scripts I used for kdb+ partition/index the data on the year and passenger_count columns. This may give it a somewhat unfair advantage over ClickHouse and LocustDB on all queries that group or filter on these columns (queries 2, 3, 4, 5 and 7). I was going to figure out how to remove that partitioning and report those results as well, but didn’t manage before my self-imposed deadline.
So yeah, maybe the guy has no idea what to do with kdb - but ultimately having a fast, free and open-source database & query language beats a fast and expensive piece of commercial software.
Looking over the queries, he made some... interesting changes that have him benchmarking oranges to apples.
Having read both this post and the documentation for TrailDB, I don’t think they are comparable past very simple use cases. This post seems to be covering a complete database system with a query parser, planner, and (previously?) a storage engine. TrailDB is more like a storage engine you can push certain kinds of filters down into, but performs neither query parsing nor planning for you.
I also wouldn't be surprised if someone has done a sqlite backend for it.
How to Analyze Billions of Records per Second on a Supercomputer
Present:
How to Analyze Billions of Records per Second on a Single Desktop PC
Future:
How to Analyze 5 Records per Second on Thousands of Machines Distributed Across the Globe Over the Blockchain.
Software, for better or worse, is a social enterprise, so you can get away with dumping a lot on your users before they pack up and look for alternatives.
There is an axiom that software gets bigger, for no particularly good reason, for at least 40 years.
This can be witnessed in how the size of /bin/true on Unix went from 0 bytes to whatever size it is today [1]. I can't find it now, but there was an article with a bar chart showing a gradual progression through time.
Of course, on-disk size doesn't translate directly to performance, but the trivial example demonstrates mechanisms by which bloat/cruft can accumulate so gradually that no outcry is ever caused.
It also refutes the sibling/downthread comment the "cruft" is (necessarily) someone else's use case.
[1] e.g. 13k MacOS native, 52k GNU coreutils from MacPorts
Sounds like a computer variant of Parkinson's law: "Work expands to fill the time available for its completion."
Alternatively, with the complaint that desktop software feels more sluggish on a modern computer than contemporaneous software did on much older hardware suggests that something is even being overfilled.
My intuition says there's a different phenomenon of psychology at work here.
Why does the Word window not appear within milliseconds of me clicking on the application icon?
I'd also suggest just using it more. If you find yourself wanting to use grep, cut or sed (especially if you need more than one!) for a one-liner, try using awk instead, if only for the practice. Once you're accustomed to some of the idioms, simplicity, and built-in looping, it will feel like a more natural, casual text-processing tool that you'd reach for automatically.
Lastly, search for "cookbooks" or libraries or other collections of useful scripts to see how others have used the language, in more advanced ways. Personally, I don't necessarily find it as something to aspire to (i.e. use of the language for complex, multi-line programs) as much as a way to better understand some of the power of the quirkier features that may, otherwise, be less obvious.
One of my favorite non-obvious behaviors is bypassing the input/loop functionality by having the only pattern be BEGIN. This results in an output-only program that makes for a convenient way of playing around with things like printf or just language syntax features (or differences between awk/mawk/gawk), without having to redirect input from /dev/null or worrying about hitting control-D twice and exiting the shell (for those of us who abhor ignoreeof):
awk 'BEGIN{print "Hello, world!"}'It always astounds me these days when someone manages to release slow software for a desktop computer despite modern systems being orders of magnitude faster than the first desktop computers while often being no more responsive.
Edit: Also I'm amused by how much of a nerve this seems to have hit. I guess some people are defensive about their high performance computing approaches... :)
Fortunately, when articles about real HPC clusters (aka Supercomputers) make it to the front page, they need no defense, only an occasional explanation that what makes them (extra) special are the high-bandwidth, low-latency interconnects. (Those interconnects make a very strong effort at Fallacies of Distributed Computing numbers 2 and 3[1] and tend to neatly take care of the remaining 6).
To be fair, though, I don't think the claim for the more common distributed systems is that they're "high performance" so much as that they're scalable (and have other benefits of being multi-node like resilience).
[1] The bandwidth may well be indistinguishable from infinite if it exceeds local (e.g. CPU-memory, CPU-CPU, CPU-GPU) bandwidths. I don't think that's yet true for something like multi-socket Intel systems, but it might be possible with enough interface cards. I didn't look at the specs on those POWER9 chips.
If you're processing an amount of data that comfortably fits in memory on a single machine, then obviously Hadoop is going to perform poorly in comparison. The costs of scheduling a job onto N mappers/reducers, transferring code to each node, waiting for the slowest mapper/reducer to finish, transferring data from mappers -> reducers, replicating output on HDFS, etc. are well-understood. It's true that many people try to use Hadoop when they'd be better served with simpler solutions, but that does not justify the amount of shade that the author throws at it.
The Adam Drake article alludes to it only in the last sentence by mentioning traditional relational databases as an alternative.
For workloads that are relatively I/O-heavy and CPU-light, it's very hard to beat local SSDs (or even HDDs in enough quantity) attached to a single [1] node, if the competition is distributed storage attached by ethernet. It only takes a couple 600MB/s SSDs to saturate a 10GE. A server with 48 lanes of PCIe 3 slots could take the I/O of 78 of them.
100Gb/s networks are getting close. For upwards of $1k per server (NIC and switch port) one can bring that ratio closer to 4:1 from 39:1. I'd expect this is the attractive route for anyone with CPU needs that can't be met by a 4-socket server.
[1] Yes, there can be more than one node with copies for redundancy, as has been complained about elsewhere in the sub-thread, or even scalability
I posit that these two assertions are contradictory.
My own understanding of the term "well understood" is that it is synonymous with "widely understood". If many people are still making the mistake of using Hadoop when those costs outweight the benefits, it seems that understanding isn't quite wide enough.
That said, although I have a grasp of when the tradeoff is so loopsided as to be obvious, I don't know where to go (or where to point other people to go) for a better understanding of where the boundary is.
Where should we go to better learn that understanding of those costs?
The problem with the article is that if it's for a general audience that doesn't understand the tradeoffs of a system like Hadoop, it really paints a picture that it is just a bad, slow tool. It barely acknowledge just how rigged the comparison is at all, aside from mentioning that you might need something like Hadoop for really big data in the conclusion, while it is peppered with unnecessarily snide comments about Hadoop that will probably be more memorable. I think it is liable to leave readers more confused about the tradeoffs involved after reading than before.
So.. It's well understood by True Scotsmen? :) I certainly understood that there are experts in the field who have a deep, even intuitive understanding of how and when to use which tools. To the extent that those experts don't communicate that knowledge to a broader audience while at the same time may be advocating for the use of the tools, they bear some responsibility for the misuse.
My point wasn't so much that you used imprecise, but rather, that the statement about how well it's understood was inaccurate or irrelevant (depending on which definition you were going for).
> But as a general rule of thumb, my feeling is that for tens to hundreds of GB, you should consider it. And for TBs or more, you almost certainly want to be doing something distributed
While a 3-4TB cutoff makes some sense if ones workload has to remain in-memory for performance reasons, that can't be anywhere near the cutoff for any kind of workload that could stand to read from SSDs.
> I don't know if there's any resource out there that really goes deep into the tradeoffs involved though. There probably is, given how popular the subject is, but I'm not aware of one.
I would hope so, but I'm not so sure. It may not even need to be very deep, something akin to the "5 minute rule" for memory/disk caching. Mostly, I'm not convinced that the subject of tradeoffs is actually popular, so much as just using the tool without considering them is.
> The problem with the article is that if it's for a general audience that doesn't understand the tradeoffs of a system like Hadoop, it really paints a picture that it is just a bad, slow tool.
If we're talking about the adamdrak.com 233x article, I have to disagree, as my read of it was that it focused on evangelizing the "under-used approach for data processing" of "standard shell tools and commands".
> it is peppered with unnecessarily snide comments about Hadoop that will probably be more memorable
That's certainly not a charitable interpretation, and I would hazard that it's not even fair or factual (as to "peppered", at least). It's mentioned only a handful of times:
> Command-line Tools can be 235x Faster than your Hadoop Cluster
I agree that click-bait can be considered snide.
> I was skeptical of using Hadoop for the task, but I can understand his goal of learning and having fun with mrjob and EMR
To me, this comment, this very first mention of Hadoop in the intro, made it clear that this was "rigged" in that the "competition" was neither competing nor truly concerned about performance.
> while the Hadoop processing took about 26 minutes (processing speed of about 1.14MB/sec).
Merely a factual summary. Nothing snide that I could detect.
> Although Tom was doing the project for fun, often people use Hadoop and other so-called Big Data ™ tools for real-world processing and analysis jobs that can be done faster with simpler tools and different techniques.
This seems like just a restatement of the admission in the introduction plus the assertion (that I believe even you agree with) that many people mis-use Hadoop when it's not called for.
> The resulting stream processing pipeline we will create will be over 235 times faster than the Hadoop implementation and use virtually no memory.
Again, just another factual summary, with no snideness I could detect.
> While we can certainly do better, assuming linear scaling this would have taken the Hadoop cluster approximately 52 minutes to process.
This next mention is after at least half of the bulk of the article. It may be an assuming-spherical-cows estimation, but it doesn't strike me as grossly misleading on its face, and there's no editorializing.
> This gets us up to approximately 77 times faster than the Hadoop implementation.
> about 174 times faster than the Hadoop implementation.
> gets us down to a runtime of about 12 seconds, or about 270MB/sec, which is around 235 times faster than the Hadoop implementation.
The next three mentions, near the end, are just comparisons of the evolving demonstration implementation to the reference implementation. No detectable snideness.
> Hopefully this has illustrated some points about using and abusing tools like Hadoop
> but more often than not these days I see Hadoop used
Here in the conclusion paragraph is where I agree that there is both snideness and where a reader may be confused about tradeoffs, if that's what they were expecting to be enlightened about.
However, because that's not what the introduction promised, my criticism would be merely that the conclusion doesn't match the introduction (and maybe goes too far into inflammatory territory with "abuses").
Pretend that section isn't even in the article, and the article still reads OK.
I still disagree about the article itself (even outside of the conclusion), but perhaps I am reading it uncharitably and other people are not getting the same impression. I do feel that it would be easy for someone who is not very familiar with these technologies to get the wrong impression. Misusage does probably go mainly in the other direction (of people overusing Hadoop rather than underusing), though, so maybe that is not so important a concern.
I've elaborated on these in another comment, as well.
Today, drawing the line at a single terabyte is way too early, even for all-in-memory workloads, if only because there exists an almost 4TB AWS instance now. Any smaller than 3.5TB (or whatever RAM is available to applications) is, at best, living in the past.
> scalability if requirements change
This reads as premature optimization, which turns the strong argument into either a weak argument or even an argument against.
Now, if you know or have reasonable certainty that your requirements will change (and will do so faster than, say Moore's Law) and change soon, then that's different. I suspect there are people who think this, but that it's little more than wishful thinking or a delusion as to how large their slice of "web scale" actually is.
> resilience to machine failures
Machine failures just aren't a legitimate consideration for modern, high-end (but still commodity) hardware. You wouldn't bet your whole business on it, of course, but a 1% chance every year of losing an hour or two of batch processing? Sure.
Sadly, the flip side of this is that I see Hadoop clusters being built with such reliable servers, including redudant PSUs and fans, instead of taking full advantage of the resilience at the software level in order to save as much as possible at the hardware level. The original company behind map-reduce is certainly not splurging on hardware.
I can't really comment on rates of machine failures, but I have seen it happen before, even just for stupid reasons like someone in a data center unplugging a machine.
Surely, though, workloads that require in-memory performance are fairly niche, and jumping from there to distributed (even in-memory) seems non-obvious, at best. Why aren't large arrays of fast SSDs a better alternative? The bandwidth is comparable, but the latency is terrible (still comparable to ethernet to a remote node, though?)
What about workloads that don't require fully-in-memory in the first place? If the cutoff is, then hundreds of TB, wouldn't that cover the vast majority of common use cases?
> I can't really comment on rates of machine failures, but I have seen it happen before, even just for stupid reasons like someone in a data center unplugging a machine.
That sort of anecdata isn't very useful, because a human can cause any failure at any layer, including someone stop a whole cluster, which I've seen happen before.
My point about it not being a legitimate concern is that what is now common practice with what is now common equipment means it's uncommon. These practices and equipment had to evolve, but that evolution happened on the order of over a decade ago.
Also, be wary of selection bias. It's very easy to remember the "fire drill" because of the one machine failure, and it makes a much more interesting story to tell that gets passed around and modified enough, eventually sounding like multiple stories and therefore multiple machines. The hundreds of servers that operated unheard and unseen for years, sometimes beyond their specs (e.g. with only only blower out of four still turning and only half-speed at that), get nary a thought, let alone mention.
We need to substitute the word "professionally" with more precise terms when talking about our industry. Because if one was to read "professionally" as "at work", then your statement is absolutely false - both the lower bound and average amounts of critical thinking and caring in this industry are extremely low. Even ignoring people who obviously have no clue, I can still imagine Hadoop and other big data stacks being sanctioned by "professionals" in management for buzzword-generating reasons, and implemented by "professional" "engineers" for CV padding reason.
I's entirely possible that 9 out of 10 people on a project using Hadoop know it's a waste but there are non-technical reasons for doing so. Resume padding and PHB demanding some technology of the week would be the two most common ones.
That said it probably is contradictory much of the time. I'd say the majority of current developers don't know about the simpler tools.
That's probably true but, cynically, I tend to think a lot of the time when people use Hadoop they're not doing it because Hadoop is the best solution: they're doing it because the solution allows them to use Hadoop.
I'm not suggesting that something like resume-padding is never a motivation, just that it seems unlikely to be the sole or even primary motivation.
Now, I may be wrong, but my reasoning behind this is that, for all its purported power as a tool, as potatoyogurt has alluded [1], it's really only the true experts who can wield that power effectively. The ones who merely succeeded at getting on their resume would only be able to use it where its power isn't truly needed (as was the case before).
If the technology falls out of fashion, then those resume-holders will be in a position of needing to deliver the cargo, rather than carved wooden headphones.
As such, I believe their motivation is actually to attempt to learn the true power of the tool (i.e. true resume building, rather than mere padding) and that they're grossly underestimating the costs through ignorance and a desire to learn by doing.
The trouble is, without well-published non-distributed reference implementations and, instead, ultra-popularity of the distributed tools instead, they never end up learning those costs, and we're in a state of perpetuated ignorance and perpetuated over-use.
[1] by saying that the cost trade-offs are well understood by the experts, which strongly implies that the experts have a pretty deep understanding of the actual mechanics of distributed computing in general.
Both because of familiarity with querying and the solidity of running a multi-tenant system.
But I do recommend that they switch to MapR [c++ core and a passable central FS for unix-based super fast queries] if they're concerned with efficiency.
[For context, in my day job we do multiple clusters of millions of network traffic summaries/sec and are often replacing Hadoop, or more recently, ELK, as people tried to use them for that use case. All well >>> will fit in ram. We have our own in-house column-store + streaming combo db done in go/c/c++ that started as clustering fastbit.]
I doubt the point of the article was to suggest that linux cli stuff would scale to hundreds of users on the same host, but, if each of those users has a host of their very own, such as a laptop, the model could scale very well indeed, for small enough datasets.
> I wouldn't recommend to most CIOs they replace Hadoop in all or maybe even most cases, even for few-TB data sets and smaller.
Well, as you point out later, regarding familiarity, once it's in, it's probably too late. What about for a new implementation?
In answering this question, don't get too hung up on a literal interpretation of "single" server being exactly one. For example, a traditional RDBMS with one or more replicas (for performance, redundancy, or both) would still fall under the single server model. Really, it's about the non-distributed-computing option.
> if they're concerned with efficiency.
The fact that this is an "if" (and I do know that it is, even for startups) is bewildering to me, even more so in the context of distributed architectures where scaling is less linear the more data that has to be shared.
So I think the latter issues dominate (culture/training) for new implementations as well as existing.
Re: efficiency overall, and of distributed systems, it's interesting to see both that MapR has made a business selling more efficient Hadoop-ishness, but also depressing how infrequently I see them deployed, even for pretty massive clusters.
Unfortunately, it's hard to see a startup focusing on single-machine solutions for GB->TB sets (even scaled out with replicas) as the way many in the space get started is with an open core model and the thing they charge for is the clustering and/or monitoring needed to become a distributed system.
But... I am optimistic we'll see a generational effect over the next 10 years in openness/interest in composable ad-hoc analytics tools, especially with Windows incorporating unix cli components.
I'm only very slightly familiar with their features/value-add and not at all with their pricing. Could the pricing model be particularly unpalatable for some reason?
Not that I expect there has to be a deeper reason beyond simply not caring about cost/efficiency. I've certainly both seen and heard described plenty of Hadoop installations that seemed to have missed the "cheap" point in Google's M-R paper and subsequent Hadoop hardware selection advice from, for example, Hortonworks, or misunderstood what it meant. There may also be some misunderstanding of "commodity" or "industry standard" to mean server hardware of a certain "class" (such as brand name or with redundancy features), even if it conflicts with cheapness.
Some of it may be that the hardware selection advice articles (e.g. Hortonworks, Cloudera) are very old, with excellent general advice, but potentially misleading specific numbers. Even extrapolating from those numbers in a naive way can easily lead to needless expense and/or sub-optimal performance (that time some Xeons had 3, not 2, not 4, memory channels).
The latest article I found in an (admittedly quick) search was https://hadoopoopadoop.com/2015/09/22/hadoop-hardware/ from late 2015, which is still remarkably long ago and is rather verbose.
"Big Data™" is unnecessary shade.
> One especially under-used approach for data processing is using standard shell tools and commands. The benefits of this approach can be massive, since creating a data pipeline out of shell commands means that all the processing steps can be done in parallel. This is basically like having your own Storm cluster on your local machine.
It is entirely unlike having a Storm cluster on your machine, and trying to do your data processing as chained shell commands will rapidly become cumbersome if you try to do actual complex processing.
Yes, I get that the author is trying to point out that simpler tools can work for many cases, but the tone of the article makes it seem like that author is just generally saying that EMR/Hadoop is bad. He does not acknowledge just how weighted the test he did is against Hadoop or give any indication of what the tipping point is where you actually want to start considering something distributed. This paints a really misleading picture for anyone who does not already know a fair amount about these technologies.
But...if you just want to run your own personal search engine say...
Then for Wikipedia/Stackoverflow/Quora size datasets (50GB with with 10GB worth of every kind of index(regex/geo/full text etc) ) you can run real time indexing on live updates with all the advanced search options you see under their "advanced search" page one any random Dell or HP Desktop with about 6-8GB of RAM.
Lots of people do this on Wall Street. People don't get what is possible on desktop cause so much of it has moved to the cloud. It will come back to desktop imho.
There will always be people who need them and use them but that proportion is going to keep decreasing (I'm somewhat sad about this, but the math is hard to argue with).
In the meantime (unless you are dealing with video) most text and image datasets out there that avg Joe needs can easily be stored/processed entirely locally thanks to cheap terabyte drives/multicore chips these days. People just haven't realized there isn't that much useable textual data OR that local computing doesn't require all the overhead of handling millions of requests a second. This is Google problem not an avg Joe problem that is being solved with cloud compute.
There is no dispute with the size or amount of compute available on desktops.
Because almost everyone already uses the model of co-locating the search index and query code on a single computer (both Wikipedia and Stackoverflow use Elastic Search which does this).
They use multiple physical servers because of the number of simultaneous requests they serve.
This has never been the use-case for Hadoop.
I've built Hadoop based infrastructure for redundantly storing multiple PB of unstructured data with Spark on top for analysis. This is completely different to search.
That's very different to the Wall St analyst running desktop analysis in Matlab, or the oil/gas exploration team doing the same thing.
That seems remarkably unlikely if there truly is only the one. Maybe I don't understand what you mean, though. How is that any different from having a single distributed cluster, instead?
> if it goes down
I hear (read) this quite a bit, especially recently, and I'm a bit mystified that it's even brought up in this day and age.
Firstly, having worked with (what is now) commodity hardware for over a decade, I believe that people who haven't grossly over-estimate how often "it goes down", especially today. This overestimation is trotted out as a reason that operating ones own hardware is such a "nightmare" and therefore one must always use cloud or a VPS.
With a "large" enough server, the risk goes up, of course. More DIMMs means more memory can fail, but we're still talking about low single digit percent for all errors (including correctable). IIRC, CPUs have even lower failure rates. Everything else tends to be redundant.
Even then, your workflow isn't dead. It's just missing a DIMM or a CPU, possibly after a reboot (which won't be, if you configured it right).
In many cases, downtime isn't actually caused by hardware but by software (or humans). That's not going to be unique to centralized processing versus distributed.
Also, if the single machine is on a cloud provider, many hardware and utilization issues can be abstracted anyway, for a huge price premium.
edit: if you're running on a cloud, you also may be able to autoscale to deal with spiky usage patterns.
I'd venture to say that "for free" is a bit of a stretch, other than under the "already paid for" definition of "free". Being built on top of a job scheduler means one always has to exert the effort/overhead of dealing with a scheduler, even for the simplest case.
On a single machine, you can decide if you want to add the complexity of something like LSF or not. (I would call a custom single-node solution a strawman in the face of pre-Hadoop schedulers. Batch processing isn't exactly new).
> And without sophisticated job scheduling, you can't really come too close to fully allocating your machine (when you're not watching carefully), because you'll accidentally exhaust resources and break random jobs.
I'm pretty sure I'm still missing something. Since my goal is to gain a better understanding, I'll put forward where I believe the disconnect is (but definitely, anyone, correct me if my assumptions are off):
You're coming from a distributed model, where the assumption is that "resources" are divisible and allocatable in such a way that, for any given job, there's an effective maximum number of nodes it can use (e.g. network-i/o-bound job that would waste any additional CPUs). Any leftover nodes need other jobs allocated to them to reach full utilization.
I, however, am considering that the single-server model assumes that there won't be any obvious bottlenecks or limiting resources, instead prefering to run a single job at one time, as fast as possible. This would completely obviate the need for any complex scheduling and avoids accidental resource exhaustion (which, I suppose, is really only disk or memory).
I don't think it's an obvious decision, and you're absolutely right that people are usually too quick to jump to distributed systems when they really don't need them. But I think there are a lot of arguments for using something like Hadoop even before it's strictly necessary. I think part of the disconnect is that we have different backgrounds, so we both look at different things and think "oh, that's easy" vs "oh, thats like a pain."
And yet I've seen it done. At least they weren't truly external, just external to Hadoop.
> In the former case, we're taking rapid steps towards YARN anyways. I'm sure non-Hadoop, pre-packaged systems exist that can handle this, although I'm not personally familiar with them
There's those goggles again :) Batch schedulers have existed for decades (e.g. PBS started in '91).
> (such as smooth scaling without having to think about hardware),
This neatly embodies what I believe is the primary fallacy in most of the decision making, including the fact that it's often parenthetical (i.e. a throway assumption).
Does anyone really value the "smoothness" of scaling? I'd expect the important feature to be, instead, that the slope of the curve doesn't decrease too fast.
The notion that Hadoop someone frees one from having to think about hardware flies in the face of hardware sizing advice from, Cloudera, Hortonworks, and others that discuss sizing node resources based on workload (mostly i/o and ram) expectations and heterogenous clusters. It does, however, explain my observation, in the wild of clusters built out of nodes that seem undersized in all respects for any workload.
>It's now really easy to spin up a cluster,
It's really easy if it's already there? That borders on the tautological. Or are you talking about an additional cluster in an organization where there already is one?
>and with small scale, costs are not that big a deal.
That's just too broad a generalization, just as "cost is a big deal" would be. Cost is always a factor, just not always the biggest one.
Small scale is often (though not always) associated with limited funds. Doubling, or even tacking on 50% to, the cost could be catastrophic to a seed startup or a scientific researcher.
>With large scale, your system is distributed in some way anyway.
This strikes me as little more than a version of the slippery slope fallacy. Even some web or app servers behind a load balancers could be considered distributed [1], but that doesn't make them a good candidate for anything that's actually considered a distribute framework.
It also hand-waves away the problem that, even if costs weren't a big deal at small scale, they don't somehow magically become less of an issue at large scale. Paying a 50% "tax" on $40k is one thing. At $400k, you could have hired someone to think about hardware for a year, instead.
[1] I just recently pointed out, slightly tongue-in-cheeck, an architectural similarity, one layer down, between app server and database https://news.ycombinator.com/item?id=17521817
> I don't think it's an obvious decision
Certainly, which is why I even bother with discussion like this, in the hopes of making the decision clearer (of not obvious) to me in the future.
A response to my footnoted (in my other comment) comment pointed out how oversimplified my understanding of distributed databases was. Well, I knew it was an oversimplification, but not in which way.
There's plenty of computer science research from the 70s and 80s covering these topics, but they're both tough to translate to practical considerations, and they're woefully out of date (e.g. don't account for SSDs or cheap commodity hardware).
> But I think there are a lot of arguments for using something like Hadoop even before it's strictly necessary.
Well, philisophically, I would disagree with such an assertion on the grounds of premature optimization, absent the "strictly".
I would advocate for switching from scaling "up" (aka "vertically", larger single machines) to scaling "out" (aka "horizontally" or distributed) around the point of cost parity, not at the point it is no longer possible to scale up a single machine (unless that point can reasonably be expected to occur first, I suppose).
> I think part of the disconnect is that we have different backgrounds, so we both look at different things and think "oh, that's easy" vs "oh, thats like a pain."
That would account for any overestimation of how difficult it is to work with hardware or how complex Hadoop is set up, administer, or use. Those are just initial conditions and may well unduly influence decision making that has far longer-lasting consequences.
However, I'd like to think I'm not often guilty of the latter overestimation when discussing solutions (and even advocating single-server), as I tend to assume that it can it least become easy enough for anyone out there, so long as the technology is popular enough (like Hadoop) or traditional/mature enough (like the tools in the original comment, or PBS) that plenty of documentation and/or experts exist.
My background also includes having seen, first-hand, over decades, various attempts at distributed processing and databases in practice, with varying degrees of success. This has included early "universal" filesystems like AFS, "sharding" MySQL to give it "web scale" performance [1], Glustre and its ilk, some NoSQLs, and of course Hadoop.
If anything, I'd say that, with most popular, new technologies, especially ones predicated on performance or scale, "it's a pain" is not the knee-jerk skeptical reaction my experience has ingrained in me. Rather, it's more like "sure, it's easy now, but you'll pay." TANSTAAFL.
[1] This worked well enough but did have a high up-front engineering cost and a high on-going operating cost for the large number of medium-small servers plus larger than otherwise needed app servers to do DB operations that could no longer be done inside the database because each one had incomplete data. Due to effort overestimation, it was unthinkable to move from a VPS to a colo so as to get a medium-large single DB server with enough attached storage to break the "web scale" I/O bottleneck for years to come.
Except that of course nowadays such hardware is just a couple of clicks away in AWS.
It is the foundation technology of almost every data science team around the world. And your misguided post (which for some weird reason only focuses on graph algorithms) doesn't change that. And not sure why you think it's inefficient. We run 30 node, autoscaling clusters which stay close to 100% for most of the time.
I don’t know about your specific workload, but i’ve seen quite a few Hadoop setups that were at 100% load most of the time, and were replaced by relatively simple non Hadoop based code that used 2% to 10% of the hardware and ran about as fast.
I didn’t spend much time evaluating the “pre”, but at least one workload spent 90% of the 100% on [de]serialization.
It’s not my link, it is Frank McSherry who is commenting in this thread - I hope he can chime in on why he chose this specific example - but it correlates very well with my experience.
I have exactly zero knowledge on Spark's efficiency as well as zero on how representative graph algorithms are, but I can confidently say that the above statement fails to refute the referenced article's thesis (which, arguably, criticizes assertions just like that).
Just because your implementation scales (even autoscales) to use more compute resources says nothing about its efficiency (overall or even marginal when adding more nodes, i.e. the shape of the curve).
Computer science has struggled with achieving even near linear-scalability ever since the advent of SMP.
I'm not sure it's fair to summarize any one thing as "the entire point" of Hadoop, but, as I recall, it was, originally, an open source implementation of Google's Map-Reduce paper. Put another way, it was a way to bring Google's compute strategy to the masses.
That said, the notion that there is "commodity, cheap, off the shelf PC hardware" and "exotic specifications" is completely a false dichotomy, especially in the face of what, for example, Google actually does.
Google goes cheap. Very cheap. It's custom and exotic, just optimized for cost, but not absolute cost per nod, rather the ratio of cost for performance.
That last part is what's missing from every single Hadoop installation I've personally seen (or that anyone I know has personally seen), the maximization of performance for cost. Instead, there's an inexplicable desire to increase node count by using cheaper nodes, no matter the performance.
> Except that of course nowadays such hardware is just a couple of clicks away in AWS.
I'm a bit unclear what the "except" means here. I don't believe AWS has the truly high-end specs available (and never has, historically, so we can reasonably assume it never will). It's also very not-cheap.
Joking snark aside, I'm actually doubtful this is true. Specifically, I don't recall the impetus for Hadoop (or Google's original Map-Reduce, as described in the '04 paper) being an all-in-memory workload.
Despite it being repeatedly brought up in this sub-thread, I maintain that it's a niche use case and that disk-based data processing workloads are far more common.
ETA: Does anyone know of a canonical or early/initial document outlining the purpose, or at least design goals, of Hadoop?
In that article, the author runs a regular expression that looks lines that start with "Result" in the pgn files. Then they do a tiny bit of processing with awk.
In the original article[0] the author completely parses the pgn files. Then they extract the result from the parsed file.
I wouldn't be surprised if the difference came completely from that.
I would love to see someone run benchmark the exact same program and see how much of a difference there really is between Hadoop and Bash.
[0] https://web.archive.org/web/20140119221101/http://tomhayden3...
Come to think of it, just benchmarking a single-threaded run of enough lines of each against each other should give you an idea, especially with even basic profiling from time(1) to estimate how much was CPU and how much was I/O.
That said, I don't think that such a discrepancy, if it exists, takes away from what I took to be the author's overall point, which is that old, simple tools, are underused, despite being much faster and good enough, i.e. worse is better [1].
[1] This an example of sacrificing correctness for simplicity in implementation, something detailed in the "rise of worse is better" essay, not merely an excuse for sloppiness in engineering.
366,000 RPS on a laptop: https://adamdrake.com/big-data-small-machine.html
And that's with non-optimized code!
I now hear actual data scientists complain that startups trying to hire them think they have "big data" problems when all they have are "data" problems.