1.1B Taxi Rides on Kdb+/q and 4 Xeon Phi CPUs(tech.marksblogg.com) |
1.1B Taxi Rides on Kdb+/q and 4 Xeon Phi CPUs(tech.marksblogg.com) |
If you like writing dense, clever regexs (which I do) then you'll love k & q. The amount that you can get done with just a few characters is unparalleled.
Which leads to, IMHO, their main drawback: k/q (like clever regexes) are often write-only code. Picking up another's codebase or even your own after some time has passed can be very hard/impossible because of how mindbendingly dense with logic the code it. Even if they were the best choice for a given domain, I'd try to steer clear of using them for anything other then exploratory work that doesn't need to be maintained.
The biggest messes I have to clean up come less from "clever" code than they do from people who try to program in K as if it were some other language. For example, somebody fond of for loops might write the following to apply a function 'f' to pairings of values and their index in a list:
result:()
i:0
do[#v
r:r,,f[v[i];i]
i:i+1
]
Ugly, complicated, but "close at hand". There is of course a much nicer and more idiomatic way to do the same thing: result: f'[v;!#v]
Most of the time conciseness isn't the goal, but you get it as a side effect of writing good code and working "with the grain" of the language. > their main drawback: k/q (like clever regexes) are often write-only code
This depends on the code reader's mentality. One line of k (or q or j or apl) would do what 10 lines of verbose languages do. For verbose languages, your expectation is spending 1 minute for 10 lines of code to fully understand it; but for terse languages, you need to change your expectation to spending 1 minute for only 1 line of code. You are not going to understand anything if you still want to spend 6 seconds per line.On the other hand, proficiency is important. You read English articles in a slow pace in your first grade; you are not going to be a speed reader without practice, even if English is the only language you speak. No one would expect you to speed read Japanese right after you can compose simple Japanese sentences.
For me it is the opposite. I like not having to type any more than necessary. I like writing as little code as possible.
I also like being able to read through a program listing without having to wade through page upon page of long, camel case function names, and trying to follow five levels of indentation.
Not because I think that is the "wrong" approach for everyone but because I personally struggle with overcoming language and documentation verbosity in order to make any progress. It is the wrong approach for me. I wonder if perhaps this is how it feels for the typical programmer, in the OP's comment, who might struggle with a lack of verbosity.
Sometimes I streamedit (search and replace) blocks of code to remove indentation and shorten long function names to two letter labels, just so I can read without distraction.
What are the chances of finding someone who has the same preference for terseness and who is as skilled as Arthur Whitney?
q.k is about 240 lines in the version I am using. Most of the functions in .q, .Q and the other namespaces fit on a single line. This is a thing of beauty, IMO. For me, this makes studying the language manageable.
Someone in this thread posted a fizzbuzz solution last week that I thought illustrated how one can "work outward", incrementally adding primitives to the left or right to keep modifying output until the desired result is reached.
I use kdb+ daily but I'm a slow learner and still not very good with composing programs from scratch in k.
What little I have written was also done by "working outward", so the fizzbuzz example gave me some hope maybe I'm not too far off track. Thank you for that example. I hope we see some more.
I am thankful k exists, even if the intepreter is not open source and there's no BSD port.
As someone else commented somewhere, perhaps the most awkward aspect of k is getting it to interface with anything else. I think this is what keeps me from using it more frequently for more daily tasks.
But this may be more of a problem with everything else, and not necessarily with k.
I wish a more competent C programmer than I would write a lightweight replacement using linenoise or libedit to substitute for "rlwrap".
I will concede that there is a culture of trying to be a bit too clever on the k4 mailing list but it's perfectly possible to write maintainable code in kdb+
I'm guessing that best way to address this issue is through liberal use of explanatory comments.
The "characters to clever/unmaintainable" ratios you achieve with k (which most kdb+ platforms end up reaching for at some point) are almost unparalleled. A famous example of how awesome/powerful/ridiculous k can be is the "4 lines of K" text editor: http://www.kparc.com/$/edit.k
I guess my point is that letting your guard down, for even a single line, can be orders of magnitude more trouble than it would be in most languages and for that reason I wouldn't base a new stack on it. Also, the developers for it are rare-ish and expensive + it has a serious learning curve for those unfamiliar with FP or Lisp.
Yes Q is dense but it can be annotated with comments.
I also have to maintain some java code written by this lot and it's nearly as unintelligible.
Reasons:
- Even in a OLAP database you end up with quite a few places that have very branchy code. Research on GPU friendly algorithms on things like (complex) JOINS and GROUP BY is pretty new. Additionally complex queries will functions and operations that you might not have a good GPU implementation for (like regex matching)
- Compression. You can use input data that compressed in anyway that there is a x86_64 library for. So you can now use LZ4, ZHUFF, GZIP, XZ. You can have 70+ independent threads decompressing input data (it's OLAP so it's pre-partitioned anyways). (Technically branching, again)
- Indexing techniques that cannot efficient implemented on the GPU can be used again. (Again branching)
- If you handle your own processing scheduling well, you will end up with near optimal IO / memory pattern (make sure to schedule the work on the core with local memory) and you not bound PCIe speed of the GPU. With enough PCIe lanes and lots of SSD drives you process as near memory speeds (esp. when we'll have Xpoint memory)
So the bottom line is if can intelligently farm out work in correct size chunks (it's OLAP so it's prob partitioned anyways) the the Phi is fantastic processor.
I'm primarily talking about the bootable package with Omni-Path interconnect (for multiple).
Lots of people complain about the conciseness of the language and that it is "write-once" code. I tend to disagree. While it might take a while to understand code you didn't write (or even code you wrote a while ago), focusing on writing in q rather than the terser k can improve readability tremendously.
My only wish is that someone would write a free/open-source 64-bit interpreter for q - with similar performance and speed to the closed version. Kona (for k) gets close https://github.com/kevinlawler/kona
- https://github.com/johnearnest/ok
It seems more like a kdb+ competitor than open-source alternative, and isn't using q.
it's a really beautiful little language once you get into it :-)
First time I noticed (mention of) recap at http://tech.marksblogg.com/benchmarks.html
But seriously, what a wonderful world it would be if all papers were this well written.
Of course, not quite, and that's discounting the (small) sync overhead but still, no need to shell out 4 big servers, overpriced phi chips and fancy wide bus memory.
The hardware doesn't seem consistent across different benchmarks. He says it's fast for a "cpu system", but for practical purposes Phi competes more with GPGPUs.
Would this be just as fast with one redis system with 512GB ram? I don't know too many apples to oranges here.
But what useful conclusions can be drawn from it?
Sort of meta, but Mark's job seems awesome. Gets all these toys and writes about configuring them. (The actual configuring is probably a pain but still)
% cat startmaster.q
k).Q.p:{$[~#.Q.D;.Q.p2[x;`:.]':y;(,/(,/.Q.p2[x]'/':)':(#.z.pd;0N)#.Q.P[i](;)'y)@<,/
Looks like line noise... :DI see that as being something I'd very much like.
[0]: https://en.wikipedia.org/wiki/GDDR5_SDRAM#Commercial_impleme...
Unlike typical computer memory architectures, where the memory bus connects multiple chips or modules to one controller, GDDR doesn't do that; every slice of the memory controller only speaks with a single chip, strictly point-to-point. (Reducing bus load and layout issues and thus allowing higher clock rates).
That's why, with GPUs, it's usually sufficient to say how wide the bus is (often 64 - 128 - 256 - 384 - 512 bits) to get a rough idea of it's performance, since memory clock frequencies occupy a rather narrow range. (However, narrow-bus, lower-end GPUs often don't use the same technology as higher-end GPUs, eg. DDR3 instead of GDDR5)
I love J compared with K, but that is because I found it first, and the differences between J and K are minimal, but a different enough to keep me using J.
https://www.walmart.com/ip/INTEL-SERVER-CPU-SC7120P-XEON-PHI...
http://tech.marksblogg.com/billion-nyc-taxi-rides-redshift.h...
Here's hoping some combo of Apache Arrow (also cache aware, much more language stack flexibilty), Aerospike (lua built in), Impala, and others, can finally take on this overpriced product, which has had a lack of serious competitors for 20 years, owing to its (price inelastic) finance client base.
kdb+ is available for raspberry pi, is that cross platform enough?
https://kx.com/2016/06/08/kx-releases-raspberry-pi-build-wit...
edit:
$ file q/l64/q
q/l64/q: ELF 64-bit LSB executable, x86-64, version 1 (SYSV), dynamically linked (uses shared libs), for GNU/Linux 2.6.18,This was my experience. I got a nice email from Eric Iverson in response along with the activation key.
Can you learn it with just the free version or realistically does it require working with it in a professional setting and learning from others who already use it.
Are there any sources you recommend for a someone just starting out?
Thanks.
http://code.kx.com/wiki/Cookbook
This includes a tutorial with a smartmeter concept which has several queries of increasing complexity:
http://code.kx.com/wsvn/code/cookbook_code/tutorial/
Step through this, look at the function definitions, cross-ref with the kdb reference page. Also read Q for mortals on their site:
http://code.kx.com/wiki/Tutorials
Yes it's not as accessible as other langs with loads of books, but on the plus side it's coherent, unlike c++.
The author has an overview of the benchmarks with various systems at http://tech.marksblogg.com/benchmarks.html – but one can't just compare rows as many factors vary (esp. hardware).
Generally with the PCIe one you're running something like OpenCL and with system CPU package you run threads and processes like you normally would.
Technically you could run software directly on the old addin cards since they boot to Linux but you had handle the distribution, running and communication of your software with the host. (you could run any x86_64 binary)
Honestly they make it pretty hard to hack with as a device. It could succeed as a device if they let devs easily create cool applications with it.
There is no reason not to use camel cased variable names and indent functions, if/else blocks etc, and when written this way the code can be perfectly legible even to non q programmers.
Something else that leads people into the write-only trap is that the usual way of working with the language is to use the REPL loop while working, where you can tend to be doing multiple things on 1 line. It's just laziness not to reformat and clean up the code afterwards though.
There is: it makes the program bigger. Program source code length is significant, and if you have more lines, you have more opportunities for bugs.
Long variable names may take more characters, but there is no way they increase bugs.
Indenting(!) doesn't even take many more characters if you use tabs...
Compiler errors do. But you catch those first run.
If tab vs 2 spaces vs 3 spaces can induce a logic error that is a purposely obtuse language.
Or do those single character tokens have special meanings in k? Which would be even worse.
It can seem that way, but there is value in dense programs: They take up less room on the screen. This means you can see where you repeat yourself, and can often find bugs by just reading your code.
I can't count the number of times I scrolled over some code and missed a bug.
A few examples of some useful conclusions:
- Just because a relatively well-optimized PostgreSQL database on a regular workstation takes 5 minutes to run a query doesn't mean you can't get special hardware to run that query faster than you can type.
- Spark + S3 + Amazon Elastic Map Reduce look like an ideal tool to work with large data, but they're pretty slow compared to better tools, and even compared to plain PostgreSQL.
- HDFS really is a lot faster than S3.
- Performance of an Xeon Phi 64-core CPU is within an order of magnitude to an NVidia Titan X.
- Loading 104 GB of compressed data into Q/kdb+ expands to 125 GB with and takes about 30 minutes, but on Redshift expands to 2 TB and takes many hours to upload on a normal connection, plus 4 hours to actually import!
- It might cost $5000 to custom-build a GPU-based supercomputer that can do these queries in under a second, but you can run similar queries if you're willing to wait for 5 minutes each by spinning up instances for a few dollars an hour plus a few more dollars an hour for storage, or by just running PostgreSQL on your workstation.
Also, not a conclusion, but it's incredibly useful to have a simple example exactly how to configure the tool and import some CSV data
Just because a relatively well-optimized PostgreSQL database on a regular workstation takes 5 minutes to run a query doesn't mean you can't get special hardware to run that query faster than you can type.
Already well established for years with systems like redis, and more recently with gpu databases, and other techniques posted on HN regularly.
Spark + S3 + Amazon Elastic Map Reduce...is pretty slow compared to better tools, and even compared to plain PostgreSQL.
Not valid because it doesn't generalize. It so much depends on type of work being done, system architecture, etc, that you can only say it may or may not be true.
HDFS really is a lot faster than S3.
This is already well established, Amazon states aa much right in the docs: http://docs.aws.amazon.com/emr/latest/ManagementGuide/emr-pl...
Performance of an Xeon Phi 64-core CPU is within an order of magnitude to an NVidia Titan X.
Not precise enough to matter because getting within 10x difference is not close to being competitive.*
Loading 104 GB of compressed data into Q/kdb+ expands to 125 GB with and takes about 30 minutes, but on Redshift expands to 2 TB and takes many hours to upload on a normal connection, plus 4 hours to actually import!
I don't see how it's possible for 104GB of csv text data to decompress into only 125GB. For cvs to compress only ~20%...doesn't make sense.
It might cost $5000 to custom-build a GPU-based supercomputer that can do these queries in under a second
No, two problems here. The hardware in question could have used 1 cheap CPU instead of two expensive Xeons and been much less expensive. Bigger problem: The MapD software itself will be $50,000.
> I don't see how it's possible for 104GB of csv text data to decompress into only 125GB. For cvs to compress only ~20%...doesn't make sense.
The CSV file itself is around 500 GB. The internal representation, which might use binary formats for numbers, or compress text, uses 125 GB. Redshift expands it to 2TB for all the indexing and mapping.
> Bigger problem: The MapD software itself will be $50,000.
Ouch. That's a rather large oversight. Is the author affiliated with MapD, perhaps?
This is actually my biggest complaint with the article. He used cstore_fdw with Postgres, which doesn't allow much real indexing, and as far as I can tell (knowing only a little bit about it) he didn't really use any of the benefits of cstore_fdw.
I'd be interested to see how plain Postgres, possibly on a compressed filesystem, with properly-indexed tables stacks up.
Bugs do hide in variable names and whitespace.
thisIsALongVariableName = True
thisIsaLongVariableName = True
is easy to spot when they're side-by-side, but not so much when they're two pages away from each other. Python's tabs and spaces can confuse the indention in cases so that code that "looks" aligned actually isn't.If you haven't run into a problem caused by a typo, you haven't been programming long enough.
Typos affect everyone. It is not a function of the language that causes you to mistype. It is that the program is long and you cannot see the whole thing that causes you to miss it.
> Is it the case of K, Q or whatever its name is?
{foo}[] generates an error because foo is unbound. What's the difference between {foox} and {fooy}? They both might be variables, and yet it is only by reading the program that you can tell which one is correct.
No, it's not supposed to be a joke.
Steve McConnell 1993 observed density is proportional to source program length, so this should be obvious to every programmer: If a small program (as measured in source code bytes) is more likely to be a correct program, then this follows.
A major issue with discussing programming is the sheer number of people who believe they know how to program, when any non-programmer could see quite obviously that they don't: A professional bridge-builder doesn't often fail to build a bridge, but a professional CMS programmer seems to unable to get much past hello world without a bug or two; Less code will therefore produce less bugs.
There is a summary of his advice here[1], but just to highlight
Describe everything the routine does And we mean literally everything. If that makes the name ridiculously long, the name isn't the problem. Your routine is.
And
Make names as long as necessary According to McConnell, the optimum name length for a variable is 9 to 15 characters; routines tend to be more complex and therefore deserve longer names. Make your names as long as they need to be in order to make them understandable.
I've read McConnell, and your claims are so completely the opposite of what he recommends that I'm still unconvinced you aren't trolling.
[1] https://blog.codinghorror.com/i-shall-call-it-somethingmanag...
"A study from Basili and Perricone found that routine size was inversely correlated with errors: as the size of routines increased (up to 200 lines of code), the number of errors per line decreased (Basili and Perricone 1984).
And the conclusion (on page 174): …None of the studies that reported decreased cost, decreased error rates, or both with larger routines distinguished among sizes larger than 200 lines, and you're bound to run into an upper limit of understandability as you pass 200 lines of code
With that in mind, consider that KDB's SQL92 interface is 35 lines (the parser is 14 lines). It may be an extreme example of what McConnell is observing, and yet was himself unable to learn from.
> I'm still unconvinced you aren't trolling.
Look at it this way: Here is software that is faster than what you can write (or maybe you want to write your own taxi problem implementation), and if you don't try hard, you will miss out in finding out how to do something that you can't do.
That downvote button is so easy, that internet person is just a troll, I've been programming for ages, so of course I know what I'm talking about, but if you look through my comment history, you might find it easier to convince yourself at least that I believe that program length matters and permit yourself a discussion about it. You might learn something.
* https://news.ycombinator.com/item?id=8476294
* https://news.ycombinator.com/item?id=8477064
* https://news.ycombinator.com/item?id=10872209
and so on.
Edit, see http://mathoverflow.net/questions/8295/origins-of-mathematic... for the confusion this causes..
- you need to write same variable name on each step of derivation of a formula
- paper and blackboards don't have autocomplete
- juxtaposition is used for multiplication, thus SwapBlue can be Swap times Blue or a single variable called SwapBlue, spacing for multiplication is not a good solution when writing to board/paper
Mathematics notation is so notoriously awful that mathematicians often can't even read equations from other disciplines without significant extra context.
In any language with mandatory variable declaration the code wouldn't even compile. And any decent IDE would even flag this while writing the code.
Nonsense. memmove and memcpy are both defined, and if you choose the wrong one your program compiles and runs, but it simply produces the wrong result.
> And any decent IDE would even flag this while writing the code.
There is no IDE that will flag this. Advanced whole-program static analysis has a hard time with this problem and I'm not aware of any general solutions.
I find it difficult to agree that using non-representative names for variables or functions improves understadability.
Notably, using something like x to represent a meaningful value means the brain has to hold the mapping between the two, which will decrease the number of useful pieces of information kept in short term memory[1].
The brain doesn't keep track of the number of characters in a variable name.
[1] http://www.psych.utoronto.ca/users/peterson/psy430s2001/Mill...
The issue, for me, is not readability but context. Even a word or two (e.g. cuts, begins, dims) helps greatly to establish the context of the code. Like having a map before hiking.
A comment can clearly provide that context without making the variable names long.
[1]: https://news.ycombinator.com/item?id=8747314
> The brain doesn't keep track of the number of characters in a variable name.
Surely you must appreciate that if we use too many characters the window will scroll?
as your opponent has mentioned, context is the key. you operate with objects in your brain, and the faster the transition from code element to the brain object, the better you understand the code. long names make this lookup unnecessarily difficult.