Data Frames in D - let's not wait for linear algebra; useful today in finance and Internet of Things
Laeeth Isharc via Digitalmars-d-learn
digitalmars-d-learn at puremagic.com
Sat Dec 27 07:33:13 PST 2014
Russell:
"I think we are agreeing. Very lightweight editor and executor of
code
fragments is as good, if not better, that the one line REPL."
Yes - the key for me is that the absence of a shell is by no
means a reason to say that D is not suited to this task. One may
wish to refine what exists, but that is another question entirely.
"Part of the problem here is tribalism. Most data science people
want to
use the same tools that other data science people use, even
though the
issue is to differentiate themselves."
Yes - we are answering two different questions. I could not care
less about persuading anyone en masse in a broad sector, those
who think of themselves as being 'data scientists' included.
It's silly, in my view, to think of it as an established field
very distinct from others, and with a fixed way of doing things.
If for no other reason that things are in flux and the sector is
growing quickly, which means that there is room for many
different approaches, and it is premature to think the popularity
of approach X or Y today means that approach 'D' can't be
productive tomorrow.
But as I said, I am less convinced in persuading anyone, and
rather more concerned with getting a basic data frame in D up and
running because I could certainly use it, and the hard work has
been done already. The basics should be an evening's work for an
advanced D hacker, but it will probably take me longer than that.
In any case, since nobody else has come forward, I will keep
working away at it.
"A BLAS library is certainly a precusor, as is very good data
visualization tools, graphs, diagrams etc."
Perhaps a prerequisite to D being seen as a contender, but I
don't see how it's a prerequisite just to have a dataframe, which
is really a very simple yet incredibly useful thing.
"Go has masses of people putting a lot of effort into Web. It's
not the ideas, it's the number of people getting on board and
doing things".
Also about the quality of the people. (I have no view about Go,
but have a very positive view on D). When things get big there
is a danger they get cluttered. That's one blessing for D.
"To get some traction in any of these areas, finance data
analysis and
model building, or systems activity, it is all about people doing
it,
publicizing it and making things available for others to use".
Yes - so do you have any thoughts on what a data frame structure
should look like? I am trying to do and after that will make
available.
"But it needs to be better than Julia in some way that makes
others sit up and take notice. There has to be the ability to
create
some hype."
Don't care ;) This concept of "what is your edge" is not my cup
of tea because I do not see the world in those terms. Something
of high quality that's highly productive will over time stand a
decent chance of becoming more widely adopted, whereas trying to
force it into some kind of marketing framework can prove
counterproductive.
Right now, the main thing I care about is solving the problem at
hand, because if it solves my problem well then I am pretty sure
it will be useful to others too, and be so better than if one had
adopted a more consciously 'commercial/marketing' mindset.
I would post the dataframe skeleton here, but it's too
embarassing right now and want to read the std.variant library to
see what tricks I can learn. (A data series seems kind of like a
variant, but with every cell the same type). Obviously in some
cases the data frame type is defined at compile time, like a
struct, and that's easy. But if you are loading from a file you
need to be able to have dynamic typing for the column.
"> I don't believe I agree that we need a perfect
multi-dimensional
> rectangular array library to serve as a backend before thinking
> and doing much on data frames (although it will certainly be
> very useful when ready).
Also, if there is a ready made C or C++ library that can be made
use of,
do it."
Well, the hard parts of arrays themselves (and it's not that
fiendishly hard, I would think) seem to need to be tightly
integrated with the language, so I don't see how a C/C++ library
will help so much. For the linear algebra, yes...
hyping it up.
"I recently discovered a number of hedge funds work solely on
moving
average based algorithmic trading. NumPy, SciPy and Pandas all
have
variations on this basic algorithm."
Well, having worked for more or less quanty hedge funds since 98,
I would think it unlikely that anyone depends only on moving
averages although basic old-school trend-following certainly does
work - it is just a hard sell to herding institutional investors,
and does not fit very well with the concept of a 'career'. (You
have to be able to see the five years of subdued returns since
2009 as just part of the cycle, which indeed may be the correct
view when one sees markets as a natural phenomenon, but is not
the view of asset allocators, or talented people one may want to
hire in other areas).
"Perceived to be fast. In fact it isn't anything like as fast as
it
should be. NumPy (which underpins Pandas and provides all the data
structures and basic algorithms), is actually quite slow."
Yes - was tired when I wrote, and meant to say Pandas is fast for
key things such as parsing large data files eg CSVs -
significantly faster than Julia, from what I have seen. And yes
- I agree about Numpy, and don't need to be persuaded of the
benefits of moving to something else if one can make it slightly
less inconvenient. Which is how this conversation started - you
really don't need a perfect BLAS implementation/wrapper to start
to benefit from a dataframe.
"Guido though is I/O bound rather than CPU bound in his work and
doesn't see a need for anything other than multiprocessing for
accessing parallelism in Python. Sadly, it can be shown that
multiprocessing is slow and inefficient at what it does and it
needs replacing".
I cannot claim deep expertise here, but this was one of the
things that got me looking at D originally. Just too frustrating
trying to fit with the restrictions to write nogil Cython code,
knowing that one might need to rewrite when one has mentally long
moved on. Ie I feel like I am short options building my platform
that way, and I don't like being short options when they don't
cost much to buy. Hence D. It also struck me that there was a
degree of complacency amongst some Python people, whereas hunger
and insecurity may be a spur to greater and more creative efforts.
"In principle this is fertile territory for a new language to
take the
stage. Hence Julia." I fear D has missed the boat of this
opportunity
now."
I really don't see why one can't just take the next boat arriving
in fifteen minutes. Or establish a new boat service going
somewhere better that hooks up with the existing network.
Conditions are changing so quickly, and the gap between the talk
about big data etc and what people have actually done so far so
large that to me the field seems wide open. I don't see an
alternative acceptable way to do what I would like, so D it will
be. And if I think that way today, probably others will have the
same thoughts in coming years. (Perhaps not).
"This is worth hyping up, it should be front and centre on teh
dlang
pages along with Facebook funding bug fixes."
I agree. Also in a few lines a punchier summary of why
Sociomantic use D, what the benefits have been, and how they deal
with the standards sorts of hurdles that might have been
objections in a more mature and conventional company ("how are
you going to hire experienced D programmers").
"But if all the libraries are C , C++ and Fortran, is there any
value add
role for D?"
I guess we vote with our feet/fingers. Sounds like you don't
find D especially useful (since you don't use it much currently),
whereas I do. De gustibus non est disputandum, particularly when
tastes reflect being in different situations.
"Lots of C++ system embed Python or Lua for dynamic scripting
capability,
lots of Python and R system call out to C. This seems a well
established
milieu. Is there a good way for D to, in an evolutionary way
establish a
permanent foothold. Certainly it cannot be a revolutionary one."
You write as if Christensen's book "The Innovator's Dilemma" had
never been written, and nor had it been a standard textbook in
business schools for some years. You may have good arguments as
to why he is wrong, or why it doesn't apply to D, but you haven't
set them out, as far as I am aware.
Not Russell
"There will sure be some algorithms where numba/cython would do
better (especially if they cannot be easily vectorized), but
that's not the point. The thing about numpy is that it provides a
unified accepted interface (plus a reasonable set of reasonably
fast tools and algorithms) for arrays and buffers for a multitude
of scientific libraries (scipy, pytables, h5py, pandas, scikit-*,
just to name a few), which then makes it much easier to use them
together and write your own ones."
Yes. But one has to start somewhere (if not happy with the
python route), and we start to have equivalents of
scipy,pytables/h5py. So why not pandas?
"Splunk stuff is just an example of using dataflow networks for
processing data rather than using SQL. The "Big Data using JVM"
community are already on this road, cf. various proprietary
frameworks
running over Hadoop and Spark."
Yes - technically, it may well be "nothing more than". But many
of the practical problems which have a high commercial return to
solving are "nothing more than" quite simple things technically.
One doesn't need to be a technical genius to make valuable
commercial contributions. And maybe Hadoop and Spark are just
the perfect solution for most people (maybe not!), but that
certainly leaves some room for others.
So... data frames!?
More information about the Digitalmars-d-learn
mailing list