Guys, thank you very much for your feedback.
I have already my own vanilla spark-based implementation of row
similarity that reads and writes into NoSQL (in my case HBase).
My intention is to profit from your effort to abstract algebraic layer
from physical backend because I find it a great idea.
There is no question that the effort to implement i/o with some NoSql
and spark is very low nowadays.
My question is more towards understanding your design.
In particular, why for instance
org.apache.mahout.math.drm.DistributedEngine has def drmFromHDFS()?
I do understand argument with "files is most basic and common" and "we
had this already in mahout 0.6 so its for compatibility purposes", but
why for instance instead of drmFromHDFS() there is no def createDRM()
and then some particular implementation of DistributedEngine (or
medium-specific helper) that then decides how DRM shall be created?
Admittedly, I do NOT understand your design fully just yet and I am
asking these questions not to criticize this design but to help me
understand it.
Another example is existance of org.apache.mahout.drivers.Schema. It
seems that there is effort to kind of make medium-specific format
flexible and abstract it away, but again the limitation is it is
file-centric.
Thank you for your hints with drmWrap and IndexedDataset. With this in
mind, maybe my error is that I am trying to reuse classes in
org.apache.mahout.drivers, maybe I should just write my own driver from
scratch and with Database in mind.
Thank you again for your hints and ideas
reinis
On 10.10.2014 01:00, Pat Ferrel wrote:
There is also the mahout Reader and Writer traits and classes that currently
work with text delimited file I/O. These were imagined as a general framework
to support parallelized read/write to any format and store using whatever
method is expedient, including the ones Dmitriy mentions. I personally would
like to do MongoDB since I have an existing app using that.
These are built to support a sort of extended-DRM (IndexedDataset) which
maintains external IDs. These IDs can be anything you can put in a string like
Mongo or Cassandra keys or can be left as human readable external keys. From an
IndexedDataset you can get a CheckpointedDRM and do anything in the DSL with it.
They are in the spark module but the base traits have been moved to the core
“math-scala” to make the concepts core with implementations in left in the
engine specific modules. This is work about to be put in a PR but you can look
at it in the master to see if it helps—expect some refactoring shortly.
I’m sure there will be changes needed for DBs but haven’t gotten to that so
would love another set of eyes on the code.
On Oct 9, 2014, at 3:08 PM, Dmitriy Lyubimov<[email protected]> wrote:
Bottom line, some very smart people decided to do all that work in Spark
and give us for free. Not sure why, but that did. If the capability already
found in Spark, there's no need for us to replicate it.
WRT specifically NoSql, Spark can read HBase trivially. I also did a bit
more advanced things with a custom rdd implementation in Spark that was
able to stream coprocessor outputs into rdd functors. In either case this
is actually a fairly small effort. I never looked at it closely, but i know
there are also Cassandra adapters for spark as well. Chances are, you
could probably load data from any thinkable distributed data store into
Spark these days via off the shelf implementations. If not, Spark actually
makes it very easy to come with one on your own.
On Thu, Oct 9, 2014 at 2:47 PM, Dmitriy Lyubimov<[email protected]> wrote:
Matrix defines structure. Not necessarily where it can be imported from.
You're right in the sense that framework itself avoids defining apis for
custom partition formation. But you're wrong in implying you cannot do it
if you wanted, our that you d have to do anything that complex as you say.
As long as you can form your own rdd of keys and row vectors, you can
always wrap it into a matrix (drmWrap api). Hdfs drm persistence on the
other hand had been around for as long as I remember, not just in 1.0. So
naturally those are provided to be interoperable with mahout .9 and before,
e g be able to load output from stuff like seq2sparse and such.
Note that if you instruct your backend to use some sort off data locality
information, it will also be able capitalize on that automatically.
There is actually far greater number of concerns of interacting with
native engine capabilities than just reading the data. For example, what if
we wanted to wrap an output of a shark query into a matrix. Instead of
addressing all those individually, we just chose to delegate those to
actual capabilities of backend. Chances are they already have (and, in
fact, do in case of spark) all that tooling far better than we will ever
have on our own.
Sent from my phone.
On Oct 9, 2014 12:56 PM, "Reinis Vicups"<[email protected]> wrote:
Hello,
I am currently looking into the new (DRM) mahout framework.
I find myself wondering why is it so that from one side there is a lot
of thought, effort and design complexity being invested into abstracting
engines, contexts or algebraic operations,
but from the other side, even abstract interfaces, are defined in a way
that everything has to be read or written from files (on HDFS).
I am considering to implement reading/writing to NoSQL database and
initially I assumed it will be enough just to implement own
ReaderWriter, but I am currently realizing that I will have to
re-implement or hack-around by derivating own versions of large(?)
portions of framework including own variant of CheckpointedDrm,
DistributedEngine and what not.
Is it because abstracting away storage type would introduce even more
complexity or because there are aspects of design that absolutely
require to read/write only to (seq)files?
kind regards
reinis
