[
https://issues.apache.org/jira/browse/MAHOUT-1507?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13964285#comment-13964285
]
Pat Ferrel edited comment on MAHOUT-1507 at 4/9/14 3:34 PM:
------------------------------------------------------------
Dictionary and Mahout Input Creation:
The solr-recommender example does a single machine (single threaded) read
through of the input data. It creates an in-memory BiHashMap to translate
external Ids to Mahout Ids and writes these to HDFS--one for rows, one for
columns. It then does a mr read through of all the data instantiating the
hasmap once on each node. This has two obvious scalability bottlenecks; the
creation of the dictionary and the need for it to be in memory.
To make this more scalable it could be done by doing a mr 'unique' but I think
that requires a single reducer anyway and would require an incrementing number
saved in HDFS for jobs to coordinate the creation of a sequential Mahout Id
from. This number would be need to be accessed often and so would be heavily IO
bound during dictionary creation. The creation of Mahout input could be done
with a join on the External Id field of the input and the appropriate
dictionaries. It would be done on every preference since we are talking about
the recommender.
To create output:
The solr-recommender does and mr job across the cluster to translate the data
from Mahout sequence files with Mahout Ids into text files with external Ids.
It creates several CSV files for Solr input. Every node machine reads both
entire dictionaries into memory once for sharing between tasks on the node.
Here I imagine another 'join' of the output Mahout Id field on both
dictionaries. Since I'm new to Spark but see it has a join operation I'll be
looking there first.
Is it worth doing this without an in-memory hashmap dictionary?
was (Author: pferrel):
Dictionary and Mahout Input Creation:
The solr-recommender example does a single machine (single threaded) read
through of the input data. It creates an in-memory BiHashMap to translate
external Ids to Mahout Ids and writes these to HDFS--one for row, one for
columns. It then does a mr read through of all the data instantiating the
hasmap once on each node. This has two obvious scalability bottlenecks; the
creation of the dictionary and the need for it to be in memory.
To make this more scalable it could be done by doing a mr 'unique' but I think
that requires a single reducer anyway and would require an incrementing number
saved in HDFS for jobs to coordinate the creation of a sequential Mahout Id
from. This number would be need to be accessed often and so would be heavily IO
bound during dictionary creation. The creation of Mahout input could be done
with a join on the External Id field of the input and the appropriate
dictionaries. It would be done on every preference since we are talking about
the recommender.
To create output:
The solr-recommender does and mr job across the cluster to translate the data
from Mahout sequence files with Mahout Ids into text files with external Ids.
It creates several CSV files for Solr input. Every node machine reads both
entire dictionaries into memory once for sharing between tasks on the node.
Here I imagine another 'join' of the output Mahout Id field on both
dictionaries. Since I'm new to Spark but see it has a join operation I'll be
looking there first.
Is it worth doing this without an in-memory hashmap dictionary?
> Support input and output using user defined ID wherever possible
> ----------------------------------------------------------------
>
> Key: MAHOUT-1507
> URL: https://issues.apache.org/jira/browse/MAHOUT-1507
> Project: Mahout
> Issue Type: Bug
> Components: Math
> Affects Versions: 0.9
> Environment: Spark Scala, Mahout v2
> Reporter: Pat Ferrel
> Labels: spark
> Fix For: 1.0
>
>
> All users of Mahout have data which is addressed by keys or IDs of their own
> devise. In order to use much of Mahout they must translate these IDs into
> Mahout IDs, then run their jobs and translate back again when retrieving the
> output. If the ID space is very large this is a difficult problem for users
> to solve at scale.
> For many Mahout operations this would not be necessary if these external keys
> could be maintained for vectors and dimensions, or for rows and columns of a
> DRM.
> The reason I bring this up now is that much groundwork is being laid for
> Mahout's future on Spark so getting this notion in early could be
> fundamentally important and used to build on.
> If external IDs for rows and columns were maintained then RSJ, DRM Transpose
> (and other DRM ops), vector extraction, clustering, and recommenders would
> need no ID translation steps, a big user win.
> A partial solution might be to support external row IDs alone somewhat like
> the NamedVector and PropertyVector in the Mahout hadoop code.
> On Apr 3, 2014, at 11:00 AM, Pat Ferrel <[email protected]> wrote:
> Perhaps this is best phrased as a feature request.
> On Apr 2, 2014, at 2:55 PM, Dmitriy Lyubimov <[email protected]> wrote:
> PS.
> sequence file keys have also special meaning if they are Ints. .E.g. A'
> physical operator requires keys to be ints, in which case it interprets
> them as row indexes that become column indexes. This of course isn't always
> the case, e.g. (Aexpr).t %*% Aexpr doesn't require int indices because in
> reality optimizer will never choose actual transposition as a physical step
> in such pipeline. This interpretation is consistent with interpretation of
> long-existing Hadoop-side DistributedRowMatrix#transpose.
> On Wed, Apr 2, 2014 at 2:45 PM, Dmitriy Lyubimov <[email protected]> wrote:
> On Wed, Apr 2, 2014 at 1:56 PM, Pat Ferrel <[email protected]> wrote:
> On Apr 2, 2014, at 1:39 PM, Dmitriy Lyubimov <[email protected]> wrote:
> I think this duality, names and keys, is not very healthy really, and
> just
> creates addtutiinal hassle. Spark drm takes care of keys automatically
> thoughout, but propagating names from name vectors is solely algorithm
> concern as it stands.
> Not sure what you mean.
> Not what you think, it looks like.
> I mean that Mahout DRM structure is a bag of (key -> Vector) pairs. When
> persisted, key goes to the key of a sequence file. In particular, it means
> that there is a case of Bag[ key -> NamedVector]. Which means, external
> anchor could be saved to either key or name of a row. In practice it causes
> compatibility mess, e.g. we saw those numerous cases where e.g. seq2sparse
> saves external keys (file paths) into key, whereas e.g. clustering
> algorithms are not seeing them because they expect them to be the name part
> of the vector. I am just saying we have two ways to name the rows, and it
> is generally not a healthy choice for the aforementioned reason.
> In my experience Names and Properties are primarily used to store
> external keys, which are quite healthy.
> Users never have data with Mahout keys, they must constantly go back and
> forth. This is exactly what the R data frame does, no? I'm not so concerned
> with being able to address an element by the external key
> drmB["pat"]["iPad'] like a HashMap. But it would sure be nice to have the
> external ids follow the data through any calculation that makes sense.
> I am with you on this.
> This would mean clustering, recommendations, transpose, RSJ would require
> no id transforming steps. This would make dealing with Mahout much easier.
> Data frames is a little bit a different thing, right now we work just with
> matrices. Although, yes, our in-core matrices support row and column names
> (just like in R) and distributed matrices support row keys only. what i
> mean is that algebraic expression e.g.
> Aexpr %*% Bexpr will automatically propagate _keys_ from Aexpr as implied
> above, but not necessarily named vectors, because internally algorithms
> blockify things into matrix blocks, and i am far from sure that Mahout
> in-core stuff works correctly with named vectors as part of a matrix block
> in all situations. I may be wrong. I always relied on sequence file keys to
> identify data points.
> Note that sequence file keys are bigger than just a name, it is anything
> Writable. I.e. you could save a data structure there, as long as you have a
> Writable for it.
> On Apr 2, 2014 1:08 PM, "Pat Ferrel" <[email protected]> wrote:
> Are the Spark efforts supporting all Mahout Vector types? Named,
> Property
> Vectors? It occurred to me that data frames in R is a related but more
> general solution. If all rows and columns of a DRM and their
> coresponding
> Vectors (row or column vectors) were to support arbitrary properties
> attached to them in such a way that they are preserved during
> transpose,
> Vector extraction, and any other operations that make sense there
> would be
> a huge benefit for users.
> One of the constant problems with input to Mahout is translation of
> IDs.
> External to Mahout going in, Mahout to external coming out. Most of
> this
> would be unneeded if Mahout supported data frames, some would be
> avoided by
> supporting named or property vectors universally.
--
This message was sent by Atlassian JIRA
(v6.2#6252)
