Comments

1) "Pig Latin is equivalent to the computational power of SQL" Pig

Latin is different than SQL in that it is a dataflow language, where

as SQL is a query language (for more on this see

http://feeds.developer.yahoo.net/~r/YDNHadoop/~3/i2Cdha8kedw/comparing_pig_latin_and_sql_fo.html ) So in your chart

I would put Pig Latin across from relational algebra. The translation

from Pig Latin to relational algebra is nearly 1-1.

2) MR should not be across from Relational Algebra. MR does not

itself present a relational algebra. It is possible to implement

relational algebras over MR, as Pig, Hive, and JAQL all do. Evaluated

just as MR I would argue that it is not relationally complete because

it presents only grouping and sorting options. It is straight forward

to implement projection and filtering via the Map function, but this

is not natively provided. Joins can be implemented via Reduce, but

doing so is not trivial. At the same time, since MR is written in

Java, which is definitely Turing complete, then MR must also be

Turning complete.

3) I'm not sure where MR goes in your chart, but the bottom seems the

wrong spot. By definition, anything you can implement in JAQL, Pig,

or Hive you can implement in MR. So to say it is the least

computationally powerful seems wrong. Perhaps I don't understand what

is meant by Computational Power.

To sum up, it seems to me that you're trying to map three separate

dimensions (relational completeness, turing completeness, and SQLness)

all onto one dimension. Instead it would make more sense to me to

present each project relative to these three dimensions: JAQL (yes,

yes, no), Pig (yes, no, no), Hive (yes, no, yes), and MR (no, yes, no).

Alan.

On Mar 20, 2010, at 4:54 PM, Rob Stewart wrote:Hi,

I have a question, in a remark that Alan Gates made a few months ago

on

these mailing lists regarding the computability and expressibility

of Pig,

Hive, and the MapReduce model.

In particular, it was a question regarding the Turing computability

of each.

Is anyone able to remark on my discussion of this in my forthcoming

paper

(this is just a small extract). I am pretty confident as to where I

have put

Pig and Hive, quite confident with JAQL (I've checked it out with

the JAQL

dev's), and far less sure of my assessment of the MapReduce model

(only

relationally complete??)

Find it here:

http://www.macs.hw.ac.uk/~rs46/dropbox/computability.pdfI would greatly appreciate any comments of pointers to any

inaccuracies.

thanks,

Rob Stewart