13
Mar
By Peter Pawlowski in Analytics, Blogroll, nPath on March 13, 2009
   

Our goal at Aster is to build a product that will answer your analytical questions sooner. Sooner doesn’t just mean faster database  performance - it means faster answers from the moment you conceive of the question to the moment you get the answer. This means allowing analysts and end-users to easily ask the questions on their mind.

Aster nCluster, our massively-parallel database, has supported SQL from birth. SQL is great in many respects: it allows people of various levels of technical proficiency to ask lots of interesting questions in a relatively straightforward way. SQL’s easy to learn but powerful enough to ask the right questions.

But, we’ve realized that in many situations SQL just doesn’t cut it. If you want to sessionize your web clicks or find interesting user paths, run a custom fraud classifier, or tokenize and stem words across documents, you’re out of luck. Enter SQL/MR, one part of our vision of what a 21st-century database system should look like.

Let’s say your data is in nCluster. If your analytic question can be answered using SQL, you don’t have to worry about writing Java or Python. But, as soon as something more complicated comes up, you can write a SQL/MR function against our simple API, upload it into the cluster, and have it start operating on your data by invoking it from SQL. How is this related to MapReduce? It turns out that these functions are sufficient to express a full MapReduce dataflow. How are SQL/MR functions different than the UDFs of yore? It’s all about scale, usability, reusability; all three contributing to you getting your answer sooner.

Scalability
SQL/MR functions play in a massively-parallel sandbox, one with terabytes and terabytes of data, so they’re designed to be readily parallelized. Yes, they just accept a table as input and produce a table as output, but they do so in a distributed way at huge scale. They can take as input either rows (think “map”) or well-defined partitions (think “reduce”), which allows nCluster to move data and/or computation around to make sure that the right data is on the right node at the right time. SQL/MR functions are table functions breaking out of the single node straight-jacket. This means you can analyze lots of data fast.

Usability
We want to make sure that developers using our SQL/MR framework spend their time thinking about the analytics, not dealing with infrastructure issues. We have a straight-foward API (think: you get a stream of rows and give us back a stream of rows) and a debugging interface that lets you monitor execution of your function across our cluster. Want to write and run a function? One command installs the function, and a single SQL statements invokes it. The data you provide the function is defined in SQL, and the output can be sliced and dices with more SQL - no digging into Java if you want to change a projection, provide the function a different slice of data, or add a sort onto the output. All this allows a developer to get a working function - sooner - and an analyst to tweak the question more readily.

Reusability
We’ve gone to great lengths to make sure that a SQL/MR function, once written, can be leveraged far and wide. As mentioned before, SQL/MR functions are invoked from SQL, which means that they can be used by users who don’t know anything about Java. They also accept “argument clauses” - custom parameters which integrate nicely with SQL. Our functions are polymorphic, which means their output is dynamically determined by their input. This means that they can be used in a variety of contexts. And, it means that any number of people can write a function which you can easily reuse over your data. A function, once written, can be reused all over the place, allowing users to ask their questions faster (since someone’s probably asked a similar question in the past).

In fact, we’ve leveraged the SQL/MR framework to build a function that ships with nCluster: nPath. But this is just the first step, and
the sky’s the limit. SQL/MR could enable functions for market basket analysis, k-means clustering, support vector machines, natural language processing, among others.

How soon will your questions be answered? I’d love to hear of any ideas you have for analytic functions you’re struggling to write in SQL which you think could be a good fit for SQL/MapReduce


Post a comment

Name: 
Email: 
URL: 
Comments: