That's indeed the general intent! There hasn't been any formal work
done; given its current operators, the algebra is basically in the class
of what one might loosely call "nested relational/schema optional
languages". In terms of computational capabilities, such languages are
probably in the same class as the relational algebra extended with
aggregation and grouping and ordering - again, roughly speaking.
There's support for ordered as well as unordered nested collections, so
that might allow certain kinds of queries that would be at least very
awkward for bag of records data models and languages. Not sure. (I
haven't followed what the ACM PODS community may have done regarding
query language capabilities since the very early days, I fear.) In
terms of things like Pregelix or graph languages, what's missing for
sure at the Algebricks level is the power offered by
recursion/transitive closure. Once upon a time we were thinking about
going there, but haven't gotten back to that yet. (And have no
immediate plans to do so.)
Cheers,
Mike
On 2/15/16 12:41 AM, Sandeep Joshi wrote:
Thanks for all the helpful answers. I will check out vxquery next.
Algebricks seems like the equivalent of LLVM for query languages. I am
wondering if Algebricks is powerful enough to map any query language, be it
graph-based, relational or hierarchical (mrql, sql, pregel). Is there a
formal proof of this expressive power ? Will there always be a one-to-one
correspondence between the plan trees of different languages or would there
be a case where one would have to expand to look at sub-trees while doing
query translation ?
-Sandeep
On Mon, Feb 15, 2016 at 4:21 AM, Mike Carey <[email protected]> wrote:
PS: There's an important point below that you shouldn't miss (Sandeep) if
you look at the Hivesterix code - if you find its approach puzzling, note
that it was designed to only add what was needed to run Hive queries on
Hyracks - and so that it could potentially be kept in upper-level sync with
Hive itself. As a result, it was not done as a "Hive lookalike done right"
- it was done as a "Hive lookalike that lets the existing Hive code do as
much of the initial work as possible".
On 2/14/16 2:48 PM, Yingyi Bu wrote:
Hi Sandeep,
Here is the Hivesterix codebase in the Apache source tree:
https://github.com/apache/incubator-asterixdb-hyracks/tree/fullstack-0.2.13
We have maintained Hivesterix up to hyracks-0.2.13, but stopped
maintaining
after that release. Mike has elaborated the reason.
Furthermore, none of these rewrite rules seem to be SQL-specific. Are
there
any SQL-specific rewrite rules which were added?
That's exactly the motivation of the Algebricks project --- most rules
that
a typical SQL compiler implemented are not SQL-specific:-)
However, there indeed are few Hive-specific rules that I added in order to
get the Hive-on-Algebricks plan work efficiently:
https://github.com/apache/incubator-asterixdb-hyracks/tree/fullstack-0.2.13/hivesterix/hivesterix-optimizer/src/main/java/edu/uci/ics/hivesterix/optimizer/rules
The Hivesterix implementation first translates a Hive-optimized MR plan
into an Algebricks logical plan, and then let Algebricks do further
optimizations and finally execute the resulting Hyracks job on the Hyracks
runtime.
Best,
Yingyi
On Sun, Feb 14, 2016 at 2:26 PM, Mike Carey <[email protected]> wrote:
Sandeep,
Just to chime in as well:
- VXQuery is indeed the best example to look at, probably, to
understand
the AsterixDB/Algebricks separation.
- Hivesterix was built by Yingyi Bu (who'll see this) early on - it
drove
the separation idea, actually, but we made a decision not to try and
maintain it. It was intended to provide a third/different proof of
separation and applicability of the approach, from a research standpoint,
but doesn't have additional value to offer the world (since Hive itself
is
a moving target and Hive on Tez now provides the non-MapReduce-runtime
value that Hivesterix initially offered). Yingi would probably be happy
to
share the code base with you if you wanted to look at it for any reason,
but the only things in the Apache AsterixDB (incubating) project are
things
deemed worthy of engineering/maintenance work.
Hope that helps too!
Cheers,
Mike
On 2/14/16 11:47 AM, Till Westmann wrote:
Hi Sandeep,
Apache VXQuery, the XQuery implementation mentioned in the SoCC paper,
is
a separate project [1].
Specifically to your questions:
1) There is no need to implement other projects that use Algebricks
inside of the AsterixDB source tree (as VXQuery shows).
2) It is clearly easier to combine a Java parser and plan tree generator
with Algebricks, but there's no reason why one couldn't connect to other
languages (e.g. by using a text-based intermediate format between the
parser and the optimizer and between the plan generator and the
runtime).
3) The reason for the different set of rules is that some are language
agnostic and some are language-specific. As you can see in figure 2 of
the
paper a language implementation has to provide language-specific rules
to
augment the language-agnostic rules provided by Algebricks.
Specifically, the rules in AsterixDB's asterix-algebra project augment
the rules in Algebricks to support AsterixDB's query language AQL.
Hope this helps,
Till
[1] http://vxquery.apache.org
On 14 Feb 2016, at 11:02, Sandeep Joshi wrote:
I had some questions about the process of mapping other query languages
to
Algebricks. The Sigmod SoCC 15 paper mentions that two languages
XQuery
and HiveQL which have been mapped to Algebricks, but the implementation
is
not found in either of the two repositories released under Apache.
I found Hivesterix and Pregelix under
https://github.com/madhusudancs/hyracks/tree/master/fullstack/hivesterix
I couldn't find the XQuery to Algebricks translator anywhere. Has this
been released ?
What is the reason these language translators are not part of the
Apache
repository ?
The Apache repositories contain the language translators for AQL and
SQL.
After comparing the implementations for Hivesterix and SQL/AQL, here
are
some questions
1) Does one have to integrate the parser for a new language within the
Apache AsterixDB source tree, or can one build the Algebricks
translator
outside the Apache tree and invoke the Hyracks job execution engine
directly, as is being done in the hivesterix implementation seen here.
https://github.com/madhusudancs/hyracks/blob/36bb1021b17b736aa1648bd439e1246ae419aa89/fullstack/hivesterix/hivesterix-dist/src/main/java/edu/uci/ics/hivesterix/runtime/exec/HyracksExecutionEngine.java
2) When a query language is converted to Algebricks, the
ICompilerFactory
converts one plan tree to another by calling Visitor::visit() on each
node
of the source query. Does this imply that the plan tree for the source
language can only be constructed in Java ? Would it be
difficult/impossible to integrate a parser and plan tree generator
which
was written in any language into Algebricks ?
3) In the Apache repositories, the query rewrite rules which are used
during optimization are found under two different repositories.
One in main asterixdb repository
https://github.com/apache/incubator-asterixdb/tree/master/asterix-algebra/src/main/java/org/apache/asterix/optimizer/rules
and the other in the hyracks repository
https://github.com/apache/incubator-asterixdb-hyracks/tree/master/algebricks/algebricks-rewriter/src/main/java/org/apache/hyracks/algebricks/rewriter/rules
Are these two sets of rules characteristically different or is this
duplication just an artifact of rapid prototyping ?
Furthermore, none of these rewrite rules seem to be SQL-specific. Are
there any SQL-specific rewrite rules which were added ?
-Sandeep
