drill-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Eric Tschetter <ched...@metamarkets.com>
Subject Re: metamarkets druid
Date Tue, 30 Oct 2012 16:56:37 GMT
> Is it jus the late hour or they are not using any code generation for their query execution
engine?
> The GroupBy execution for instance looks quite complex, and doesn't look very efficient,
since it was written generically,

Sorry if this doesn't actually thread properly, I apparently fail at
being able to reply to messages that were sent before I joined the
list.

Just wanted to respond here.  Yes, it's true that we aren't using code
generation for our queries.  It's also true that the GroupBy
implementation is not optimal.  Our benchmarks are based on a
timeseries-specific query that is still tangled up with internal stuff
and so has not been added to the OSS release yet.  The GroupBy
implementation is there as a first pass at something to allow for it
to just function and us to explore and squash the higher order bugs
related to memory management and streaming of results back up to the
query broker.  Code generation is definitely on our roadmap for
potential ways to optimize query execution, but it has not been the
highest priority especially given that for our other query types (like
the timeseries query that is yet to open up), just relying on
JIT-compilation of pure Java has been plenty fast and probably easier
to maintain (I say probably, 'cause I don't really know how hard it
would be to maintain the code-generated solution as we haven't done
it).

--Eric

Mime
View raw message