+1. We on the Kudu team are interested in exposing data blocks using this in-memory format as the results of our scan operators (via RPC or shared memory transport). Standardizing it will make everyone's lives easier (and the performance much better!)
-Todd On Mon, Oct 26, 2015 at 5:22 PM, Wes McKinney <[email protected]> wrote: > hi all, > > I am excited about this initiative and I personally am looking forward > to seeing a standard in-memory columnar representation made available > to data science languages like Python, R, and Julia, and it's also the > ideal place to build out a reference vectorized Parquet implementation > for use in those languages (lack of Python/R Parquet support has been > a sore spot for the data science ecosystem in recent times). This will > also enable us to create an ecosystem of interoperable tools amongst > SQL (Drill, Impala, ...) and other compute systems (e.g. Spark) and > columnar storage systems (e.g. Kudu, Parquet, etc.). > > Having richer in-memory columnar data structures alone will be a boon > for the data science languages, which are working also to improve both > in-memory analytics and out-of-core algorithms, and any distributed > compute or storage system that can interoperate with these tools will > benefit. > > thanks, > Wes > > On Mon, Oct 26, 2015 at 2:19 PM, Jacques Nadeau <[email protected]> > wrote: > > > > Drillers, > > > > > > > > A number of people have approached me recently about the possibility of > collaborating on a shared columnar in-memory representation of data. This > shared representation of data could be operated on efficiently with modern > cpus as well as shared efficiently via shared memory, IPC and RPC. This > would allow multiple applications to work together at high speed. Examples > include moving back and forth between a library. > > > > > > > > As I was discussing these ideas with people working on projects > including Calcite, Ibis, Kudu, Storm, Herron, Parquet and products from > companies like MapR and Trifacta, it became clear that much of what the > Drill community has already constructed is very relevant to the goals of a > new broader interchange and execution format. (In fact, Ted and I actually > informally discussed extracting this functionality as a library more than > two years ago.) > > > > > > > > A standard will emerge around this need and it is in the best interest > of the Drill community and the broader ecosystem if Drill’s ValueVectors > concepts and code form the basis of a new library/collaboration/project. > This means better interoperability, shared responsibility around > maintenance and development and the avoidance of further division of the > ecosystem. > > > > > > > > A little background for some: Drill is the first project to create a > powerful language agnostic in-memory representation of complex columnar > data. We've learned a lot over the last three years about how to interface > with these structures, manage memory associated with them, adjust their > sizes, expose them in builder patterns, etc. That work is useful for a > number of systems and it would be great if we could share the learning. By > creating a new, well documented and collaborative library, people could > leverage this functionality in wider range of applications and systems. > > > > > > > > I’ve seen the great success that libraries like Parquet and Calcite have > been able to achieve due to their focus on APIs, extensibility and > reusability and I think we could do the same with the Drill ValueVector > codebase. The fact that this would allow higher speed interchange among > many other systems and becoming the standard for in-memory columnar > exchange (as opposed to having to adopt an external standard) makes this a > great opportunity to both benefit the Drill community and give back to the > broader Apache community. > > > > > > > > As such, I’d like to open a discussion about taking this path. I think > there would be various avenues of how to do this but my initial proposal > would be to propose this as a new project that goes straight to a > provisional TLP. We then would work to clean up layer responsibilities and > extract pieces of the code into this new project where we collaborate with > a wider group on a broader implementation (and more formal specification). > > > > > > Given the conversations I have had and the excitement and need for this, > I think we should do this. If the community is supportive, we could > probably see some really cool integrations around things like high-speed > Python machine learning inside Drill operators before the end of the year. > > > > > > > > I’ll open a new JIRA and attach it here where we can start a POC & > discussion of how we could extract this code. > > > > > > Looking forward to feedback! > > > > > > Jacques > > > > > > -- > > Jacques Nadeau > > CTO and Co-Founder, Dremio > > > -- Todd Lipcon Software Engineer, Cloudera
