quickstep-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Harshad Deshmukh <hars...@cs.wisc.edu>
Subject Re: Do we need 3 different InsertDestination classes?
Date Wed, 15 Jun 2016 14:31:32 GMT
Hi Navneet,

I was always in favor of getting rid of AlwaysCreateBlock derived class. 
This issue was raised long time back, but Craig insisted that we keep it.

I guess the PartitionAware class can't be an efficient drop in 
replacement for the BlockPool class. The reason is that it checks for 
each input tuple, which output partition should it be sent to. This is 
an unnecessary step when we don't have partitioning enabled.

An alternative is something called PartitionPreserving derived class, 
that Gerald wrote. It skips the above check that I mentioned.

Gerald, please feel free to correct if I didn't get it right.

On 06/15/2016 03:20 AM, Jignesh Patel wrote:
> A huge +1 from me to get rid of dead code!
> +CC: Gerald who also worked on this a while back, in case he has some input.
> Thanks!
> Jignesh
>> On Jun 14, 2016, at 11:36 PM, Navneet Potti <navsan@gmail.com> wrote:
>> Hi Harshad
>> I’m just kicking off this discussion based on a conversation I had with Jignesh
this morning. We have 3 different implementations of InsertDestination, and I think we only
really need one, the PartitionAwareBlockPoolInsertDestination.
>> The AlwaysCreateBlockInsertDestination seems to be entirely useless practically.
I’m guessing it’s a relic of early development. The normal BlockPoolInsertDestination
seems like a degenerate case of the PartitionAware one, with a single partition.
>> If my understanding of the design is right, then we should refactor this code to
only have one InsertDestination. There is a lot of duplication of code and complex logic in
there. Anything we can do to simplify and clean up the code is probably a good move.
>> Thoughts?
>> Cheers,
>> Navneet


View raw message