Reviews for the Revision of the paper Data-Parallel Query Processing on Non-Uniform Data, submitted to VLDB 2020.
Overall Rating: accept
1. Final and Overall Recommendation
Accept
3. Did the authors satisfactorily address the revision requirements identified in the meta-review of the original submission?
Only partially
5. Justify your answer to the above question.
The revision addressed most of the comments this reviewer raised. The additional experiments seem promising and done quite well considering the little time for revision.
6. Additional comments to the authors on the revised version of the paper
I applaud authors' efforts to improve the paper in a short period.
There are still a couple minor improvements you can do before camera ready, though.
Figure 4 should be placed in the same page as figure 5 and its walkthrough.
Figure 9 gives more question on the significance of load balancing. pk-8/32-fk shows significant improvements due to "coalesced" memory access. This effect seems as big as the load balancing or pushdown parallelism in general. Wasn't the load imbalance the main challenge? If coalescing memory accesses was the main challenge, there would be other, simpler solutions this paper should compare against. I might suggest one more experiment probably in appendix to further dig into these effects.
1. Final and Overall Recommendation
Accept
3. Did the authors satisfactorily address the revision requirements identified in the meta-review of the original submission?
Yes
5. Justify your answer to the above question.
The authors made a significant effort to address the comments raised during the original review phase; the edited manuscript has improved significantly.
6. Additional comments to the authors on the revised version of the paper
Nothing in particular.
1. Final and Overall Recommendation
Accept
3. Did the authors satisfactorily address the revision requirements identified in the meta-review of the original submission?
Yes
5. Justify your answer to the above question.The authors addressed all of my comments:
- An end-to-end performance comparison with OmniSci was added (addresses W1/D2)
- Unclear parts that relied on references were expanded (addresses W2/D1)
- The new Section 1.2 explains how DogQC works and the answers document elaborates on the vision for DogQC (addresses W3/D4).
- An explanation was added to why lane-refill is outperformed by the naive setup for high selectivities in Section 5.2 (addresses D3).
6. Additional comments to the authors on the revised version of the paper
I was already positive about the paper in its initial form, and I am happy to see that the revised paper has significantly improved upon it.
Regarding the vision for DogQC, I think it would be worth narrowing down even more the target use cases -- medium databases is not precise enough. The database market is shifting more and more towards the cloud, and databases requiring special hardware are at a disadvantage in the cloud, because it is a world ruled mainly by cost and flexibility. I personally believe that GPU databases will continue serving niche markets, but it will be very hard to expand any further, except maybe as accelerators and not full-blow DB engines.