|
SIGMOD 2008
2008 ACM SIGMOD International Conference on Management of Data
June 9-12 2008
Vancouver
Canada
|
Aggregates over review questions.
Min Is this paper relevant to the Industrial Track? I.e., is it about commercial information management software, or about industrial-strength prototypes of information management software in widespread use? | 0 |
Avg Overall rating | 2.00 |
Spread Overall rating | 4 |
Review: 1 |
Reviewer:
| Timos Sellis |
Email:
| timos@dblab.ece.ntua.gr |
Organization:
| NTUA University |
Review:
| |
| Question | Response |
1 | Is this paper relevant to the Industrial Track? I.e., is it about commercial information management software, or about industrial-strength prototypes of information management software in widespread use? |
No
|
2 | Originality: How novel is the work described? |
Very novel
|
3 | Technical Quality: How complete, deep, and correct is this work? |
Good
|
4 | Significance: How much can this paper teach the community? |
Some; community would benefit from reading
|
5 | Presentation: how readable is the paper? |
Good
|
6 | Overall rating |
Neutral
|
7 | Reviewer confidence |
Low
|
8 | Justification for rating (max 3 lines, please) |
Even though the work presented in this paper is not about an industrial project or prototype, the research ideas and results presented by the authors could possibly have an impact in the design of the query optimizer module of commercial systems. Moreover, two commercial database optimizers are used for the experiments and possible improvements are proposed for the Foreign-Plan-Costing feature.
|
9 | Detailed comments to authors: |
This paper addresses the problem of identifying robust query plans that can be used on operational database environments. A new algorithm (SEER) is proposed for plan diagram reduction that is non-intrusive with respect to the optimizer and provides plan performance guarantees that are applicable to individual points in the selectivity space (i.e. on an individual query basis). Using SEER, the optimizer’s choices are replaced by alternative plans from the parametric optimal set over the selectivity space. Those alternative plans are chosen in order to be more resilient to errors in selectivity estimates.
The paper is well presented, technically sound and the experimental results are consistent with the analysis from the authors. The main drawback of the proposed algorithm is that the execution time of the reduction process varies for the presented query templates from 2 minutes to more than 10 minutes, a cost maybe too high for an industrial application. The authors underline this problem and propose that possible improvements must be done to the Foreign-Plan-Costing feature of the commercial query optimizers examined.
Some comments:
1. During the plan diagram reduction process, in SEER algorithm (figure 12): It is possible that the costs fre(Pi, x, y) and foe(Pi, x, y) are computed more than one time for each optimal plan Pi and for some or, in the worst case, even all the points {x,y} of the selectivity space S.
This is a minor remark if f(x,y) is O(1), as it is implied in your analysis. But if fre(Pi, x, y) and foe(Pi, x, y) are calculated in real time through calls to FPC, those redundant calls could add a significant overhead to the overall running time. Is there a possible improvement of SEER algorithm without those redundant calls?
On the other hand, if fre(Pi, x, y) and foe(Pi, x, y) for every Pi are analytically modeled through surface-fitting techniques in the beginning of the reduction process, it should be explicitly stated in algorithm of figure 12.
2. Section (2.2) is well presented and self explanatory and I don’t believe that Figure 3 is needed for ease of understanding.
3. In section (2.3): “…the costs of Poe (P17), Pre (P45) and Poa are shown in Figure 4…”.
But in figure 4 you have Poe (P26) and Pre (P34)
4. In section (3.2), Selectivity Nodes: Coefficients c1 and c2 are a1 and a2 in figure 8.
|
10 | Comments for PC (will not be seen by authors) |
|
11 | Name of external reviewer, if any: |
some help from Ioannis Roussos
|
|
Review: 2 |
Reviewer:
| Vivek Narasayya |
Email:
| viveknar@microsoft.com |
Organization:
| Microsoft Research US |
Review:
| |
| Question | Response |
1 | Is this paper relevant to the Industrial Track? I.e., is it about commercial information management software, or about industrial-strength prototypes of information management software in widespread use? |
Yes
|
2 | Originality: How novel is the work described? |
Very novel
|
3 | Technical Quality: How complete, deep, and correct is this work? |
Good
|
4 | Significance: How much can this paper teach the community? |
Some; community would benefit from reading
|
5 | Presentation: how readable is the paper? |
Good
|
6 | Overall rating |
Accept
|
7 | Reviewer confidence |
Medium
|
8 | Justification for rating (max 3 lines, please) |
Follow on work relative to the VLDB 2007 paper (and prototype) on plan diagram reduction. The idea is to develop checks against arbitrary degradation due to plan reduction. The paper has novel technical contributions and takes a rigorous approach for proving correctness of checks. Experiments appear to validate the effectiveness of results.
|
9 | Detailed comments to authors: |
1. The cost model used for proving the sufficiency of perimeter checking, although a simplification of the actual cost model, is a plausible one; and such simplification is often necessary to derive analytical results as in this paper.
2. The experimental evaluation appears to be done for certain query templates. However, it is not mentioned how many dimensions (i.e., selectivity variables) were varied in these templates. For example, do the trends look similar for larger number of dimensions e.g., 4 or 5 instead of just 2?
3. In general, some of the selectivity variables could be on intermediate query expressions or views (although these are typically not avaiable in today's commercial DBMSs). Do the techniques of plan reduction presented here extend to such cases?
|
10 | Comments for PC (will not be seen by authors) |
The paper probably could also have been submitted to the research track, but there is sufficient justification for industrial track as well, since the prototype is developed over two different commercial DBMSs.
|
11 | Name of external reviewer, if any: |
|
|
Review: 3 |
Reviewer:
| Eric Simon |
Email:
| esimon@businessobjects.com |
Organization:
| Business Objects |
Review:
| |
| Question | Response |
1 | Is this paper relevant to the Industrial Track? I.e., is it about commercial information management software, or about industrial-strength prototypes of information management software in widespread use? |
No
|
2 | Originality: How novel is the work described? |
Slightly novel
|
3 | Technical Quality: How complete, deep, and correct is this work? |
Good
|
4 | Significance: How much can this paper teach the community? |
Some; community would benefit from reading
|
5 | Presentation: how readable is the paper? |
Good
|
6 | Overall rating |
Reject
|
7 | Reviewer confidence |
Medium
|
8 | Justification for rating (max 3 lines, please) |
This is a good and interesting research paper but it is not appropriate for the industrial track. It should be submitted to the research track of a forthcoming conference (I recommend VLDB).
|
9 | Detailed comments to authors: |
The topic of this paper is the robustness of a query plan when the statistics are erroneous. Indeed, existing optimizers build an expected good query plan by using heuristics and statistics on the input relations. In particular, an important factor governing the decisions taken by an optimizer is the selectivity of the query on the input relation. A robust optimizer is an optimizer which built plans that do not dramatically degrade when selectivity estimates are erroneous.
This problem has been recently studied in several research papers including the autjors who published a paper at VLDB 2007. This paper builds upon previously published propositions but in this paper the authors present a solution to the robustness problem which is based on a mathematic model, which is both simple and efficient. Experiments compare the efficiency of the proposed solution with a panel of other solution already proposed in the literature. The paper is clearly presented. It provides proof of the theorical results.
It is not clear to see how the proposed method could be easily implemented without deeply modifying existing optimizers. This aspect which is a key argument of the paper should be discussed.
|
10 | Comments for PC (will not be seen by authors) |
Despite the experimental evaluation and the integration of the proposed method in Picasso, the paper is more eligible as a research paper than an industrial paper.
|
11 | Name of external reviewer, if any: |
|
|
Review: 4 |
Reviewer:
| Mohamed Zait |
Email:
| mohamed.zait@oracle.com |
Organization:
| oracle |
Review:
| Not Available |
|
|