FASTune: Towards Fast and Stable Database Tuning System with Reinforcement Learning
Round 1
Reviewer 1 Report (Previous Reviewer 1)
As this is a resubmitted manuscript, it would have been very nice to get a document where I can see the changes. Now, I had to read the whole paper again and was not able to focus on the important parts that have been improved.
The new version of the paper seems to address all issues stated in tha last version. Except for some minor language errors, I have no further points to address.
Author Response
Please see the attachment.Author Response File: 
Author Response.pdf
Reviewer 2 Report (New Reviewer)
This paper deals with the problem of automatic database tuning, and in particular, aims to perform database tuning faster than the existing RL-based tuning techniques. This paper is well written with enough experiments, suggested methods are well described, and results look promising.
My two criticisms:
1) it is not clear whether authors repeated their experiments a few times to see if there is variance in the outcomes.
2) In the experiment, I/O was simply separated into read and write operations, but I doubt whether indexed read, write operations, and I/O types according to various index types(hashing, B+-tree, etc) should also be considered.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report (New Reviewer)
Please check the expression of the formula(1)and formula(2).
Please add the quantitative analysis describing the results about Figure 5 on lines(579-585).
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 4 Report (New Reviewer)
The reviewer became acquainted with the research results contained in the article with great interest. Fulfilling the duties of a reviewer, it should be pointed out that in Figure 2 in the Agent block there is a minor error, it is Crtic instead of Critic. Moreover, the comparative studies concerned only two relational database management systems, namely MySQL and OpenGauss. PostgeSQL reviewer is missing from this statement.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
This manuscript is a resubmission of an earlier submission. The following is a list of the peer review reports and author responses from that submission.
Round 1
Reviewer 1 Report
In the given paper, the authors present "FASTune", a new algorithm that utilizes Deep Reinforcement Learning for automatically tune database configurations.
All in all, the paper is nicely written. The authors provide a good introduction to the topic. The main idea of the algorithm and their idea of using virtual environments that are simulating the database though a MLP that is trained on historical data are explained into detail.
The results are indicating that the presented algorithm is able to achieve tuning results of a similar quality than state-of-the-art algorithms, but in much shorter time.
Nevertheless there are some parts that should be discussed in a more detailed way.
My main concern, however is regarding the reproducability of the results.
Also the paper explaines the main idea of their algorithm, there are no explicit information about the internal structure of the FASTune algorithm (number of layers, number of neurons, activation functions, optimizer). There are no information that explaines the concrete training process and corresponding hyperparameters, e.g. as learning rate.
The paper lists many values that are tunable hyperparamters, but there is no which values are utilized in the experiments.
The authors have written that they intend to publish the source code and the dataset in the future. However, I don't think that - in its current state - researchers that are reading this paper would be able to reproduce the given findings and use them as the basis for further research in this area.
While I fully understand that it is a lot of work to clean the source code and to publish it on a platform like www.github.com, I would like to encourage the authors to provide enough information in their paper for the readers to reproduce the presented model.
Another point that I want to address are some language issues. Here are some examples:
- l. 72: To address this issue, the dispatcher is proposed[. The] dispatcher divides
- l. 81: A Mu[l]ti-State Soft Actor-...
- l. 115: In the beginning, [the] agent randomly recommends configurations and applies them to the database[. Then the ] agent
- l. 117 RL can achieve high performance and does not require training data, but they...
- What 'they'? Do you mean 'the' RL agent or 'it' which relates to RL?
- l.148-158: Here are a lot of articles (the) missing
- ...
In the following, I list some other issues that catched my eye:
- l. 81: Here I miss the reference for the Muti-State Soft Actor-Critic model, or at least the Actor-Critic model
- l. 124: Here, you are explaining Reinforcement Learning. I would have expected at least one reference of the important works in that field, i.e. the book "Reinforcement Learning: An Introduction" of Sutton and Barto. In addition the description misses the essential part of this kind of methods like, "Inspired by the way biological agents act", "Optimize problems using a reward signal", "Markov Descision Process with state, action, reward".
Later in the text, you are also mentioning the Q-Value, but you never explain what this value ist, i.e. the action-state value, and what its characteristics are. You also not explain completely how the actor-critic algorithm works.
- l. 189: Where ist the reference for the DDPG method
- l. 214: FASTune combines the Workload State with and Database State as the Muti-State, which is provided to the Agent when generating an Action;
- I don't fully understand what you want to emphasize. Can you explain this in a more detailed way.
- l. 265: Reference for SVM?
- l. 284: Why doing this in future and not publishing it now as additional material to this paper?
- l. 331: Reference for FAST-MDC
- l. 356: More details on virtual environment. You write that it is an MLP with 4 layers.. Which kind of activation function, which optimizer, what hyperparameters? Who long was it trained. How big was you dataset of historical data?
Reviewer 2 Report
This paper proposes a fast and stable RL-based database tuning system, named FASTune. The FASTune is a virtual environment used to evaluate configurations. It is an equivalent yet more efficient scheme than the cloned database. Experimental results indicate that, compared with the state-of-the-art methods, FASTune can achieve performance improvement while maintaining stability in the tuning
The topic is interesting, but the paper’s clarity and completeness have to be improved.
1. The literature review is not sufficient to support that reinforcement learning is a powerful method to tune database systems.
2. In section 5, the core functions used in the environment proxy shall explain clearer, for example, DB-BERT, support vector machines, radial basis function, outlier detection technique, and minimum covariance determinant, etc.
3. In the whole paper, many technical terms and methods are not explained clearly, resulting in low readability of the paper.
4. The parameters in the paper and algorithm are not clearly defined and are not rigorous. It is recommended to organize and present in tables.
5. This paper seems to be in the experimental stage and cannot provide a comprehensive research contribution.
