* DBSherlock Datasets

- Terms of use

Please cite the following papers if you use any part of this dataset:

@inproceedings{sigmod/YoonNM16,
  author    = {Dong Young Yoon and
               Ning Niu and
               Barzan Mozafari},
  title     = {{DBSherlock}: {A} Performance Diagnostic Tool for Transactional Databases},
  booktitle = {Proceedings of the 2016 {ACM} {SIGMOD} International Conference on
               Management of Data},
  year      = {2016}
}

@article{DBLP:journals/pvldb/YoonMB15,
  author    = {Dong Young Yoon and
               Barzan Mozafari and
               Douglas P. Brown},
  title     = {{DBSeer}: Pain-free Database Administration through Workload Intelligence},
  journal   = {{PVLDB}},
  volume    = {8},
  number    = {12},
  year      = {2015},
}


- Questions

For all inquiries please contact dyoon AT umich.edu


- Datasets

All datasets are available as MAT-files, which can be loaded in MATLAB or Octave.
There are three datasets used for conducting experiments with DBSherlock:

  * dbsherlock_dataset_tpcc_16w.mat: an original dataset that is obtained by
  running TPC-C benchmark with scale factor of 16 (i.e. 16 warehouses) as a
  normal workload. We mostly used this dataset to generate experiment results in
  the paper.

  * dbsherlock_dataset_tpcc_500w.mat: an additional dataset that is generated by
  running TPC-C benchmark with scale factor of 500 (i.e. 500 warehouses) as a
  normal workload. This dataset was generated in order to demonstrate DBSherlock's
  scalability on larger workloads, since it was one of the issues that had been
  raised by reviewers. We used this dataset to update experiment results wherever
  applicable.

  * dbsherlock_dataset_tpce_3000.mat: an additional dataset that is generated by
  running TPC-E benchmark with scale factor of 3000.

Please note that only 'dbsherlock_dataset_tpcc_16w.mat' contains datasets for
compound situations. Also, 'dbsherlock_datase_tpcc_16w.mat' contains one extra
test case called 'Database maintenance', which has been excluded in the paper
due to the test case being hard to reproduce in different systems. The other two
additional datasets do not contain the test case 'Database maintenance'.

'dbsherlock_dataset_tpcc_16w.mat' contains seven fields:

  1. causes: a cell structure that contains the name of each performance anomaly
  (i.e., test case).

  2. test_datasets: 11-by-11 cell structure, where each row represents different
  test cases and each column represents different duration or start times of the
  anomaly, ranging from 30 seconds to 80 seconds with increment of 5.
  For example, test_datasets{2,4} is a dataset with the anomaly 'I/O Saturation'
  with the anomaly duration of 45 seconds.

  3. compound_datasets: 6-by-11 cell structure. It is similar to 'test_datasets'.
  Only difference is that each row represents different compound test cases.

  4. normal_regions: this contains the normal region of each dataset in
  'test_datasets' with corresponding indices.

  5. abnormal_regions: this contains the abnormal region of each dataset in
  'test_datasets' with corresponding indices.

  6. normal_regions_compound: this contains the normal region of each dataset in
  'compound_datasets' with corresponding indices.

  7. abnormal_regions_compound: this contains the abnormal region of each dataset
  in 'compound_datasets' with corresponding indices.

'dbsherlock_dataset_tpcc_500w.mat' and 'dbsherlock_dataset_tpce_3000.mat' have
identical fields except they are missing fields related to the compound test
cases.
