FeiGSSS
commented
5 days ago This fault injection record was considered to be unjustified and thus ignored by the authors. Millions of seconds of fault injection is outrageous.
By the way, this data cleaning result didn't come from this paper either, I first saw it proposed in GAIA's github repository . Later in the DiagFusion paper, this labeled data was removed and the cleaned labeled data was made public in their repository. TVDiag should have followed their labeling file
Dear Authors,
Dataset A contains CPU anomalies, e.g.:
2021-07-29,logservice1,"2021-07-29 22:09:57,277 | WARNING | 0.0.0.3 | 172.17.0.3 | logservice1 | [cpu_anomalies] trigger a parallel fast sorting program, start at 2021-07-29 22:09:57.274933 and lasts 1985016.0505759716 seconds.These are not present in the preprocessed data attached as an example, and neither the article address them (e.g., Fig. 13 does not depict this failure type, Table 3. also states that there are 5 failure types only in the dataset).
What is the reason for this? Can TVDiag handle these failures, or should the Evaluation section be interpreted without considering these failures?