- Rating
- 1
- Confidence
- 3
- Informativeness
- 1
- Presentation
- 1
- Soundness
- 1
Read review
Paper summary
The paper proposes RippleRCA, a two-stage root-cause analysis pipeline for microservices designed to separate upstream root causes from downstream services that respond with a delay. In the first stage, lag-weighted multi-lag devisation features are encoded with a graph encoder. For each fault case, top candidates are retrieved from a structural anomly channel and propagation response channel. Then aggregate these across fault cases into a global score for this stage. In stage 2, a ridge-linear lagged SCM is fitted on normal windows. For each candidate, residuals are abducted, a normal-state intervention is applied at the candidate's dominant lag, and the RMS change of its descendants is used as a counterfactual effect. This is fused with a lag-aligned local anomaly score and with the Stage-1 score. It is tested on AIOps2025 and RCABench.
Critical strengths
- The paper targets a real failure mode. Ranking delayed downstream symptoms above the true source is a practical problem in RCA. The paper includes component ablations, sensitivity analyses, and a robustness study under missing or spurious topology edges, which is a realistic deployment concern.
- Code is provided for reimplementation.
Critical weaknesses
- The proposed counterfactual quantity does not establish that a candidate is the fault source, and the temporal alignment lacks sufficient justification. Eq 11 measures how much restoring a candidate changes its descendants. This can identify an influential intervention point, but it does not necessarily identify where the fault originated. It also measures change rather than whether downstream behavior becomes more normal.
- The reported evaluation contains inconsistencies and leaves the diagnostic protocol unclear. For the fusion weight, Eq 14 reduces to the Stage 1 ranking when . Appendix F and Figure 4 report unchanged RCABench performance across . Table 2 reports Stage 1-only AC@1 of 0.122, compared with 0.370 for the full model. Without an interpretation on the difference in the settings across these expeirments, they seem to be inconsistent. Also, the counterfactual contribution is not isolated. Stage 2 only still includes local anomaly evidence, and removing lag alignment retains counterfactual inference. An ablation here is needed.
- Novelty over LagRCA is incremental, since lag-aware causal RCA is that paper's main contribution. LagRCA already explicitly models delayed fault propagation, so RippleRCA’s additional contribution needs to be established through its candidate-selection and counterfactual-refinement mechanisms. While I do not see this justified in RippleRCA's method.
- The clarity of this paper is poor. Terms like are not defined. It's not clear if is trained and how. It's not clear whichi run is Fig. 4 showing.
Questions for authors
- Can the author please explain the inconsistency of RCABench performance?
- What makes Eq 11 discriminate an originating fault from an influential downstream mediator?
- Why does a lag aggregated from incoming edges determine the appropriate intervention time for explaining downstream responses? How are heterogeneous outgoing delays and cumulative path delays handled under the default response horizon?
- Which fault instances contribute to Stage 1’s cross-instance statistics, and when are those statistics computed relative to evaluation?
- Does the method localize services or individual instances?
- How were the baselines configured and tuned, and did all methods use the same evaluation cases, candidate universe, and root-cause identifier mapping?
- How were the hyperparameters selected, and which data were used for selection? The paper reports sensitivity analyses for , , and , but does not sufficiently specify the settings or selection procedures for , the SCM ridge regularization, and the PCMCI+ test and thresholds. How dependent are the reported gains on the relative weight of the counterfactual term?
- How are contemporaneous edges oriented when constructing the SCM, and does the topology constraint enforce caller-to-callee causal direction or only restrict adjacency? Appendix D explicitly assigns to leaf nodes. How does the method handle genuine root causes at these nodes, and is its localization performance different for this subset?