acceptodds
Under review as a conference paper at ICLR 2027

Root Cause or Delayed Symptom? Lag-Calibrated Counterfactual Inference for Microservice Fault Localization

Abstract

In microservice systems, fault propagation often involves non-negligible time lag, and faults in upstream services may manifest as anomalies in downstream services only after several time steps. Therefore, relying solely on anomaly signals observed at the current timestamp often makes it difficult to distinguish the true upstream root cause from downstream delayed response services. Most approaches rely on anomaly intensity or statistical correlations at individual timestamps, without explicitly modeling propagation delays or temporal causal relationships. As a result, they are prone to errors when anomaly signals are not synchronized across services. To address this issue, we propose RippleRCA, a hierarchical lag-calibrated causal inference framework for root cause analysis in microservice systems. Specifically, RippleRCA begins by capturing global cross-service propagation dependencies through edge-level and node-level lag-aware causal priors constructed with PCMCI+ under service call topology constraints. It then performs coarse-grained candidate recall from both the structural anomaly and propagation response perspectives. This is achieved through a dual-channel mechanism based on the features of temporal difference. Furthermore, RippleRCA further performs lag-calibrated temporal alignment and estimates counterfactual effects via recovery interventions, enabling fine-grained causal root cause ranking. Extensive experiments on two public datasets, AIOps2025 and RCABench, demonstrate the effectiveness of RippleRCA. Ablation studies and parameter sensitivity analysis further confirm the effectiveness of lag-aware causal priors, propagation response modeling, and the hierarchical two-stage design.

Primary area:
causal reasoning
Keywords:
Microservice Systems, Root Cause Analysis, Propagation Delay, Causal Inference, Counterfactual Reasoning

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_GyRUOpenReview ↗
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 KK 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
  1. 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.
  2. Code is provided for reimplementation.
Critical weaknesses
  1. 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.
  2. The reported evaluation contains inconsistencies and leaves the diagnostic protocol unclear. For the fusion weight, Eq 14 reduces to the Stage 1 ranking when β=1\beta = 1. Appendix F and Figure 4 report unchanged RCABench performance across β∈[0,1]\beta \in [0, 1]. 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.
  3. 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.
  4. The clarity of this paper is poor. Terms like ϕ,rv(τ)\phi, r_v(\tau) are not defined. It's not clear if ΦG\Phi_G is trained and how. It's not clear whichi run is Fig. 4 showing.
Questions for authors
  1. Can the author please explain the inconsistency of RCABench performance?
  2. What makes Eq 11 discriminate an originating fault from an influential downstream mediator?
  3. 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?
  4. Which fault instances contribute to Stage 1’s cross-instance statistics, and when are those statistics computed relative to evaluation?
  5. Does the method localize services or individual instances?
  6. How were the baselines configured and tuned, and did all methods use the same evaluation cases, candidate universe, and root-cause identifier mapping?
  7. How were the hyperparameters selected, and which data were used for selection? The paper reports sensitivity analyses for τmax⁡\tau_{\max}, KK, and β\beta, but does not sufficiently specify the settings or selection procedures for λ1,λ2\lambda_1,\lambda_2, the SCM ridge regularization, and the PCMCI+ test and thresholds. How dependent are the reported gains on the relative weight of the counterfactual term?
  8. 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 Scf=0S_{\mathrm{cf}}=0 to leaf nodes. How does the method handle genuine root causes at these nodes, and is its localization performance different for this subset?
Reviewer_dDwoOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

RippleRCA is a two-stage microservice root cause analysis framework. It constructs topology-constrained lag priors and combines structural anomalies with propagation responses to retrieve candidates. A lagged structural causal model then performs temporally aligned restoration interventions to refine their ranking. Experiments on AIOps2025 and RCABench show improvements over six baselines, supported by ablation and robustness studies.

Critical strengths
  • Distinguishing root causes from delayed downstream symptoms addresses an important challenge in microservice diagnosis.
  • Combining topology, historical observations, lag priors, and counterfactual reasoning provides complementary evidence for candidate retrieval and refinement.
  • Results improve localization accuracy and ranking quality on two public benchmarks. Ablation, sensitivity, and topology perturbation experiments provide additional empirical support.
Critical weaknesses
  • Existing methods, including LagRCA, already exploit propagation delays. The paper insufficiently explains its additional contribution or why incoming lag priors determine appropriate intervention times. Ablations demonstrate utility but do not validate this mechanism.
  • No incident illustrates how observations become lag priors, candidates, counterfactual effects, and final rankings. This makes the complete diagnostic process difficult to follow and its causal explanations difficult to assess.
  • RCABench relies on injected faults, while the provenance of AIOps2025 incidents remains unclear despite its production-environment description. Validation on naturally occurring incidents and unseen failure scenarios is needed to substantiate practical generalization.
Questions for authors

Please refer to the Weakness section above.

Reviewer_owF3OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes RippleRCA, a lag-calibrated causal inference framework for root cause analysis in microservice systems, addressing the challenge of delayed fault propagation across services. The framework first constructs lag-aware causal priors using PCMCI+ under service call topology constraints to capture cross-service dependencies. It then employs a dual-channel mechanism based on temporal-difference features to identify potential root cause candidates from structural anomaly and propagation response perspectives. Finally, it performs lag-calibrated temporal alignment and counterfactual effect estimation through recovery interventions to rank candidate root causes. The proposed framework is evaluated on two public datasets, AIOps2025 and RCABench, with experimental results demonstrating its effectiveness.

Critical strengths
  1. The paper addresses an important challenge in root cause localization, where downstream symptoms may be more prominent than the initial fault due to propagation delays. Distinguishing root causes from delayed responses is important for reliable diagnosis.

  2. The proposed framework is well structured, combining topology-constrained lag discovery and coarse-grained candidate selection with counterfactual ranking. This provides a reasonable overall design for root cause localization.

Critical weaknesses
  1. The framework uses the given service topology as a constraint and applies the existing PCMCI+ method to identify lagged dependencies between services. It is unclear how much improvement comes from the proposed components rather than PCMCI+ itself. A simpler PCMCI+-based baseline, where anomalous services are ranked using outgoing dependency strengths and anomaly timing, would help demonstrate the contribution.

  2. In the sensitivity analysis, τmax=0\tau_{max} = 0 is explored, which removes nonzero lags (Fig. 2(a)). However, it achieves comparable performance to τmax=5\tau_{max} = 5 on AIOps2025. This raises questions about whether explicitly modeling propagation delays provides substantial benefits, given that it is one of the main motivations of the paper.

  3. According to Eq. 14, setting β=1\beta=1 should make the ranking depend entirely on Stage 1. However, Table 2 reports an AC@1 of 0.122 for Stage 1 alone on RCABench, while Appendix F (Fig. 4(c)) shows unchanged performance across β\beta values from 0 to 1. Can the authors clarify why these results differ?

  4. In Eq. 7, Stage 1 constructs a global candidate ranking by aggregating information across different fault samples. However, different fault incidents may have different root causes, and frequently selected services are not necessarily responsible for a particular incident.

  5. In Eq. 11, the counterfactual effect measures changes in downstream trajectories after restoring a candidate node. However, restoring a downstream service may also produce large changes without resolving the original fault. It is unclear whether this score measures root cause responsibility or simply downstream influence.

Questions for authors

Please see the "Critical Weaknesses" section.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.