acceptodds
Under review as a conference paper at ICLR 2027

PFMR: ALLOCATING SUPPORT BY LOCAL EFFECT FOR TRACE-BASED ROOT CAUSE LOCALIZATION

Abstract

A single microservice failure can produce anomalies across a request path, making it difficult to decide which service to inspect first. Statistical support for a change and the magnitude of local deterioration carry different diagnostic information. We propose PFMR, a trace-based method that uses positive local effect to allocate two-window rank support across services. It updates a change reference with effect-weighted support, pooling the two signals within each service. Under additive execution, we derive a condition for separating local deterioration from propagated anomalies; a finite support ceiling ensures that weak local effects contribute vanishingly little at fixed sample counts. PFMR achieves 78.13% and 87.92% service-level Top-1 accuracy on FSE1422 and RCAEval240. At a fixed total support contribution, it improves Top-1 by 5.70 and 8.33 percentage points over a common coefficient and by 3.38 and 3.33 points over normalized additive fusion. Prediction-defined groups show the largest gains when the reference and support prefer different services and the support-preferred service has the larger local effect.

Primary area:
causal reasoning
Keywords:
Root cause analysis

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_3Rj6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

PFMR is a trace-based root cause localization method that uses positive local deterioration to weight statistical support when ranking services. It combines a change reference with separately pooled effect and support signals. The paper derives conditional separation results and reports Top-1 accuracies of 78.13% on FSE1422 and 87.92% on RCAEval240, alongside controlled comparisons of alternative ranking rules.

Critical strengths
  • PFMR substantially outperforms the listed baselines on both datasets, achieving consistently high service-level localization accuracy.
  • Shared service statistics and fixed-total-contribution controls help isolate the benefit of effect-dependent support allocation.
  • The paper formalizes its ranking rule and provides sufficient conditions for separating local deterioration from propagated changes.
Critical weaknesses
  • The contribution centers on how local effect weights statistical support. The paper does not sufficiently establish why this specific refinement constitutes a substantial research advance beyond existing combinations of local deviation and propagation evidence.
  • Although ranking ablations and subgroup analyses are provided, they do not directly establish how frequently strong support accompanies weak propagated effects in actual incidents. The detailed analytical observations need concrete examples and measurements of their underlying conditions.
  • Both datasets use injected faults in experimental microservice environments. Strong results therefore leave uncertainty about naturally occurring production failures and imperfect tracing conditions.
  • No incident is followed from raw spans through local-effect estimation, support pooling, and final ranking. Such a walkthrough would clarify the method and demonstrate how it corrects a misleading initial diagnosis.
Questions for authors

Please refer to the Weakness section above.

3 comments

If you have any questions about our responses during the discussion period, we would be happy to address them. We look forward to your further feedback.

Dear Reviewer 3Rj6,

Thank you very much for your help with reviewing for ICLR'27!

The authors have now responded to your reviews. I'm wondering if you could have a look at their responses and see whether they have addressed your concerns. And could you also leave a comment to let the authors know your thoughts? If any concerns remain, please explain which ones have not been addressed and why.

Warm regards, Your AC

With one day remaining in the discussion period, we have designed experiments specifically to address your questions and provided our responses. We would greatly appreciate your feedback to help us understand whether your concerns have been adequately addressed. Thank you for your time and consideration.

Reviewer_Fj5FOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes PFMR for service-level root cause localization from traces. For each service, it computes a change reference DvD_v, a positive local effect EvE_v, and a rank-based support score IvI_v. It then ranks services by logit(Dv)+Evlog⁡(1+Iv)\mathrm{logit}(D_v) + E_v \log(1+I_v). The main claim is that the size of local deterioration should decide how much the statistical support counts.

Critical strengths

The observation that a service with strong statistical evidence is not always the one where the delay starts makes sense, and the caller/child example in the introduction shows it well. The accuracy numbers are high on both datasets.

Critical weaknesses

My main concern is that I could not find a clear statement of the problem. The abstract says that statistical support and local deterioration "carry different diagnostic information," and the introduction says that "their roles are distinct." Neither part explains what makes this hard, or where current methods fail. Without this, it is difficult to tell what PFMR is meant to fix. The same issue affects the contributions: the first one describes a ranking rule, but I could not see in what way it improves on prior work.

The writing makes the method hard to follow. Terms such as "two-window rank support" and "change reference" appear in the abstract without explanation, and the reader has to reach Section 2.3 or the appendix to understand them. Section 2 begins directly with Equation (1). There is no short overview that links each part to the motivation. Several design choices are not explained, for example the use of the 1-D Wasserstein distance for duration change. Section 2.3 is especially confusing because the definition of the statistic, its upper bound, and comments on effect versus support are mixed together.

The method uses the call graph only to subtract child durations. Proposition 1 assumes additive execution and unchanged local execution at competing services. Real systems often have asynchronous calls, retries, shared resources, and multi-hop effects, and these break such assumptions. The paper does not discuss this.

Questions for authors
  1. Can the authors give a formal definition of the problem, and show a concrete case where existing methods fail but PFMR succeeds?
  2. Why use the Wasserstein distance here? Was a simpler measure, such as a shift in the mean or quantiles, tested?
3 comments

We would greatly appreciate any further feedback you may have on our responses during the discussion period. Should there be any points you would like us to clarify, we would be more than happy to do so.

Dear Reviewer Fj5F,

Thank you very much for your help with reviewing for ICLR'27!

The authors have now responded to your reviews. I'm wondering if you could have a look at their responses and see whether they have addressed your concerns. And could you also leave a comment to let the authors know your thoughts? If any concerns remain, please explain which ones have not been addressed and why.

Warm regards, Your AC

With one day remaining in the discussion period, we have designed experiments specifically to address your questions and provided our responses. We would greatly appreciate your feedback to help us understand whether your concerns have been adequately addressed. Thank you for your time and consideration.

Reviewer_ufJ3OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PFMR, a trace-based root cause localization method that combines positive local deterioration with statistical support. Its key innovation is to weight two-window rank support by service-specific effect magnitude. The paper provides conditional theoretical analysis and evaluates the method on FSE1422 and RCAEval240, achieving Top-1 accuracy of 78.13% and 87.92%, respectively. Controlled ablations further demonstrate improvements over common-coefficient weighting and normalized additive fusion.

Critical strengths

Well-motivated methodology. The distinction between local effect magnitude and statistical support addresses a fundamental challenge in separating root causes from propagated anomalies. The effect-dependent allocation provides a principled ranking mechanism.

Strong empirical validation. The method substantially outperforms the reported baselines on two benchmarks. Particularly convincing are the fixed-total contribution controls, which demonstrate Top-1 improvements of 5.70 and 8.33 percentage points over common-coefficient weighting, with positive paired confidence intervals.

Critical weaknesses

Limited validity of local execution attribution. Equation (14) subtracts summed child durations, which may double-count overlapping calls and underestimate local execution time. Since the theoretical separation guarantee assumes additive execution, its applicability to concurrent and asynchronous microservice traces remains unclear. This directly affects the validity of the proposed local-effect measure.

Limited robustness to disappearing operations. Equation (4) excludes operations observed only in the normal window, potentially suppressing evidence from services that fail completely during an incident. The very low Top-1 accuracy for PodFailure (4.17%) raises concerns about this limitation, although the paper does not establish that it causes these errors.

Questions for authors

Overlapping executions: How frequently do child spans overlap in the evaluated traces? How does PFMR perform when local duration is calculated using the union of child intervals instead of the sum of child durations?

Missing operations: What proportion of root-cause services have no observations in the fault window? Can the authors analyze whether the exclusion of normal-only operations contributes to the poor PodFailure results?

Sensitivity of effect weighting: How robust are the improvements to alternative effect transformations and support-weighting functions? This would clarify whether the gains arise from the general principle of effect-dependent weighting or the specific proposed formulation.

Comments on the paper

3 comments
AuthorsOpenReview ↗

Summary of Responses to the Reviewers

We thank all three reviewers for their careful evaluation and constructive feedback, which have helped us further improve our paper. We have aimed to address their questions and concerns clearly and concisely. Below, we briefly summarize the reviewers’ main comments and our responses.

  • Reviewer Fj5F recognized the motivating observation and strong accuracy, and requested clearer problem formulation, design rationale, and theoretical assumptions. We clarified the problem and differences from prior work, explained the Wasserstein choice and theoretical conditions, and provided comparative experiments and concrete examples.
  • Reviewer ufJ3 valued the distinction between local effect and statistical support and the controlled experiments, and asked about overlapping calls, disappearing operations, and weighting sensitivity. We addressed these through trace audits and controlled comparisons of residual calculation, the treatment of disappearing operations, and alternative weighting rules.
  • Reviewer 3Rj6 recognized the strong results, controlled comparisons, and theoretical analysis, and requested stronger evidence of the contribution and practical relevance. We quantified how often support and local effect disagree, provided additional application results, and followed an incident from raw spans to the final ranking.

Thank you again for your time and careful review. We will continue to address any questions you may have during the discussion period and would appreciate your feedback on our responses.

AuthorsOpenReview ↗

Dear Area Chair and Reviewers

To facilitate a clearer understanding of our work, we would like to briefly summarize the central idea of PFMR, starting from the common questions raised in the reviews. We hope this concise explanation will help clarify the motivation behind our method and how its key components work together.

PFMR grew out of a practical difficulty in service operations: a single fault can produce changes across many services, while the strongest statistical signal may not identify where the problem began. Our aim is to provide a relatively simple computational framework with a clear path from trace observations to root-cause ranking. The central idea is to distinguish how much a service’s own behavior deteriorates from how strongly the observations support a change. Local deterioration then determines how much that support influences the ranking. We hope this explicit connection between observations and decisions helps engineers understand why a service is prioritized and focus their investigation on the source of a failure.

As the discussion period draws to a close, we sincerely hope to receive any further feedback from the Area Chair and Reviewers. Your comments would be invaluable in helping us further improve the clarity, presentation, and overall quality of our paper. We would be more than happy to address any remaining questions or concerns.

Thank you again for your time and consideration.

AuthorsOpenReview ↗

Submission of the Revised Manuscript

Thank you to the reviewers and AC for your thoughtful feedback. We have submitted a revised manuscript, with changes highlighted in blue for ease of review. We would appreciate your assessment of whether our responses address your concerns, as it will help us further refine the manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.