acceptodds
Under review as a conference paper at ICLR 2027

KLD: Kinetic Log-Aware Joint Learning for Dynamic Microservice Diagnosis

Abstract

Dynamic microservice diagnosis enables timely anomaly detection and rapid fault remediation, laying essential foundations for the availability, performance and reliability of microservice systems. Nevertheless, frequent changes of service instances and invocation topologies, noisy and sparse monitoring data, and pervasive metric drift pose significant diagnostic challenges. We present KLD, an evolution- and runtime-dependency-aware framework for joint anomaly detection and root-cause localization in dynamic microservices. Instead of treating multimodal observations as static system states, KLD explicitly models the temporal evolution of runtime signals and fault evidence propagation over execution-aware dependency structures. KLD extracts multi-order temporal and trajectory features from metrics and traces, aligns these features with log semantics via cross-modal interaction and cooperative gating, and propagates fused evidence over runtime dependency graphs reconstructed from distributed traces to model fault cascading. Leveraging margin-based contrastive representation learning, KLD distinguishes root-cause nodes corresponding to injected faults from downstream affected services. Evaluations on SN and TT datasets show that our method achieves detection F1 scores of 0.995 / 0.993 and localization HR@1 of 0.995 / 0.994, outperforming state-of-the-art baselines. Ablation studies validate the effectiveness of each component, and robustness experiments demonstrate stable performance under controlled perturbations on logs, traces and metric baselines.

Primary area:
learning on time series and dynamical systems
Keywords:
Dynamic Microservices, Anomaly Detection, Root Cause Analysis, Kinetic Temporal Modeling, Cross-Modal Representation Learning

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_41NmOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes KLD, a multimodal framework that combines kinetic temporal features, log-guided cross-modal fusion, and graph propagation over trace-derived dependency graphs for joint anomaly detection and root-cause localization in microservices. It reports near-perfect detection and localization results on two benchmarks built from SocialNetwork and TrainTicket.

Critical strengths

A kinetic and trajectory-aware representation of metrics and traces that uses first- and second-order differences and curvature to separate sharp fault changes from gradual drift.

A cross-modal co-gating module that aligns log semantics with metric and trace signals and suppresses noise from routine logs and sparse traces.

A joint training objective that combines anomaly detection, root-cause ranking, and a margin-based contrastive loss to separate root-cause services from downstream affected services.

Critical weaknesses

The presentation needs improvement. Text in several figures is too small to read, especially the pipeline overview in Figure 1 and the screenshots in Figure 4.

All experiments use datasets the authors collected themselves on SocialNetwork and TrainTicket. The paper does not evaluate on any public RCA benchmark, so the results are hard to compare with prior work.

The novelty is limited. Most components, such as temporal differences, cross-attention, depthwise convolution, are standard modules, and the paper does not explain why this particular combination is needed.

The detection and localization tables use different sets of baselines. For example, FAMOS appears only in the detection table, while LagRCA appear only in the localization table, with no explanation for this choice.

Questions for authors

See the above weaknesses

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

This paper is a potential LLM generated paper with clear hallucinations in references and evaluation results, which violates ICLR research integrity policy.

Critical strengths

None

Critical weaknesses

This paper is a potential LLM generated paper with clear hallucinations in references and evaluation results, which violates ICLR research integrity policy.

Questions for authors

This paper is a potential LLM generated paper with clear hallucinations in references and evaluation results, which violates ICLR research integrity policy.

Reviewer_qP7DOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses anomaly detection and root cause localization in dynamic microservice systems by proposing the KLD framework. KLD extracts temporal features from metrics and call chains, including higher-order differences, trajectory orientation, and curvature. It then integrates these features with log semantics through cross-modal interaction and gating mechanisms. The model reconstructs runtime dependencies between services using call chains and propagates fault information along these dependencies. It also employs contrastive learning to distinguish the actual faulty service from affected downstream services, allowing anomaly detection and root cause localization to be performed jointly.

Experiments were conducted on two datasets, SN and TT. KLD achieved anomaly detection F1 scores of 0.995 and 0.993, along with root cause localization HR@1 scores of 0.995 and 0.994, respectively. These results outperform those of the baseline methods. The paper also presents ablation studies and evaluates the model’s robustness by introducing log noise, removing portions of call chains, and simulating metric drift.

Critical strengths
  1. The problem addressed in this paper is of practical significance. In dynamic microservice environments, factors such as topological changes, metric drift, and fault propagation pose substantial challenges for anomaly detection and root cause localization. This paper proposes an approach that integrates temporal features, multi-modal information, and service dependencies to tackle these issues; the overall methodology is sound and holds clear practical value.

  2. The experimental results are competitive. KLD achieves strong performance in both anomaly detection and root cause localization tasks across the SocialNetwork and TrainTicket datasets, demonstrating a significant improvement over existing methods.

Critical weaknesses
  1. The three proposed innovations are not clearly connected to specific research questions or limitations of existing methods. The paper does not explain which problems require these particular designs or why existing temporal modeling, multimodal fusion, and graph propagation methods are insufficient. This makes it difficult to determine what is fundamentally new beyond combining several features and modules.

  2. The ablation study does not isolate the main design components. The individual effects of the first-order differences, second-order differences, trajectory direction, and curvature proxies are not reported. It is therefore unclear which components account for the performance gains and whether some of them provide overlapping information.

  3. The description of KLD as learning topological evolution does not fully match its implementation. KLD uses traces to reconstruct the current runtime dependency graph and then propagates information over that graph. It does not appear to learn how the topology changes over time. “Runtime-observed Dependency-aware Propagation” would describe this operation more accurately.

  4. The robustness experiments introduce log noise, missing call chains, and metric drift into preprocessed features. Some perturbations are also included during both training and testing. This setup mainly measures how well the model adapts to perturbations already seen during training. Applying perturbations only at test time would provide clearer evidence of its ability to handle unseen disturbances.

Questions for authors
  1. Which specific limitations of existing methods do the three innovations proposed in the paper address? Why do these issues necessitate the current design, rather than being resolvable through existing approaches for temporal modeling, cross-modal fusion, or graph-based propagation? This point directly influences my assessment of the paper’s novelty.

  2. Does KLD actually learn the patterns of how the topology evolves over time, or does it simply reconstruct the current dependency graph from traces within each time window? If the latter, could the authors clarify how this differs from "Runtime-observed Dependency-aware Propagation," and justify why it is still characterized as modeling topological evolution?

  3. Could the authors provide ablation results for the first-order difference, second-order difference, trace direction, and the proxy measure for curvature, both when used individually and when combined incrementally? These results are crucial for determining the effectiveness of each design component and whether there is any functional redundancy.

  4. Were the perturbations in the robustness experiments applied to both the training and test sets? Could the authors provide additional results where log noise, missing call chains, or metric drift are introduced only during the testing phase? This would help clarify whether the model is merely adapting to perturbations it encountered during training, or if it is truly capable of handling previously unseen noise.

Reviewer_8YuhOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

KLD jointly performs anomaly detection and root-cause localization using logs, metrics, and distributed traces. It combines difference-based and trajectory-geometric representations, log-conditioned cross-modal co-gating, GAT propagation over the current trace-derived call graph, and supervised detection, localization, and root–non-root separation losses. The paper reports strong window-level results on three injected fault types in SN and TT, supported by ablations, event aggregation, and perturbation experiments.

Critical strengths
  1. The combination of local and dependency evidence is well motivated. Equation (19) concatenates the raw local representation with the graph-propagated representation, retaining a route to the localization head that bypasses neighbor aggregation. This matches the concern that propagation may dilute weak root-cause evidence. Graph-level detection and node-level localization also match the output granularity of the two tasks.
  2. The separation objective addresses a straightforward degenerate solution. L2 normalization in Equations (23)–(24) prevents satisfying the margin merely by increasing embedding norms. Removing this loss while retaining localization cross-entropy is an informative control: Table 3 reports HR@1 dropping to 0.948/0.957 on SN/TT.
  3. The evaluation includes useful controls. Section 4.1.1 specifies splitting complete runs before window generation and fitting normalization on training data only. Comparing dynamic, static, and no-graph settings, and supplementing window-level metrics with event-level results, also helps characterize the approach.
Critical weaknesses
  1. The problem positioning and the advances over prior work could be articulated more clearly. For example, the cited LagRCA already analyzes multi-lag propagation and downstream victim misranking. It would help to explain which remaining limitation KLD addresses and why its design is suited to that limitation.
  2. The implementation and functional interpretation of co-gating are underspecified, and the ablations do not sufficiently isolate the key designs. Equation (12) does not specify the layers, activations, dimensions, or parameter sharing of "f" _"1" and "f" _"2" . The axes retained after pooling in Equation (11), and those processed by the two-dimensional depthwise convolutions, are also unclear. Similarly, Table 8 removes differences and geometry together and reports only SN, where HR@1 falls from 0.9954 to 0.9821.
  3. Baseline task adaptation and supervision conditions are insufficiently documented. FAMOS originally performs multiclass fault-type classification, whereas Table 1 evaluates it as a binary anomaly detector. The paper should explain the introduction of a normal class or detection head, changes to the loss, conversion from request-level to window-level inputs, preservation of Gaussian association, and validation-based threshold selection.
  4. The perturbation and evaluation protocols provide limited support for the broader dynamic-diagnosis claims. Appendix C.4 perturbs trace features only at non-root nodes, using the true root identity to exclude an important failure mode, missing root observations.
Questions for authors
  1. Please provide the metric implementation or per-sample ranks for Table 2 and per-seed event counts, detection counts, and Top-k hits for Table 7 to explain W4.
  2. Please specify the tensor-level implementation of "f" _"1" "/" "f" _"2" , pooling, and convolution, and clarify whether conventional-gate controls and incremental difference/geometry ablations support the claimed mechanisms.
  3. Please document the FAMOS detection adaptation and baseline-specific supervision/prior resources, including whether graph baselines can use the same current trace relationships.
  4. What happens when trace perturbations are selected without root labels? Please clarify clean-training/corrupted-testing performance and evaluation policies for missed detections, incomplete metrics, and roots absent from the graph.
  5. The 10-second windows with 10-second stride in Section 4.1.1 conflict with the description of heavily overlapping windows in Appendix D.1.
  6. In Table 10, “Best HR@1” is below “HR@1 at best F1” for kernel={2,3,5}. Please specify the selection sets and validation/test roles.
  7. Positive-temperature softmax in Equation (21) preserves single-window rankings for fixed logits, whereas training temperature in Equation (22) and averaging probabilities across windows can affect results. Does Table 10 retrain each temperature setting? Equation (25) is the overall training objective and is not subject to this ranking-invariance argument.
  8. Do the latency and memory figures in Table 4 include BERT-large encoding and graph construction? Please use consistent measurement boundaries across methods.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.