Topological Analysis for Identifying Anomalies in Serverless Platforms

Gianluca Reali, Mauro Femminella · 2026

Read Topological Analysis for Identifying Anomalies in Serverless Platforms by Gianluca Reali, Mauro Femminella (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-10850.txt.

Introduction

Gianluca Reali and Mauro Femminella model information flows in serverless platforms with topology. They ask which parts can be fixed locally, and which persist at every scale. With Hodge decomposition, they split the flows they observe. One part can be fixed locally. The rest are harmonic modes. They argue these flows arise from how functions interact, not from configuration errors. A step-by-step method points to practical remedies. These include dumping effects that contain the harmonic inefficiencies, not a full redesign.

Catalog record

CREATORGianluca Reali, Mauro Femminella
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE62120
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2603.10850

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Topological Analysis for Identifying Anomalies in Serverless Platforms Gianluca Reali, Mauro Femminella Abstract The information flows in serverless platforms are complex and non-conservative. This is a direct result of how independently deployed functions interact under the platform coarse-grained control mechanisms. To manage this complexity, we introduce a topological model for serverless services. Using Hodge decomposition, we can separate observed operational flows into two distinct categories. They include components that can be corrected locally and harmonic modes that persist at any scale. Our analysis reveals that these harmonic flows emerge naturally from different types of inter-function interactions. They should be understood as structural properties of serverless systems, not as configuration errors. Building on this insight, we present an iterative method for analyzing inter-function flows. This method helps deriving practical remediation strategies. One such strategy is the introduction of "dumping effects" to contain harmonic inefficiencies, offering an alternative to completely restructuring the service’s topological model. Our experimental results confirm that this approach can uncover latent architectural structures. I Introduction This paper investigates the operational challenges that frequently arise during the implementation and deployment of complex cloud services using serverless computing platforms [ 38 , 37 ] . Despite the increasing maturity of orchestration technologies and the emergence of numerous design patterns, implementing large-scale, multi-tenant cloud services in Function-as-a-Service (FaaS) mode may generate unforeseen issues that compromise correct service delivery [ 48 , 39 , 9 ] . These complications stem from multiple converging factors whose simultaneous management poses significant complexity. First, serverless architectures typically require coordinating numerous independent or loosely coupled components. Unlike traditional deployment approaches, serverless computing structures application logic as chains of elementary stateless functions, where service emerges from the distributed coordination of functions through the well-known choreography pattern [ 20 ] . Elementary functions execute a piece of code in response to an event and remain active only transiently to serve application needs. This design minimizes network footprint, operational costs, energy consumption, and attack surface. However, these benefits come with trade-offs, such as increased code complexity, debugging difficulties, intricate function orchestration, and additional latency [ 9 , 32 , 30 ] . A second critical factor involves interaction complexities that may generate uncontrolled loops beyond implementers’ oversight [ 4 ] . For instance, if function is triggered by an event in a service and, during its execution, interacts with the same service, generating a new identical event and starting the cycle again, circular processes may trigger unexpectedly. Such behavior introduces delays and malfunctions whose root causes prove difficult to trace. Similarly, compensation mechanisms in design patterns [ 12 ] [ 10 ] introduce reverse action loops and additional invocations. For instance, in e-commerce applications, failed payments trigger compensation chains involving fund verification, card confirmation, and bank communications. Compensation loops are an essential and designed part of Saga transactions in distributed systems [ 18 ] . However, compensation loops can degenerate into implicit loops under specific pathological circumstances. They become implicit loops only when they lose their ability to terminate due to persistent errors, lack of idempotence, or bugs in the retry logic. At that point, they transform from resilience tools into cost and complexity generators. Such loops can be detected and managed when using microservices. Instead, lack of observability in FaaS deployments, due to short function lifetime, may generate uncontrollable cycles that create logical "holes" corresponding to continuous circular data flows. Additional complexity arises from on-demand function activation, introducing cold start latencies absent in traditional deployments [ 25 , 21 , 16 ] . Despite intensive research into mitigation strategies, cold starts remain sometimes unavoidable, particularly during horizontal scaling. These events trigger cascading complications: function A invoking function B during cold start may encounter timeouts, leading to retry attempts that, combined with delayed B activation, produce duplicate executions. This creates topological local curls requiring explicit management. Moreover, a cold start in a compensation loop can contribute to the creation of an implicit loop, or more likely, worsen an existing loop. Although it could not the primary cause of the loop, cold start can acts as a cost multiplier, making the system more vulnerable to anomalous behavior. Further issues include backend consistency problems, cache coherence challenges, and related complications, exemplified in Table I , which will be illustrated in detail below in the paper. TABLE I : Topological Characterization of Cycles in Distributed Serverless Architectures Service / Saga Example Architecture Logical Flow Topological Model Cycle Type Betti Inv. Harmonic Component \beta Reducible Observable Metrics E-commerce checkout Orders \rightarrow Payments \rightarrow Shipping \rightarrow Inventory \looparrowright Orders Nodes = functions, Edges = RPC, 2D workflow cell Closed transactional loop \beta_{0}=1 Global harmonic on L_{0} No (only by breaking sync cycle) P99 end-to-end latency, retry rate, error amplification Saga orchestration (FaaS) API \rightarrow Validate \rightarrow Enrich \rightarrow Persist \looparrowright Validate Nodes = functions, events/RPC, saga surface Compensation cycle \beta_{2}=1 Harmonic on L_{2} Partial (reduce compensations) P99/P999 latency, compensation retries ML training pipeline Ingest \rightarrow Preprocess \rightarrow Train \rightarrow Validate \looparrowright Preprocess Pipeline DAG with feedback edge Structural feedback loop \beta_{1}=1 High harmonic on L_{1} No (only dampenable) Loop period, queue depth, training latency Cold start + retries Invoke A \rightarrow Cold start B \rightarrow Timeout \looparrowright Invoke A Functions, invocation edges Causal time loop \beta_{0}=1 Local harmonic on L_{0} Yes (warm pools, backoff) Cold start latency, retry rate, P99 Fan-out / fan-in timeout Orchestrator \rightarrow Fan-out \rightarrow Wait \rightarrow Merge \looparrowright Orchestrator Fork/join with implicit 2-cell Waiting surface loop \beta_{2}=1 Harmonic on L_{2} Partial (reduce fan-out) Max branch latency, merge P99 Eventual consistency backend Write \rightarrow Read stale \rightarrow Detect \looparrowright Write State graph with async edges Convergence loop \beta_{1}=1 Persistent harmonic on L_{1} No Staleness window, update lag Rate-limited provider Request \rightarrow 429 \rightarrow Backoff \looparrowright Request Endpoint with external control External control loop \beta_{1}=1 Harmonic on L_{1} No 429 rate, retry storms Incoherent caches Read A \rightarrow Update \rightarrow Invalidate \rightarrow Read B \looparrowright Read A Caches as nodes, invalidation faces Drift surface \beta_{2}=1 Harmonic on L_{2} Partial Cache hit ratio, inconsistency rate Partially idempotent workflow F1 (idem) \rightarrow F2 (non-idem) \looparrowright F1 Nodes with side effects Causal duplication loop \beta_{0}=1 Structural harmonic on L_{0} Yes (full idempotency) Duplicate operations Autoscaling feedback Load \rightarrow Timeout \rightarrow Retry \looparrowright Load Groups with feedback edges Instability loop \beta_{1}=1 Harmonic on L_{1} Partial (predictive scaling) Scale oscillation, P99 spikes In this context, the paper demonstrates how algebraic techniques from Topological Signal Processing (TSP) can identify operational criticalities in application execution. Leveraging Hodge theory, which establishes isomorphism between de Rham cohomology classes of differential forms and discrete Laplacian harmonic solutions on graphs, we decompose graph-defined functions into three orthogonal components: gradient, curl, and harmonic [ 42 ] . The latter two components prove particularly relevant for analyzing application execution, as they directly associate with local service saga closures [ 28 ] and structural loops that generate inefficiencies, performance degradation, and energy consumption problems. While Hodge decomposition provides a lens for analyzing flows, its practical efficacy is tied to the metric structure underpinning the orthogonal projection onto gradient, curl, and harmonic components. In many applications, the standard combinatorial graph Laplacian, which typically employs unit edge weights, may fail to capture heterogeneous interaction strengths or domain-specific importance,

Read the complete text