State of the Art: Resilience and Scalability Patterns in Distributed Systems

Critical review of indexed literature (2016–2026), prepared as academic support for Volume 4: Scale Patterns

1. Main Research Lines

Line 1: Systematic taxonomies of recovery patterns in microservices. The most direct and most recent reference for this volume is a PRISMA-aligned systematic review (arXiv:2512.16959, December 2025) of recovery patterns in microservices, which analyzed an initial corpus of 412 records in IEEE Xplore, ACM Digital Library, and Scopus, retaining 26 high-quality studies published between 2014 and 2025. It identifies nine recurring resilience themes: circuit breakers, retries with jitter and budget, sagas with compensation, idempotency, bulkheads, adaptive backpressure, observability, and validation through chaos engineering. Notably, these nine themes match, almost one to one, the mechanisms that Chapter 8 of this volume (AOP for agents) and Chapter 5 (cases, degradation budget) propose for error handling in agent swarms, evidence that this book doesn’t invent a new taxonomy but carries over one already empirically validated from the microservices domain to the agent domain.

Line 2: Architectural pattern catalogs for microservices. Taibi et al. (cited in the catalog of the Brazilian Symposium on Software Quality) conducted a systematic mapping study that identified widely adopted architectural patterns in microservices, organized by aspect (migration, orchestration, storage, deployment). Márquez et al. complement it with a systematic review of academic and industry literature on architectural patterns and tactics for microservices. Both directly precede Chapter 8 (Vol. 1) of this series, which adapts that same catalog logic to agent orchestration.

Line 3: Empirical studies of failures in production microservice systems. An empirical study of problems, causes, and solutions in microservice systems (arXiv:2302.01894) cites relevant prior work: Soldani, Tamburri, and Van Den Heuvel (2018, Journal of Systems and Software) carried out a systematic grey literature review on the “pains and gains” of microservices, an important methodological antecedent because, unlike most of the SLRs in this review, it deliberately incorporates non-academic sources (engineering blogs, industry conference publications), recognizing that much of the knowledge of resilience in real production lives outside the peer-reviewed literature, the same reason this book cites engineering blogs from Spotify, Netflix (through its historic Simian Army tool, cited in the 2025 SLR), and other non-academic primary sources.

Line 4: Resilience of deep learning systems to hardware failures. A systematic review of 220 scientific papers published between January 2019 and March 2024 (arXiv:2309.16733) examines the resilience of Deep Learning applications to hardware failures, with a classification framework by failure model and reproducibility. It is an adjacent, not central, line for this volume: it addresses resilience at the level of compute infrastructure, while this book focuses on resilience at the level of logical orchestration, a layer distinction that delimits this volume’s scope.

2. Theoretical Frameworks Used

The microservices resilience literature doesn’t start from a unified theory but from the empirical consolidation of design patterns already practiced in industry before being systematized academically, a “practice precedes theory” pattern consistent with the very genesis of several of the patterns this book documents (the circuit breaker, popularized by engineering practice (Nygard, Release It!, 2018, already cited in this book) before academia). The 2025 SLR introduces a genuinely new theoretical contribution: a Constraint-Aware Decision Matrix that explicitly maps the latency, consistency, and cost trade-offs to appropriate recovery mechanisms, a formal framework this volume doesn’t yet cite but that could strengthen, in a future revision, the degradation policy section of Chapter 5 (Case 8).

3. Methodologies Applied

Systematic review with a PRISMA protocol on primary studies published in indexed engineering databases (IEEE Xplore, ACM DL, Scopus) predominates, complemented (explicitly and with methodological justification, not by accident) by engineering grey literature (technical blogs from Meta, Netflix). Systematic mapping studies (Taibi et al.) differ from SLRs in that they seek an overview and categorization more than a synthesis of quantitative evidence. It is a methodological difference this book doesn’t always make explicit when citing patterns, and one that a future edition could sharpen.

4. Recurring Findings

• The nine resilience themes of the 2025 SLR (Section 1) represent, as far as this review could determine, the most recent and empirically validated catalog of recovery patterns in distributed systems, and retroactively validate the selection of patterns this book adapts for agents.
• Multiple sources (Taibi et al.; Márquez et al.; the empirical study in arXiv:2302.01894) agree that the lack of rigorous systematization of fault-tolerance patterns (beyond knowledge scattered across forums and specialized publications) is a recognized problem in the field, unresolved until the most recent reviews (2024–2025). This book inherits that same risk of dispersion by extending those patterns to a domain (agents) even younger than microservices.

5. Disagreements Among Authors

The disagreement most relevant to this volume is one of scope: the microservices resilience literature (lines 1–3) treats failure as a discrete, observable event (a request that times out, a service that is down); the Deep Learning resilience literature (line 4) treats failure as a silent degradation of the quality of the result, with no explicit error signal. The agentic systems this book describes combine both failure regimes (an agent can fail explicitly (exception, timeout) or silently (a technically valid but incorrect response)), and neither of the two literature traditions reviewed here, separately, fully covers that combined space. The capture aspect of Chapter 8 (Vol. 1) of this book, by normalizing both kinds of failure under a single AgentError type, is a direct response to that gap between traditions.

6. Methodological Limitations of the Literature Reviewed

• Small corpus in the most rigorous SLRs. The 2025 SLR on recovery in microservices retained only 26 studies out of an initial corpus of 412, a low retention rate (~6%) that its own authors attribute to strict quality criteria, but which also limits the generalization of its nine themes as a definitive catalog.
• Almost total absence of academic literature on resilience specific to LLM-based multi-agent systems (as opposed to conventional microservices): none of the sources in this section directly addresses AI agents as the unit of failure; all refer to deterministic microservices or to isolated deep learning models.
• Disconnect between the design pattern literature and the operating cost literature: none of the SLRs reviewed here integrates economic cost (tokens, compute) as a dimension of the resilience decision, something this volume does explicitly by linking the CIU (Chapter 8) with degradation policies.

7. Knowledge Gaps

1. No resilience pattern catalog specific to LLM-based multi-agent systems: the nine themes of the 2025 SLR are the soundest base available, but none was formulated with semantic failure (a plausible but incorrect response), characteristic of LLMs, in mind: the exact gap this book fills by adapting those nine patterns to the agentic context.
2. A lack of research on versioned migrations applied to AI artifacts (prompts, tool schemas): the schema migration literature (Flyway/Liquibase, tools already established in database practice) has, as far as this review could verify, no published academic antecedent extending it to agent compounds; the proposal in Chapter 8 (Vol. 4) of this book is, in that sense, an original contribution and not a synthesis of existing literature.
3. Little empirical research on the exact point at which an agent swarm stops being auditable at scale (the problem that motivates “RACI at scale” in Vol. 3): the multi-agent systems literature treats scale as a performance problem, not a governance one.

8. Relationship to the Book’s Central Thesis

This volume’s thesis (that the resilience patterns that already govern conventional distributed systems carry over, with deliberate adaptation, to agent systems) finds direct empirical support in the most recent and rigorous literature available (the 2025 SLR on recovery in microservices): the nine patterns identified there are not a construction of this book; they are an already validated catalog that this book reuses and extends. This volume’s original contribution lies not in inventing the patterns but in identifying precisely where the semantic failure of agents demands a real adaptation of those patterns (the capture aspect, the degradation policy by orchestration pattern type) and not a direct copy, a distinction the literature reviewed, centered on deterministic microservices, could not anticipate because the problem it solves (LLM-based agents) postdates most of those studies.