# Latency

A computing performance measure of delay between a request and its response, including delay on network connections.

This is one page of public article previews, not the complete archive. Follow Next page to continue. Summaries are not the original full articles.

## Causal Overtaking in Fair Transaction Ordering

DevFeed: [Causal Overtaking in Fair Transaction Ordering](<https://devfeed.tech/articles/causal-overtaking-in-fair-transaction-ordering-146054.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.26100>)

Author: Haozhe Feng, Di Wu, Jian Liu

Published: 2026-11-20T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Cryptography and Security](<https://devfeed.tech/sources/arxiv-cryptography-and-security.md>)

Topics: [ordering](<https://devfeed.tech/topics/ordering.md>), [Simulation](<https://devfeed.tech/topics/simulation.md>), [causality](<https://devfeed.tech/topics/causality.md>), [Latency](<https://devfeed.tech/topics/latency.md>)

Tags: [attack](<https://devfeed.tech/tags/attack.md>), [aws](<https://devfeed.tech/tags/aws.md>), [cs-cr](<https://devfeed.tech/tags/cs-cr.md>), [latency](<https://devfeed.tech/tags/latency.md>), [measurements](<https://devfeed.tech/tags/measurements.md>), [mitigation](<https://devfeed.tech/tags/mitigation.md>), [model](<https://devfeed.tech/tags/model.md>), [network](<https://devfeed.tech/tags/network.md>), [ordering](<https://devfeed.tech/tags/ordering.md>), [protocols](<https://devfeed.tech/tags/protocols.md>), [simulations](<https://devfeed.tech/tags/simulations.md>), [transaction](<https://devfeed.tech/tags/transaction.md>)

### AI overview

The paper identifies causal overtaking in fair transaction ordering: corrupted replicas can submit identical signed copies of a responsive transaction through multiple client entries, letting their first arrivals collectively provide ordering support that no single path achieves. The authors formalize receipt-order protocols including Themis, characterize when the attack is feasible, and report that delayed gossip does not prevent it. Under their network model, opposite required transaction orders can yield identical receipt evidence, so receipt-only mechanisms cannot always preserve earlier-submission priority while preventing responsive overtaking. They propose a mitigation and evaluate the attack and mitigation on Themis using simulations driven by AWS and WonderProxy latency measurements; the abstract does not describe the mitigation's design.

### Source excerpt

arXiv:2611.26100v1 Announce Type: new Abstract: Fair transaction ordering protocols aim to mitigate front-running using replica-local arrival orders, but prior work often assumes a single submission origin for each responsive transaction. We identify causal overtaking, in which corrupted replicas observe a target and submit identical signed copies of a responsive transaction through multiple client entries. The copies' first arrivals at different replicas can collectively provide ordering support that no individual path achieves, allowing the response to execute before its target. We formalize a class of receipt-order protocols, including Themis, characterize the timing and network conditions under which the attack is feasible, and show that delayed gossip does not rule it out. We further prove that, under our network model, histories requiring opposite transaction orders can produce identical receipt evidence. Thus, receipt-only mechanisms cannot always preserve earlier-submission priority while preventing responsive overtaking. We propose a mitigation and analyze its assumptions and limitations. Finally, we evaluate the attack and mitigation on Themis through simulations driven by AWS and WonderProxy latency measurements.

## What Is Agent Tracing? How to See Every Step an Agent Takes

DevFeed: [What Is Agent Tracing? How to See Every Step an Agent Takes](<https://devfeed.tech/articles/what-is-agent-tracing-how-to-see-every-step-an-agent-takes-141927.md>)

Original publisher: [Read original article](<https://montecarlo.ai/blog-agent-tracing>)

Author: Ruslan Sultanov

Published: 2026-11-18T17:20:39Z

Content type: article

Language: en

Sources: [Monte Carlo](<https://devfeed.tech/sources/monte-carlo.md>)

Topics: [ai observability](<https://devfeed.tech/topics/ai-observability.md>), [tracing](<https://devfeed.tech/topics/tracing.md>), [AI Agent](<https://devfeed.tech/topics/ai-agent.md>), [Multi-step AI processes](<https://devfeed.tech/topics/multi-step-ai-processes.md>), [Latency](<https://devfeed.tech/topics/latency.md>)

Tags: [accountability](<https://devfeed.tech/tags/accountability.md>), [agent](<https://devfeed.tech/tags/agent.md>), [agent-spans](<https://devfeed.tech/tags/agent-spans.md>), [agent-tracing](<https://devfeed.tech/tags/agent-tracing.md>), [agent-tracing-vs-logging](<https://devfeed.tech/tags/agent-tracing-vs-logging.md>), [agent-trust](<https://devfeed.tech/tags/agent-trust.md>), [agent-trust-glossary](<https://devfeed.tech/tags/agent-trust-glossary.md>), [ai-agent-tracing](<https://devfeed.tech/tags/ai-agent-tracing.md>), [black-box](<https://devfeed.tech/tags/black-box.md>), [customer](<https://devfeed.tech/tags/customer.md>), [glossary](<https://devfeed.tech/tags/glossary.md>), [how-to](<https://devfeed.tech/tags/how-to.md>), [latency](<https://devfeed.tech/tags/latency.md>), [llm-tracing](<https://devfeed.tech/tags/llm-tracing.md>), [logs](<https://devfeed.tech/tags/logs.md>), [model](<https://devfeed.tech/tags/model.md>), [opentelemetry-genai-semantic-conventions](<https://devfeed.tech/tags/opentelemetry-genai-semantic-conventions.md>), [order](<https://devfeed.tech/tags/order.md>), [request](<https://devfeed.tech/tags/request.md>), [response](<https://devfeed.tech/tags/response.md>), [service](<https://devfeed.tech/tags/service.md>), [tools](<https://devfeed.tech/tags/tools.md>), [trace-ai-agents](<https://devfeed.tech/tags/trace-ai-agents.md>), [tracing](<https://devfeed.tech/tags/tracing.md>), [what-is](<https://devfeed.tech/tags/what-is.md>), [what-is-agent-tracing](<https://devfeed.tech/tags/what-is-agent-tracing.md>)

### AI overview

Agent tracing records an AI agent's full run as linked spans, capturing steps such as planning, retrieval, model calls, and tool use alongside inputs, outputs, timing, and metadata. This makes silent failures, retries, quality issues, and cost or latency drivers easier to investigate than ordinary logs or service-focused distributed traces. The excerpt recommends OpenTelemetry GenAI conventions, trace IDs and parent-span IDs, versioning prompts and models, linking retrieved items to source data, and scoring traces. Capture sensitive prompts carefully: redact personal data and set retention policies. OpenTelemetry conventions remain in development, so pin versions. Traces also cannot reveal upstream data freshness problems without lineage or data monitoring. Teams can store traces in their own warehouse, but need tooling to query, score, and alert on them.

### Source excerpt

An agent tells a customer their order shipped. It didn't. Every service returned 200 OK, latency was normal, and the logs show a request in and a response out. What happened in between, which documents the agent read, which tools it called, what the model was told, is a black box. Agent tracing opens that ...

## Fivetran vs Stitch: Performance, Pricing, and More

DevFeed: [Fivetran vs Stitch: Performance, Pricing, and More](<https://devfeed.tech/articles/fivetran-vs-stitch-performance-pricing-and-more-140231.md>)

Original publisher: [Read original article](<https://wf.fivetran.com/blog/fivetran-vs-stitch>)

Author: Fivetran

Published: 2026-11-18T04:57:01Z

Content type: comparison

Language: en

Sources: [Fivetran Blog](<https://devfeed.tech/sources/fivetran-blog.md>)

Topics: [build vs buy database solutions](<https://devfeed.tech/topics/build-vs-buy-database-solutions.md>), [performance-engineering](<https://devfeed.tech/topics/performance-engineering.md>), [Latency](<https://devfeed.tech/topics/latency.md>), [Service-Level Agreements](<https://devfeed.tech/topics/service-level-agreements.md>)

Tags: [automation](<https://devfeed.tech/tags/automation.md>), [batch](<https://devfeed.tech/tags/batch.md>), [compare](<https://devfeed.tech/tags/compare.md>), [comparison](<https://devfeed.tech/tags/comparison.md>), [compatibility](<https://devfeed.tech/tags/compatibility.md>), [compliance](<https://devfeed.tech/tags/compliance.md>), [connectors](<https://devfeed.tech/tags/connectors.md>), [coverage](<https://devfeed.tech/tags/coverage.md>), [data-modeling](<https://devfeed.tech/tags/data-modeling.md>), [data-pipelines](<https://devfeed.tech/tags/data-pipelines.md>), [data-readiness](<https://devfeed.tech/tags/data-readiness.md>), [dbt](<https://devfeed.tech/tags/dbt.md>), [elt](<https://devfeed.tech/tags/elt.md>), [features](<https://devfeed.tech/tags/features.md>), [fivetran](<https://devfeed.tech/tags/fivetran.md>), [performance](<https://devfeed.tech/tags/performance.md>), [pricing](<https://devfeed.tech/tags/pricing.md>), [stitch](<https://devfeed.tech/tags/stitch.md>), [use-cases](<https://devfeed.tech/tags/use-cases.md>), [vs](<https://devfeed.tech/tags/vs.md>)

### AI overview

Fivetran emphasizes managed operations and broad, in-house-maintained connector coverage, which the text presents as reducing engineering work and supporting scalable, near-real-time syncs. Stitch's Singer-based open framework gives developer-oriented teams more ability to customize connectors, but requires more manual maintenance and scaling; its batch syncs may suit less frequent data needs. The text also says Fivetran standardizes schemas and streamlines post-load transformations, while Stitch leaves more modeling and transformation work to teams. Fivetran is described as having stronger support, documentation, and enterprise security features. Its active-row pricing may be efficient, though complex data flows can make monthly costs harder to predict. The comparison offers no specific prices, and its performance and compliance claims are presented without supporting measurements or citations.

### Source excerpt

Choosing between Fivetran and Stitch for automated ELT? Compare features, pricing, and use cases to find the right platform.

## Distribute Amazon RDS traffic between read replicas using Network Load Balancer

DevFeed: [Distribute Amazon RDS traffic between read replicas using Network Load Balancer](<https://devfeed.tech/articles/distribute-amazon-rds-traffic-between-read-replicas-using-network-load-balancer-138652.md>)

Original publisher: [Read original article](<https://aws.amazon.com/blogs/database/distribute-amazon-rds-traffic-between-read-replicas-using-network-load-balancer/>)

Author: Vijay Chakraborty

Published: 2026-11-17T15:51:53Z

Content type: tutorial

Language: en

Sources: [AWS Database Blog](<https://devfeed.tech/sources/aws-database-blog.md>)

Topics: [Load Balancing](<https://devfeed.tech/topics/load-balancing.md>), [Amazon RDS](<https://devfeed.tech/topics/amazon-rds.md>), [read-replica](<https://devfeed.tech/topics/read-replica.md>), [Virtual Private Network](<https://devfeed.tech/topics/vpn.md>), [Cloud Networking](<https://devfeed.tech/topics/cloud-networking.md>), [Latency](<https://devfeed.tech/topics/latency.md>), [connection draining](<https://devfeed.tech/topics/connection-draining.md>)

Tags: [advanced-300](<https://devfeed.tech/tags/advanced-300.md>), [affinity](<https://devfeed.tech/tags/affinity.md>), [amazon-aurora](<https://devfeed.tech/tags/amazon-aurora.md>), [amazon-cloudwatch](<https://devfeed.tech/tags/amazon-cloudwatch.md>), [amazon-ec2](<https://devfeed.tech/tags/amazon-ec2.md>), [amazon-elastic-container-service](<https://devfeed.tech/tags/amazon-elastic-container-service.md>), [amazon-elastic-kubernetes-service-amazon-eks](<https://devfeed.tech/tags/amazon-elastic-kubernetes-service-amazon-eks.md>), [amazon-eventbridge](<https://devfeed.tech/tags/amazon-eventbridge.md>), [amazon-rds](<https://devfeed.tech/tags/amazon-rds.md>), [amazon-route-53](<https://devfeed.tech/tags/amazon-route-53.md>), [amazon-s3](<https://devfeed.tech/tags/amazon-s3.md>), [aws-cloudformation](<https://devfeed.tech/tags/aws-cloudformation.md>), [aws-direct-connect](<https://devfeed.tech/tags/aws-direct-connect.md>), [aws-lambda](<https://devfeed.tech/tags/aws-lambda.md>), [aws-privatelink](<https://devfeed.tech/tags/aws-privatelink.md>), [aws-site-to-site-vpn](<https://devfeed.tech/tags/aws-site-to-site-vpn.md>), [elastic-load-balancing](<https://devfeed.tech/tags/elastic-load-balancing.md>), [health](<https://devfeed.tech/tags/health.md>), [lambda](<https://devfeed.tech/tags/lambda.md>), [load-balancer](<https://devfeed.tech/tags/load-balancer.md>), [read-replicas](<https://devfeed.tech/tags/read-replicas.md>), [technical-how-to](<https://devfeed.tech/tags/technical-how-to.md>), [traffic](<https://devfeed.tech/tags/traffic.md>)

### AI overview

An NLB can provide one stable endpoint for distributing connections across Amazon RDS read replicas, with health checks and Availability Zone affinity. An EventBridge rule invokes a Lambda function every minute to resolve replica DNS names and synchronize their IPs into the NLB target group; S3 JSON state and retry logic help avoid premature deregistration during transient DNS failures. With zonal DNS affinity enabled and cross-zone load balancing disabled, traffic can remain AZ-local in steady state, reducing inter-AZ transfer. The endpoint also supports hybrid access over Direct Connect or Site-to-Site VPN without PrivateLink or Route 53 Resolver endpoints when a custom domain is unnecessary. The approach works with RDS engines that support read replicas, but the NLB balances connections rather than queries, so long-lived connections can limit redistribution. In EKS, CoreDNS placement can undermine zonal affinity; NodeLocal DNSCache is suggested. Deregistration delay allows existing connections to drain.

### Source excerpt

Learn how to put Amazon RDS read replicas behind a Network Load Balancer with health-aware, Availability Zone affinity routing. An AWS Lambda function keeps the load balancer target group in sync with replica IP addresses, giving clients a single stable endpoint over Direct Connect or Site-to-Site VPN without AWS PrivateLink overhead.

## Why operational visibility is becoming more valuable than performance scores

DevFeed: [Why operational visibility is becoming more valuable than performance scores](<https://devfeed.tech/articles/why-operational-visibility-is-becoming-more-valuable-than-performance-scores-139522.md>)

Original publisher: [Read original article](<https://kinsta.com/blog/operational-visibility-vs-performance-scores/>)

Author: Carlo Daniele

Published: 2026-11-17T12:45:27Z

Content type: article

Language: en

Sources: [WordPress Guides, Tutorials & Tips - Kinsta® Blog](<https://devfeed.tech/sources/wordpress-guides-tutorials-tips-kinsta-blog.md>)

Topics: [Production Engineering](<https://devfeed.tech/topics/production-engineering.md>), [Application Performance Management (APM)](<https://devfeed.tech/topics/apm.md>), [real user monitoring](<https://devfeed.tech/topics/real-user-monitoring.md>), [synthetic monitoring](<https://devfeed.tech/topics/synthetic-monitoring.md>), [Concurrency](<https://devfeed.tech/topics/concurrency.md>), [Latency](<https://devfeed.tech/topics/latency.md>), [Website](<https://devfeed.tech/topics/website.md>)

Tags: [ai](<https://devfeed.tech/tags/ai.md>), [alert](<https://devfeed.tech/tags/alert.md>), [analytics](<https://devfeed.tech/tags/analytics.md>), [api](<https://devfeed.tech/tags/api.md>), [apm](<https://devfeed.tech/tags/apm.md>), [application-performance-monitoring](<https://devfeed.tech/tags/application-performance-monitoring.md>), [automated](<https://devfeed.tech/tags/automated.md>), [automation](<https://devfeed.tech/tags/automation.md>), [background-work](<https://devfeed.tech/tags/background-work.md>), [cache](<https://devfeed.tech/tags/cache.md>), [checkout](<https://devfeed.tech/tags/checkout.md>), [completion](<https://devfeed.tech/tags/completion.md>), [core-web-vitals](<https://devfeed.tech/tags/core-web-vitals.md>), [database](<https://devfeed.tech/tags/database.md>), [operational](<https://devfeed.tech/tags/operational.md>), [performance](<https://devfeed.tech/tags/performance.md>), [visibility](<https://devfeed.tech/tags/visibility.md>)

### AI overview

A Lighthouse score of 98 can accurately describe a fast homepage while missing checkout delays caused by server-side work. In the example, afternoon WooCommerce stock imports overlap with checkout traffic; APM traces reveal repeated database lookups in a custom product-matching function. Batching SKU lookups improves the importer, with staging tests and live timings used to check checkout behavior and stock accuracy. The article distinguishes lab scores, available real-user Core Web Vitals, hosting analytics, logs, APM traces, and job records: each reveals different evidence, and a successful HTTP response does not prove a queued task completed. Kinsta Analytics can show timing, PHP thread-limit events, response codes, and cache patterns; Kinsta APM can trace server-side queries and calls when enabled, but adds overhead and cannot inspect browser-to-payment-service traffic or recover traces from before it was running. Agencies should report checkout behavior during imports, completion and failed rows, recurring incidents, and unresolved decisions, with workload and environment limits in view.

### Source excerpt

Featured image for Why operational visibility is becoming more valuable than performance scores

## Delay-Aware Reinforcement Learning for O-RAN Control under Stochastic Reward Feedback

DevFeed: [Delay-Aware Reinforcement Learning for O-RAN Control under Stochastic Reward Feedback](<https://devfeed.tech/articles/delay-aware-reinforcement-learning-for-o-ran-control-under-stochastic-reward-feedback-138859.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.19344>)

Author: Xingqi Wu, Junaid Farooq, Tao Li, Juntao Chen, Ying Wang

Published: 2026-11-17T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [Reinforcement learning](<https://devfeed.tech/topics/reinforcement-learning.md>), [AI Networking](<https://devfeed.tech/topics/ai-networking.md>), [Latency](<https://devfeed.tech/topics/latency.md>)

Tags: [actions](<https://devfeed.tech/tags/actions.md>), [ai](<https://devfeed.tech/tags/ai.md>), [asynchronous](<https://devfeed.tech/tags/asynchronous.md>), [control](<https://devfeed.tech/tags/control.md>), [convergence](<https://devfeed.tech/tags/convergence.md>), [credit-assignment](<https://devfeed.tech/tags/credit-assignment.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [cs-ma](<https://devfeed.tech/tags/cs-ma.md>), [decision](<https://devfeed.tech/tags/decision.md>), [deep](<https://devfeed.tech/tags/deep.md>), [delay](<https://devfeed.tech/tags/delay.md>), [faster](<https://devfeed.tech/tags/faster.md>), [feedback](<https://devfeed.tech/tags/feedback.md>), [framework](<https://devfeed.tech/tags/framework.md>), [latency](<https://devfeed.tech/tags/latency.md>), [learning](<https://devfeed.tech/tags/learning.md>), [long-short-term-memory](<https://devfeed.tech/tags/long-short-term-memory.md>), [lstm](<https://devfeed.tech/tags/lstm.md>), [markov](<https://devfeed.tech/tags/markov.md>), [module](<https://devfeed.tech/tags/module.md>), [reinforcement-learning](<https://devfeed.tech/tags/reinforcement-learning.md>), [reward](<https://devfeed.tech/tags/reward.md>)

### AI overview

In O-RAN control, stochastic delays can separate an action from the reward used to train the controller, misaligning credit assignment and destabilizing learning. The proposed off-policy deep RL framework predicts delay with an LSTM and realigns rewards over time. In experiments using delay traces from the Open AI Cellular (OAIC) testbed, it reportedly converges faster and reduces latency violations and packet drops compared with conventional RL that ignores feedback delay. The supplied excerpt does not provide quantitative results or further implementation details.

### Source excerpt

arXiv:2611.19344v1 Announce Type: cross Abstract: Reinforcement learning (RL) is increasingly employed for real-time control in Open Radio Access Networks (O-RANs). However, the asynchronous nature of O-RAN feedback introduces stochastic delays between actions and their observable rewards, leading to misaligned credit assignment and unstable training. This letter formulates the O-RAN control problem as a Markov decision process (MDP) with delayed and stochastic rewards, where the true reward corresponding to an action is received after an unknown delay. To address this, we propose a delay-aware deep RL framework that integrates a long short-term memory (LSTM)-based delay predictor and a temporal reward realignment module within an off-policy learning pipeline. Using empirical delay traces collected from the Open AI Cellular (OAIC) testbed, we demonstrate that the proposed framework achieves faster convergence and significantly reduces latency violations and packet drops compared to conventional RL methods that ignore feedback delay. The results confirm that reward-delay compensation is critical for reliable AI-driven O-RAN control.

## PQC Migration: End-to-End Performance Profiling of Post Quantum Cryptographic Algorithms in mTLS

DevFeed: [PQC Migration: End-to-End Performance Profiling of Post Quantum Cryptographic Algorithms in mTLS](<https://devfeed.tech/articles/pqc-migration-end-to-end-performance-profiling-of-post-quantum-cryptographic-algorithms-in-mtls-139318.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.21206>)

Author: Rahul, Manjesh K. Hanawal, Sameer G. Kulkarni, Saravanan V, K. Tirumaleswar Reddy

Published: 2026-11-17T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Cryptography and Security](<https://devfeed.tech/sources/arxiv-cryptography-and-security.md>)

Topics: [Post-quantum cryptography](<https://devfeed.tech/topics/post-quantum-cryptography.md>), [TLS (Transport Layer Security)](<https://devfeed.tech/topics/tls.md>), [Latency](<https://devfeed.tech/topics/latency.md>)

Tags: [1-3](<https://devfeed.tech/tags/1-3.md>), [algorithms](<https://devfeed.tech/tags/algorithms.md>), [code](<https://devfeed.tech/tags/code.md>), [cpu](<https://devfeed.tech/tags/cpu.md>), [cs-cr](<https://devfeed.tech/tags/cs-cr.md>), [energy](<https://devfeed.tech/tags/energy.md>), [hybrid](<https://devfeed.tech/tags/hybrid.md>), [migration](<https://devfeed.tech/tags/migration.md>), [mtls](<https://devfeed.tech/tags/mtls.md>), [network](<https://devfeed.tech/tags/network.md>), [nist](<https://devfeed.tech/tags/nist.md>), [performance](<https://devfeed.tech/tags/performance.md>), [post-quantum](<https://devfeed.tech/tags/post-quantum.md>), [pqc](<https://devfeed.tech/tags/pqc.md>), [profiling](<https://devfeed.tech/tags/profiling.md>), [setup](<https://devfeed.tech/tags/setup.md>), [speed](<https://devfeed.tech/tags/speed.md>), [throughput](<https://devfeed.tech/tags/throughput.md>), [tls](<https://devfeed.tech/tags/tls.md>), [tls-1-3](<https://devfeed.tech/tags/tls-1-3.md>), [transport-layer-security](<https://devfeed.tech/tags/transport-layer-security.md>)

### AI overview

The reported tests found no measurable TLS 1.3 performance degradation from standardized post-quantum algorithms, despite their larger keys. Hybrid algorithms did degrade performance, with the impact becoming more pronounced under adverse network conditions. The authors therefore suggest making hybrid migration sensitive to network conditions. Their setup measured throughput, connection time, energy, and CPU cycles across client-server connections; the supplied excerpt gives no detailed measurements or algorithm-by-algorithm results.

### Source excerpt

arXiv:2611.21206v1 Announce Type: new Abstract: Emerging quantum computers threaten to break the classical key exchange algorithms. To safeguard against these threats, migration to new quantum-safe algorithms is underway and expected to be expedited. Though resistant to attacks, the quantum safe algorithms may impact the real-world performance due to their large key lengths. In this work, we analyze the end-to-end performance of the Transport Layer Security (TLS) protocol with the post-quantum (PQ) and hybrid cryptographic algorithms standardized by NIST. We develop a test setup to evaluate the performance of TLS connections between a client and a server under various network conditions. We focus on the impact of the new algorithms on the speed (throughput, connection time) and sustainability (energy, CPU cycles) of TLS 1.3. Our analysis demonstrates that the standardized post-quantum algorithms do not introduce any measurable performance degradations in TLS 1.3, despite their larger key lengths. However, hybrid algorithms introduce noticeable performance degradation, which becomes increasingly pronounced under adverse network conditions. Our analysis suggests that migration to hybrid algorithms should be ``network condition aware'' as poor network conditions can severely degrade performance in TLS 1.3. We provide anonymised link to the code and a web interface to analyse performance of various combination of algorithms.

## Disaggregated LLM Inference in Space: A Communication Feasibility Study

DevFeed: [Disaggregated LLM Inference in Space: A Communication Feasibility Study](<https://devfeed.tech/articles/disaggregated-llm-inference-in-space-a-communication-feasibility-study-137034.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.20893>)

Author: Arno Troch, Roger Sanchez-Vital, Seyyidahmed Lahmer, Bartek Kozicki, Jeroen Famaey, Michael Peeters

Published: 2026-11-17T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Networking](<https://devfeed.tech/sources/arxiv-networking.md>)

Topics: [LLM optimization](<https://devfeed.tech/topics/llm-optimization.md>), [Large Language Model](<https://devfeed.tech/topics/llm.md>), [communications](<https://devfeed.tech/topics/communications.md>), [Latency](<https://devfeed.tech/topics/latency.md>), [AI Architecture](<https://devfeed.tech/topics/ai-architecture.md>), [Distributed Systems & Parallel Computing](<https://devfeed.tech/topics/distributed-systems-parallel-computing.md>), [Mixture of Experts (MoE)](<https://devfeed.tech/topics/mixture-of-experts-moe.md>), [Data transmission](<https://devfeed.tech/topics/data-transmission.md>)

Tags: [ai-inference](<https://devfeed.tech/tags/ai-inference.md>), [architecture](<https://devfeed.tech/tags/architecture.md>), [arxiv](<https://devfeed.tech/tags/arxiv.md>), [attention](<https://devfeed.tech/tags/attention.md>), [cache](<https://devfeed.tech/tags/cache.md>), [capacity](<https://devfeed.tech/tags/capacity.md>), [communication](<https://devfeed.tech/tags/communication.md>), [compute](<https://devfeed.tech/tags/compute.md>), [cs-ni](<https://devfeed.tech/tags/cs-ni.md>), [data-centers](<https://devfeed.tech/tags/data-centers.md>), [decode](<https://devfeed.tech/tags/decode.md>), [delay](<https://devfeed.tech/tags/delay.md>), [disaggregation](<https://devfeed.tech/tags/disaggregation.md>), [distributed](<https://devfeed.tech/tags/distributed.md>), [experts](<https://devfeed.tech/tags/experts.md>), [fast](<https://devfeed.tech/tags/fast.md>), [frontier](<https://devfeed.tech/tags/frontier.md>), [inference](<https://devfeed.tech/tags/inference.md>), [llm](<https://devfeed.tech/tags/llm.md>), [llm-inference](<https://devfeed.tech/tags/llm-inference.md>), [space](<https://devfeed.tech/tags/space.md>)

### AI overview

The abstract finds that orbital disaggregated LLM inference depends on the serving method's communication pattern. Prefill-decode disaggregation (PDD) transfers a large KV cache once per request, so link capacity is its main constraint; with fast enough links, it can tolerate satellite separations up to 1,000 km. Attention-FFN disaggregation (AFD) exchanges small activations at every mixture-of-experts layer for every generated token, so propagation delay accumulates layer by layer and requires attention and expert instances to remain within about 1 km. The study proposes a hierarchical orbital inference architecture and network design guidelines that also account for ground access latency.

### Source excerpt

arXiv:2611.20893v1 Announce Type: new Abstract: Orbital data centers could enable large-scale AI inference, but efficiently serving a frontier language model to many users may require several compute satellites to cooperate over optical links. While disaggregated inference, which splits large language model (LLM) serving across machines, is well established in terrestrial data centers, its viability under the unique constraints of orbital deployments remains an open question. This article investigates how link capacity and satellite separation affect two types of disaggregated LLM serving, in both compact satellite formations and distributed constellations. Prefill-decode disaggregation (PDD) migrates a large KV cache once per request, while attention-FFN disaggregation (AFD) exchanges small activations at every mixture-of-experts layer for every generated token. The two respond to distance in different ways. PDD is limited mainly by link capacity and, with fast enough links, tolerates separations of up to a thousand kilometers. AFD is limited mainly by propagation delay, which accumulates layer by layer and forces attention and expert instances within about a kilometer of each other. Based on these results, we propose a hierarchical orbital inference architecture and provide network design guidelines for future orbital data centers, including the role of ground access latency.

## Top-Level Operational Equilibrium: When is the response of a distributed system predictable from its own telemetry?

DevFeed: [Top-Level Operational Equilibrium: When is the response of a distributed system predictable from its own telemetry?](<https://devfeed.tech/articles/top-level-operational-equilibrium-when-is-the-response-of-a-distributed-system-predictable-from-its-own-telemetry-136992.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.20780>)

Author: V\\'ictor Casamayor Pujol, Nefeli-Marina Rouska, Boris Sedlak

Published: 2026-11-17T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Distributed Computing](<https://devfeed.tech/sources/arxiv-distributed-computing.md>)

Topics: [observability](<https://devfeed.tech/topics/observability.md>), [Latency](<https://devfeed.tech/topics/latency.md>)

Tags: [aggregate](<https://devfeed.tech/tags/aggregate.md>), [capacity](<https://devfeed.tech/tags/capacity.md>), [change](<https://devfeed.tech/tags/change.md>), [control](<https://devfeed.tech/tags/control.md>), [cs-dc](<https://devfeed.tech/tags/cs-dc.md>), [cs-pf](<https://devfeed.tech/tags/cs-pf.md>), [distributed](<https://devfeed.tech/tags/distributed.md>), [load](<https://devfeed.tech/tags/load.md>), [operational](<https://devfeed.tech/tags/operational.md>), [prediction](<https://devfeed.tech/tags/prediction.md>), [response](<https://devfeed.tech/tags/response.md>), [stage](<https://devfeed.tech/tags/stage.md>), [telemetry](<https://devfeed.tech/tags/telemetry.md>)

### AI overview

The paper defines top-level operational equilibrium (TLOE) to identify when passive telemetry can predict a running distributed system's response to small control changes. TLOE requires stationarity with time-scale separation, a top-level aggregate sufficient for the declared controls, and time-reversibility of that aggregate. When the first two conditions hold, a steady-state fluctuation-response identity gives the response, using total work in progress for load changes and stage occupancy for capacity changes. For Poisson input to chains of exponential single-server stages, TLOE holds for load control at every stable throughput, but not for capacity control at the first or last stage. Reference-model results also show that a per-stage conjugate can overestimate early response about threefold; interstage blocking weakly breaks both conditions, and a hysteretic autoscaler has no admissible aggregate at computed reaction times. The paper proposes tests for the conditions and reports required window lengths on simulated telemetry. Its evidence is limited to reference models; real telemetry remains future work.

### Source excerpt

arXiv:2611.20780v1 Announce Type: new Abstract: Distributed systems deployed across the edge-cloud continuum are adapted continuously: scaled, re-routed, given more or less capacity. Each adaptation rests on an expectation of how the system will respond, yet operators cannot tell from the telemetry they already collect whether that response is predictable. We define top-level operational equilibrium (TLOE), a notion of equilibrium for a running system, stated relative to the controls its operator declares. TLOE has three conditions: stationarity with time-scale separation, sufficiency of a top-level aggregate for the controls, and time-reversibility of that aggregate. We prove that an aggregate satisfying the last two exists if and only if the minimal sufficient statistic of the controls is time-reversible, so the coarsest admissible aggregate is derived rather than chosen. Under the first two conditions, the response to a small control change follows from passive telemetry through the steady-state fluctuation-response identity, with a conjugate observable fixed by the type of control: the total work in progress for a change of load, the occupancy of a stage for a change of its capacity. For chains of exponential single-server stages with Poisson input, TLOE holds exactly for load control at every stable throughput, but never for a control on the capacity of the first or last stage. On reference models, a per-stage conjugate overestimates the early response about threefold, blocking between stages breaks both conditions weakly, and a hysteretic autoscaler leaves no admissible aggregate at any reaction time computed. We specify a test for each condition, show that a static sufficiency test with an estimated top-level conjugate has no power, and quantify on simulated telemetry the window lengths the prediction and the reversibility test need. The evidence is on reference models; real telemetry is left to future work.

## Latency, Speech Recognition and Turn-Taking in Voice Agents for Spanish Phone Negotiation: An Externally Measured Comparison of Pipecat, LiveKit and GPT-Live

DevFeed: [Latency, Speech Recognition and Turn-Taking in Voice Agents for Spanish Phone Negotiation: An Externally Measured Comparison of Pipecat, LiveKit and GPT-Live](<https://devfeed.tech/articles/latency-speech-recognition-and-turn-taking-in-voice-agents-for-spanish-phone-negotiation-an-externally-measured-comparison-of-pipecat-livekit-and-gpt-live-137399.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.19118>)

Author: Jovan Popoca Moran, Eduardo de Avila-Armenta, Ra\\'ul Alejandro Aquino-Ibarra, Jorge Octavio Nicol\\'as-D\\'iaz, Umberto Leon-Dominguez

Published: 2026-11-17T05:00:00Z

Content type: comparison

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>)

Topics: [Speech Processing](<https://devfeed.tech/topics/speech-processing.md>), [Latency](<https://devfeed.tech/topics/latency.md>), [livekit](<https://devfeed.tech/topics/livekit.md>), [asr](<https://devfeed.tech/topics/asr.md>)

Tags: [agents](<https://devfeed.tech/tags/agents.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [cs-cl](<https://devfeed.tech/tags/cs-cl.md>), [cs-sd](<https://devfeed.tech/tags/cs-sd.md>), [gpt-live](<https://devfeed.tech/tags/gpt-live.md>), [latency](<https://devfeed.tech/tags/latency.md>), [livekit](<https://devfeed.tech/tags/livekit.md>), [negotiation](<https://devfeed.tech/tags/negotiation.md>), [phone](<https://devfeed.tech/tags/phone.md>), [pipecat](<https://devfeed.tech/tags/pipecat.md>), [speech-recognition](<https://devfeed.tech/tags/speech-recognition.md>)

### AI overview

In simulated Spanish freight-rate calls, GPT-Live had the lowest median listener-measured response latency (2,444.8 ms), but its 171.2 ms advantage over Pipecat was not statistically conclusive; LiveKit measured 3,090.0 ms. Tool calls added 0.9-1.1 seconds across all three architectures, and framework-reported latency understated listener wait. GPT-Live also had lower word and amount error rates than the two cascades. With default settings, Pipecat and LiveKit often treated an overlapping "ok" as the caller's answer, sometimes recording agreement; no such case occurred for GPT-Live in the reported calls. The study recommends measuring latency from the listener's side and adding safeguards for amounts and backchannels in cascades. Results come from 77 calls per architecture using recordings from one speaker without a telephone network, so they do not establish performance in live phone negotiations.

### Source excerpt

arXiv:2611.19118v1 Announce Type: cross Abstract: A voice agent that negotiates by phone must answer quickly, understand spoken amounts and cope with a caller who talks over it. We compared three architectures for such an agent, negotiating freight rates in Spanish: two cascades (speech recognition, language model and speech synthesis) orchestrated by Pipecat and by LiveKit Agents, and GPT-Live, an OpenAI speech model that listens and speaks without separate stages. A program played the carrier with the same 47 recordings of one speaker for all agents, without the telephone network, in 77 calls per architecture over five negotiation scenarios, and measured latency on the listener's side with a validated instrument. Median latency to the content response was 2,616.0 ms (Pipecat), 3,090.0 ms (LiveKit) and 2,444.8 ms (GPT-Live); the confidence interval of GPT-Live's 171.2 ms advantage over Pipecat includes zero. According to its own timestamps, GPT-Live began to answer 800 ms after the end of the line, close to the 0.8 s published by OpenAI; most of the remaining wait, about 1,367 ms and nearly constant, is the audio's round trip through OpenAI's service. A tool call added 0.9-1.1 s in every architecture, more than any difference between them, and the frameworks' self-reported latency fell short of the listener's wait by 9% (Pipecat) and 22% (LiveKit). Word error rates were low (6.03, 4.93 and 1.47%), but the error on amounts was 13.01% and 12.61% in the cascades, against 2.86% in GPT-Live. With default settings, Pipecat and LiveKit stopped at almost every backchannel and often took it as the carrier's answer: an "ok" spoken over the agent ended in a registered agreement in 3 of 15 Pipecat calls, 2 of 15 LiveKit calls and no GPT-Live call (descriptive counts, not tested). Latency should therefore be measured on the listener's side, and a cascade needs safeguards for spoken amounts and backchannels.

## Google Adds Confidential TPU 8tx and Dual-Chip TPU 8i Max to Its Eighth-Generation TPU Lineup

DevFeed: [Google Adds Confidential TPU 8tx and Dual-Chip TPU 8i Max to Its Eighth-Generation TPU Lineup](<https://devfeed.tech/articles/google-adds-confidential-tpu-8tx-and-dual-chip-tpu-8i-max-to-its-eighth-generation-tpu-lineup-132451.md>)

Original publisher: [Read original article](<https://www.storagereview.com/news/google-tpu-8tx-tpu-8i-max-confidential-computing-dual-chip>)

Author: Brian Beeler

Published: 2026-11-13T14:59:38Z

Content type: news

Language: en

Sources: [StorageReview.com](<https://devfeed.tech/sources/storagereview-com.md>)

Topics: [Google](<https://devfeed.tech/topics/google.md>), [Confidential Computing](<https://devfeed.tech/topics/confidential-computing.md>), [Gke](<https://devfeed.tech/topics/gke.md>), [Chip design](<https://devfeed.tech/topics/chip-design.md>), [Latency](<https://devfeed.tech/topics/latency.md>), [Google Cloud Platform (GCP)](<https://devfeed.tech/topics/google-cloud.md>)

Tags: [ai](<https://devfeed.tech/tags/ai.md>), [amd](<https://devfeed.tech/tags/amd.md>), [amd-epyc](<https://devfeed.tech/tags/amd-epyc.md>), [attestation](<https://devfeed.tech/tags/attestation.md>), [bandwidth](<https://devfeed.tech/tags/bandwidth.md>), [compute-density](<https://devfeed.tech/tags/compute-density.md>), [confidential-computing](<https://devfeed.tech/tags/confidential-computing.md>), [cryptographic-attestation](<https://devfeed.tech/tags/cryptographic-attestation.md>), [encryption](<https://devfeed.tech/tags/encryption.md>), [enterprise](<https://devfeed.tech/tags/enterprise.md>), [generation](<https://devfeed.tech/tags/generation.md>), [google](<https://devfeed.tech/tags/google.md>), [inference](<https://devfeed.tech/tags/inference.md>), [long-context](<https://devfeed.tech/tags/long-context.md>), [package](<https://devfeed.tech/tags/package.md>), [silicon](<https://devfeed.tech/tags/silicon.md>), [tpu](<https://devfeed.tech/tags/tpu.md>), [tpu-8i](<https://devfeed.tech/tags/tpu-8i.md>), [tpu-8i-max](<https://devfeed.tech/tags/tpu-8i-max.md>), [tpu-8t](<https://devfeed.tech/tags/tpu-8t.md>), [tpu-8tx](<https://devfeed.tech/tags/tpu-8tx.md>), [training](<https://devfeed.tech/tags/training.md>)

### AI overview

Google's TPU 8tx and TPU 8i Max adapt existing eighth-generation chips for different workloads. TPU 8tx uses TPU 8t silicon with 5th Generation AMD EPYC host CPUs, hardware confidential computing, and the Boardfly network; Google reports up to 2x lower latency than TPU 8t with near-zero performance penalty. It targets secure inference and small-scale training on sensitive data. TPU 8i Max packages two TPU 8i chips together, doubling compute, SRAM and HBM bandwidth; its 768MB SRAM figure follows from TPU 8i's listed 384MB, but Google has not stated HBM capacity. Google also announced TPU Sync, Cloud Pathways availability on GKE, and GKE dynamic slicing, alongside claimed performance and recovery gains. The text gives no launch dates or pricing for the two new systems.

### Source excerpt

Google added two systems to its eighth-generation TPU lineup: TPU 8tx, a confidential computing system built on TPU 8t silicon, and TPU 8i Max, which puts two TPU 8i chips on a single package for long-context inference. They join the TPU 8t training chip and TPU 8i inference chip Google introduced at Cloud Next in The post Google Adds Confidential TPU 8tx and Dual-Chip TPU 8i Max to Its Eighth-Generation TPU Lineup appeared first on StorageReview.com.

## Is Gemini 3.8 Flash a new workhorse? Part 2: Priority inference and Provisional throughput.

DevFeed: [Is Gemini 3.8 Flash a new workhorse? Part 2: Priority inference and Provisional throughput.](<https://devfeed.tech/articles/is-gemini-3-8-flash-a-new-workhorse-part-2-priority-inference-and-provisional-throughput-131169.md>)

Original publisher: [Read original article](<https://medium.com/google-developer-experts/is-gemini-3-8-flash-a-new-workhorse-part-2-priority-inference-and-provisional-throughput-8d10006853b7?source=rss----a67bd6fa7d58---4>)

Author: Artem Nikulchenko

Published: 2026-11-12T15:44:52Z

Content type: news

Language: en

Sources: [Google Developer Experts](<https://devfeed.tech/sources/google-developer-experts-medium.md>)

Topics: [LLM optimization](<https://devfeed.tech/topics/llm-optimization.md>), [Latency](<https://devfeed.tech/topics/latency.md>), [performance-engineering](<https://devfeed.tech/topics/performance-engineering.md>), [API Business Models](<https://devfeed.tech/topics/api-business-models.md>)

Tags: [api](<https://devfeed.tech/tags/api.md>), [application](<https://devfeed.tech/tags/application.md>), [billing](<https://devfeed.tech/tags/billing.md>), [flash](<https://devfeed.tech/tags/flash.md>), [gemini](<https://devfeed.tech/tags/gemini.md>), [gemini-3-8-flash](<https://devfeed.tech/tags/gemini-3-8-flash.md>), [genai](<https://devfeed.tech/tags/genai.md>), [google-gemini](<https://devfeed.tech/tags/google-gemini.md>), [inference](<https://devfeed.tech/tags/inference.md>), [infrastructure](<https://devfeed.tech/tags/infrastructure.md>), [latency](<https://devfeed.tech/tags/latency.md>), [performance](<https://devfeed.tech/tags/performance.md>), [priority](<https://devfeed.tech/tags/priority.md>), [provisioned-throughput](<https://devfeed.tech/tags/provisioned-throughput.md>), [reliability](<https://devfeed.tech/tags/reliability.md>), [standard](<https://devfeed.tech/tags/standard.md>), [throughput](<https://devfeed.tech/tags/throughput.md>), [time-to-first-token](<https://devfeed.tech/tags/time-to-first-token.md>), [vertex-ai](<https://devfeed.tech/tags/vertex-ai.md>)

### AI overview

The tests found that Gemini 3.8 Flash's Priority inference and Provisioned Throughput generally reduce response times and variability compared with Standard, but neither prevents long outliers or "freezes." Priority was faster on average than Standard, while Provisioned Throughput changed average latency little and narrowed the spread; long-prompt outliers still occurred. In 24-hour Provisioned Throughput tests, first-token time at Low effort did not depend on prompt size, while full response time generally increased with prompt size. The author describes Standard as flexible pay-per-use, Priority as a higher-priced pay-per-use tier with overflow downgraded to Standard, and Provisioned Throughput as a fixed-term commitment with overflow sent to a lower tier. Under an idealized usage estimate, Provisioned Throughput cost more than twice Priority for a week, but about 25% more for a month; fluctuating demand and idle capacity could raise real costs. The author suggests covering stable demand with provisioned capacity and allowing peaks to spill to lower tiers. Results are specific to the reported tests, and outliers remained across all three models.

### Source excerpt

Is Gemini 3.8 Flash a new workhorse? Part 2. In Part 1, we discovered that Gemini 3.8 Flash has a large spread when it comes to response time. In this part, I want to share my investigation of it further. First of all, let's see if this delay is time-dependent. In the next section, we are running Gemini 3.8 Flash with Low effort on three prompt sizes. Time to first token by hour: Time to first token by hour Total response time by hour: Total response time by hour Based on those results, we can see that USA morning hours have the biggest spread. But unfortunately, no particular time is protected against unexpected "freezes". Considering that "freezes" are more prevalent during the USA morning hours, one can assume that freezes are correlated with higher load on infrastructure serving Gemini 3.8 Flash, since this model is currently the default model for Gemini Workspace and personal Gemini. Priority inference To battle the problem above, GCP has introduced a concept of Priority inference: Highest criticality: Priority requests get the highest criticality, leading to higher reliability, even during peak load. Graceful downgrade: If your traffic exceeds your Priority limits, overflow requests are automatically served at the Standard tier instead of failing. This keeps your application online and helps to ensure business continuity. Transparent response: The API response indicates which tier served your request, giving you full visibility into your performance and billing. Let's see if switching to Priority inference will help with response time. During our test, we need to monitor to ensure we do not exceed Priority limits. For this test, we are running the same prompts at the same time (making parallel calls) on Gemini 3.8 Flash on Standard inference and Priority inference using Low effort. Time to first token: Standard vs Priority Time to first token: Standard vs Priority Total response time: Standard vs Priority Total response time: Standard vs Priority As expected,

## Tail latency amplification: why p99 of one service becomes p50 of the aggregate

DevFeed: [Tail latency amplification: why p99 of one service becomes p50 of the aggregate](<https://devfeed.tech/articles/tail-latency-amplification-why-p99-of-one-service-becomes-p50-of-the-aggregate-131023.md>)

Original publisher: [Read original article](<https://ankit-rana.com/logs/74-tail-latency-amplification/>)

Author: hello@ankit-rana.com

Published: 2026-11-12T00:00:00Z

Content type: article

Language: en

Sources: [Ankit Rana | Mechanical Sympathy](<https://devfeed.tech/sources/ankit-rana-mechanical-sympathy.md>)

Topics: [Latency](<https://devfeed.tech/topics/latency.md>), [least connections](<https://devfeed.tech/topics/least-connections.md>)

Tags: [aggregate](<https://devfeed.tech/tags/aggregate.md>), [application](<https://devfeed.tech/tags/application.md>), [architecture](<https://devfeed.tech/tags/architecture.md>), [arithmetic](<https://devfeed.tech/tags/arithmetic.md>), [backend](<https://devfeed.tech/tags/backend.md>), [backends](<https://devfeed.tech/tags/backends.md>), [budget](<https://devfeed.tech/tags/budget.md>), [capacity](<https://devfeed.tech/tags/capacity.md>), [churn](<https://devfeed.tech/tags/churn.md>), [code](<https://devfeed.tech/tags/code.md>), [component](<https://devfeed.tech/tags/component.md>), [delay](<https://devfeed.tech/tags/delay.md>), [distributed-systems](<https://devfeed.tech/tags/distributed-systems.md>), [fan-out](<https://devfeed.tech/tags/fan-out.md>), [latency](<https://devfeed.tech/tags/latency.md>), [load](<https://devfeed.tech/tags/load.md>), [observability](<https://devfeed.tech/tags/observability.md>), [p99](<https://devfeed.tech/tags/p99.md>), [parallel](<https://devfeed.tech/tags/parallel.md>), [performance](<https://devfeed.tech/tags/performance.md>), [probability](<https://devfeed.tech/tags/probability.md>), [request](<https://devfeed.tech/tags/request.md>), [service](<https://devfeed.tech/tags/service.md>), [tail-latency](<https://devfeed.tech/tags/tail-latency.md>)

### AI overview

In a fan-out request that waits for every backend, aggregate latency is set by the slowest response: with 60 independent calls each slow 1% of the time, about 45% of aggregate requests encounter a slow call. The excerpt recommends tracking per-request maximum latency and identifying the slowest backend, since averages can hide the delay users experience. Mitigations include reducing fan-out, hedging to another replica after a delay (with bounded extra load, safe-to-duplicate requests, and enough capacity), returning partial results where correctness allows, propagating deadlines, and routing away from slow replicas. Correlated slowdowns cluster in time; hedging helps when the delay is local to one replica.

### Source excerpt

A request that waits on N backends completes at the speed of the slowest one, so the probability of hitting a slow response is one minus the probability of avoiding it N times. At a hundred parallel calls to a service whose p99 is 100ms, roughly 63 percent of requests contain at least one 100ms response, which turns a rare event at the component level into the common case at the aggregate. Sending a duplicate request after a short delay collapses that tail for a few percent extra load.

## SLO vs SLI vs SLA

DevFeed: [SLO vs SLI vs SLA](<https://devfeed.tech/articles/slo-vs-sli-vs-sla-126753.md>)

Original publisher: [Read original article](<https://blog.levelupcoding.com/p/slo-vs-sli-vs-sla>)

Author: Nikki Siapno

Published: 2026-11-10T15:49:26Z

Content type: article

Language: en

Sources: [Level Up Coding System Design Newsletter](<https://devfeed.tech/sources/level-up-coding-system-design-newsletter.md>)

Topics: [Service-Level Agreements](<https://devfeed.tech/topics/service-level-agreements.md>), [Production Engineering](<https://devfeed.tech/topics/production-engineering.md>), [observability](<https://devfeed.tech/topics/observability.md>), [performance-engineering](<https://devfeed.tech/topics/performance-engineering.md>), [Latency](<https://devfeed.tech/topics/latency.md>)

Tags: [actions](<https://devfeed.tech/tags/actions.md>), [aggregation](<https://devfeed.tech/tags/aggregation.md>), [ai](<https://devfeed.tech/tags/ai.md>), [api](<https://devfeed.tech/tags/api.md>), [architecture](<https://devfeed.tech/tags/architecture.md>), [availability](<https://devfeed.tech/tags/availability.md>), [b2b](<https://devfeed.tech/tags/b2b.md>), [breach](<https://devfeed.tech/tags/breach.md>), [budget](<https://devfeed.tech/tags/budget.md>), [burn](<https://devfeed.tech/tags/burn.md>), [business](<https://devfeed.tech/tags/business.md>), [calendar](<https://devfeed.tech/tags/calendar.md>), [capacity](<https://devfeed.tech/tags/capacity.md>), [change](<https://devfeed.tech/tags/change.md>), [checkout](<https://devfeed.tech/tags/checkout.md>), [client](<https://devfeed.tech/tags/client.md>), [reliability](<https://devfeed.tech/tags/reliability.md>), [service](<https://devfeed.tech/tags/service.md>), [slo](<https://devfeed.tech/tags/slo.md>), [target](<https://devfeed.tech/tags/target.md>)

### AI overview

SLIs record user-facing service performance, SLOs set acceptable internal targets, and SLAs define contractual commitments and remedies. The article explains that measurement definitions and collection points affect SLI results, while an SLO's error budget and burn rate can guide release and incident decisions. It recommends keeping an intentional margin between internal targets and customer guarantees; dependencies, very tight targets, low traffic, and too many SLOs can limit usefulness or create misleading alerts.

### Source excerpt

Your service can meet its SLA and still miss its reliability target. Here's why.

## Reducing Cold-Start Latency on the WebAssembly Edge via Locality and Network-Aware Scheduling

DevFeed: [Reducing Cold-Start Latency on the WebAssembly Edge via Locality and Network-Aware Scheduling](<https://devfeed.tech/articles/reducing-cold-start-latency-on-the-webassembly-edge-via-locality-and-network-aware-scheduling-126805.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.12620>)

Author: Michiel Van Kenhove, Emiel Sucaet, Jos\\'e Santos, Filip De Turck, Bruno Volckaert, Merlijn Sebrechts

Published: 2026-11-10T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Networking](<https://devfeed.tech/sources/arxiv-networking.md>), [arXiv: Software Engineering](<https://devfeed.tech/sources/arxiv-software-engineering.md>)

Topics: [WebAssembly](<https://devfeed.tech/topics/web-assembly.md>), [scheduling](<https://devfeed.tech/topics/scheduling.md>), [Latency](<https://devfeed.tech/topics/latency.md>), [wasmCloud](<https://devfeed.tech/topics/wasmcloud.md>), [round robin](<https://devfeed.tech/topics/round-robin.md>)

Tags: [artifacts](<https://devfeed.tech/tags/artifacts.md>), [cache](<https://devfeed.tech/tags/cache.md>), [cold-start](<https://devfeed.tech/tags/cold-start.md>), [cs-ni](<https://devfeed.tech/tags/cs-ni.md>), [cs-se](<https://devfeed.tech/tags/cs-se.md>), [e2e](<https://devfeed.tech/tags/e2e.md>), [edge](<https://devfeed.tech/tags/edge.md>), [edge-computing](<https://devfeed.tech/tags/edge-computing.md>), [hybrid](<https://devfeed.tech/tags/hybrid.md>), [latency](<https://devfeed.tech/tags/latency.md>), [locality](<https://devfeed.tech/tags/locality.md>), [network](<https://devfeed.tech/tags/network.md>), [raspberry-pi](<https://devfeed.tech/tags/raspberry-pi.md>), [round-robin](<https://devfeed.tech/tags/round-robin.md>), [routing](<https://devfeed.tech/tags/routing.md>), [scheduler](<https://devfeed.tech/tags/scheduler.md>), [scheduling](<https://devfeed.tech/tags/scheduling.md>), [wasm](<https://devfeed.tech/tags/wasm.md>), [webassembly](<https://devfeed.tech/tags/webassembly.md>), [worker](<https://devfeed.tech/tags/worker.md>)

### AI overview

A hybrid scheduler for a WebAssembly edge orchestrator combines worker cache state with network proximity to reduce cold-start and routing delays. On a Raspberry Pi testbed with emulated regional latency, it reduced mean end-to-end latency by up to 40.8% versus round-robin. Locality-aware placement was most effective for large or frequently requested artifacts; network-aware placement mattered more for small artifacts when routing latency had a larger relative impact. Combining both signals produced the lowest latency across the evaluated workloads.

### Source excerpt

arXiv:2611.12620v1 Announce Type: new Abstract: Edge computing places computation closer to users and data sources, but introduces heterogeneous hardware, limited resources, and variable network conditions. WebAssembly (Wasm) is increasingly considered for these environments because modules are compact, portable, and quick to start. However, cold-start latency remains a challenge: when a worker node does not yet hold the required artifact, it must fetch it before execution, adding transfer overhead that directly increases the end-to-end (E2E) latency, especially for large artifacts or slow network links. At the same time, the geographical distance between the requester and the assigned worker independently contributes to latency through network routing overhead. Together, artifact locality and network proximity do not always point to the same placement decision, and naive scheduling strategies fail to account for either. This work extends a Wasm-native edge orchestrator with state-aware scheduling strategies that jointly optimize for both objectives. Evaluation on a Raspberry Pi testbed with emulated regional network latency shows that a hybrid scheduler combining cache-state and network-proximity signals reduces mean E2E latency by up to 40.8% compared to a default round-robin baseline. Artifact locality-aware placement is most effective when artifacts are large or repeatedly requested, while network-aware placement dominates when artifacts are small and routing latency has a larger relative effect, but combining both objectives consistently yields the lowest E2E latency across diverse edge workloads.

## Beyond Query Similarity: Safety-Conditioned Cache Routing for LLM Response Reuse

DevFeed: [Beyond Query Similarity: Safety-Conditioned Cache Routing for LLM Response Reuse](<https://devfeed.tech/articles/beyond-query-similarity-safety-conditioned-cache-routing-for-llm-response-reuse-125939.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.10418>)

Author: Shilong Pan, Zhiliang Tian, Weiguo Chen, Zihan Chen, DongSheng Li

Published: 2026-11-10T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Artificial Intelligence](<https://devfeed.tech/sources/arxiv-artificial-intelligence.md>)

Topics: [Cache](<https://devfeed.tech/topics/cache.md>), [LLM security](<https://devfeed.tech/topics/llm-security.md>), [Large Language Model](<https://devfeed.tech/topics/llm.md>), [Latency](<https://devfeed.tech/topics/latency.md>)

Tags: [budget](<https://devfeed.tech/tags/budget.md>), [cache](<https://devfeed.tech/tags/cache.md>), [caching](<https://devfeed.tech/tags/caching.md>), [cs-ai](<https://devfeed.tech/tags/cs-ai.md>), [guard](<https://devfeed.tech/tags/guard.md>), [inference-efficiency](<https://devfeed.tech/tags/inference-efficiency.md>), [llm](<https://devfeed.tech/tags/llm.md>), [llm-inference](<https://devfeed.tech/tags/llm-inference.md>), [query](<https://devfeed.tech/tags/query.md>), [response-caching](<https://devfeed.tech/tags/response-caching.md>), [retrieval](<https://devfeed.tech/tags/retrieval.md>), [reuse](<https://devfeed.tech/tags/reuse.md>), [risk](<https://devfeed.tech/tags/risk.md>), [routing](<https://devfeed.tech/tags/routing.md>), [routing-policy](<https://devfeed.tech/tags/routing-policy.md>), [safety](<https://devfeed.tech/tags/safety.md>), [similarity](<https://devfeed.tech/tags/similarity.md>), [uncertainty](<https://devfeed.tech/tags/uncertainty.md>), [utility](<https://devfeed.tech/tags/utility.md>), [validation](<https://devfeed.tech/tags/validation.md>), [verification](<https://devfeed.tech/tags/verification.md>)

### AI overview

SafeCache treats semantic response reuse as a safety-conditioned routing decision because similar queries can have different safety conditions. It separates candidate retrieval from approval, compares each cached response's safety conditions with the new request, and combines mismatches into a risk score. A separate contradiction guard catches explicit conflicts, while matching-uncertainty estimates help route requests to direct reuse, selective verification, or regeneration. The policy is calibrated to a required unsafe-reuse target and verifier-call budget. The abstract reports a better safety-utility trade-off than evaluated baselines, with safer reuse at comparable cache utility; it does not give quantitative results or details of the experiments.

### Source excerpt

arXiv:2611.10418v1 Announce Type: new Abstract: Semantic response caching reuses a cached response for a semantically similar query to improve LLM inference efficiency. However, query similarity alone can lead to unsafe reuse when safety conditions change. Existing approaches improve cache safety through stricter similarity checks, context-aware validation, or response verification, yet they may still miss safety conflicts between the cached response and the new request or introduce unnecessary verification overhead. We argue that safe and efficient semantic caching requires a reuse policy that makes the cached response's safety conditions explicit and uses them to guide the reuse decision with a balance between safety and utility. In this paper, we propose SafeCache, a safety-aware framework that separates candidate retrieval from safety approval and formulates semantic response reuse as a safety-conditioned routing problem. SafeCache matches each cached-response condition against the new request and aggregates the per-condition mismatches into an overall risk score. A separate contradiction guard prevents explicit conflicts from being diluted by strong matches. It further estimates matching uncertainty and calibrates the routing policy under a required unsafe-reuse target and a verifier-call budget to choose among direct reuse, selective verification, or regeneration. Experiments show that SafeCache achieves the better safety-utility trade-off among the evaluated baselines, enabling safer response reuse at comparable cache utility.

## Google claims its new AI software cuts inference latency by over 90% -- but there's a catch

DevFeed: [Google claims its new AI software cuts inference latency by over 90% -- but there's a catch](<https://devfeed.tech/articles/google-claims-its-new-ai-software-cuts-inference-latency-by-over-90-but-there-s-a-catch-124172.md>)

Original publisher: [Read original article](<https://thenewstack.io/google-tpu-sync-inference/>)

Author: Amanda Caswell

Published: 2026-11-09T18:40:01Z

Content type: news

Language: en

Sources: [The New Stack](<https://devfeed.tech/sources/the-new-stack.md>)

Topics: [performance-engineering](<https://devfeed.tech/topics/performance-engineering.md>), [Latency](<https://devfeed.tech/topics/latency.md>), [vllm](<https://devfeed.tech/topics/vllm.md>), [Open Source](<https://devfeed.tech/topics/open-source.md>)

Tags: [ai](<https://devfeed.tech/tags/ai.md>), [ai-inference](<https://devfeed.tech/tags/ai-inference.md>), [ai-infrastructure](<https://devfeed.tech/tags/ai-infrastructure.md>), [ai-software](<https://devfeed.tech/tags/ai-software.md>), [benchmarks](<https://devfeed.tech/tags/benchmarks.md>), [c-plus-plus](<https://devfeed.tech/tags/c-plus-plus.md>), [cache](<https://devfeed.tech/tags/cache.md>), [caching](<https://devfeed.tech/tags/caching.md>), [cloud](<https://devfeed.tech/tags/cloud.md>), [cloud-tpu](<https://devfeed.tech/tags/cloud-tpu.md>), [cuda](<https://devfeed.tech/tags/cuda.md>), [disaggregated-serving](<https://devfeed.tech/tags/disaggregated-serving.md>), [engines](<https://devfeed.tech/tags/engines.md>), [for-developers](<https://devfeed.tech/tags/for-developers.md>), [frameworks](<https://devfeed.tech/tags/frameworks.md>), [gb200](<https://devfeed.tech/tags/gb200.md>), [google](<https://devfeed.tech/tags/google.md>), [hardware](<https://devfeed.tech/tags/hardware.md>), [inference](<https://devfeed.tech/tags/inference.md>), [latency](<https://devfeed.tech/tags/latency.md>), [open-source](<https://devfeed.tech/tags/open-source.md>), [software](<https://devfeed.tech/tags/software.md>)

### AI overview

Google's open-source TPU Sync moves KV-cache data across accelerator memory, host DRAM, storage, and network using zero-copy transfers and PjRtBuffer descriptors. Google reports up to 223 GB/s combined host-device throughput per Ironwood TPU host. In its benchmarks, cache transfer for Qwen3 235B fell from several seconds to about 230 ms, cutting average time to first token (TTFT) by 60% at low request load; a 480B model and 64,000-token prefix-caching tests showed TTFT reductions up to 90%. The prefix tests also reported over 50% higher throughput and 30% lower P99 inter-token latency. Results depend on workload, cache reuse, and network bandwidth. SemiAnalysis data places Nvidia GB200 and GB300 NVL72 ahead in parts of the disaggregated-serving curve, though comparison with a fully tuned TPU v7 setup is unavailable. TPU Sync ships with TorchTPU and supports vLLM and SGLang, but is not production-ready for everyone: APIs are changing, and support is limited to Cloud TPU v7x and v6e with specific JAX and PyTorch versions.

### Source excerpt

Google's eighth-generation TPU announcement on Monday included more than new chips, introducing TPU Sync, an open-source system designed to speed The post Google claims its new AI software cuts inference latency by over 90% -- but there's a catch appeared first on The New Stack.

## Latency-Aware Adaptive Best-of-$N$

DevFeed: [Latency-Aware Adaptive Best-of-$N$](<https://devfeed.tech/articles/latency-aware-adaptive-best-of-n-120615.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.07453>)

Author: Alon Jacobson, Qian Xie, Peter I. Frazier

Published: 2026-11-06T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [Latency](<https://devfeed.tech/topics/latency.md>), [LLM evaluation / benchmarking](<https://devfeed.tech/topics/llm-evaluation-benchmarking.md>), [Simulation](<https://devfeed.tech/topics/simulation.md>)

Tags: [adaptive](<https://devfeed.tech/tags/adaptive.md>), [algorithm](<https://devfeed.tech/tags/algorithm.md>), [alignment](<https://devfeed.tech/tags/alignment.md>), [best-of](<https://devfeed.tech/tags/best-of.md>), [compute](<https://devfeed.tech/tags/compute.md>), [cost](<https://devfeed.tech/tags/cost.md>), [cs-cl](<https://devfeed.tech/tags/cs-cl.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [distribution](<https://devfeed.tech/tags/distribution.md>), [generation](<https://devfeed.tech/tags/generation.md>), [inference](<https://devfeed.tech/tags/inference.md>), [language](<https://devfeed.tech/tags/language.md>), [latency](<https://devfeed.tech/tags/latency.md>), [load](<https://devfeed.tech/tags/load.md>), [online](<https://devfeed.tech/tags/online.md>), [prompt](<https://devfeed.tech/tags/prompt.md>), [quality](<https://devfeed.tech/tags/quality.md>), [regression](<https://devfeed.tech/tags/regression.md>), [sampling](<https://devfeed.tech/tags/sampling.md>), [server](<https://devfeed.tech/tags/server.md>), [serving](<https://devfeed.tech/tags/serving.md>), [simulation](<https://devfeed.tech/tags/simulation.md>)

### AI overview

For Best-of-$N$ sampling, the proposed two-stage method uses the first batch's rewards and latency to decide whether and how large a second batch should be. Its threshold for continuing falls as first-batch latency rises, accounting for serving cost without directly measuring server load. The method maps rewards through a prompt-specific reward-distribution CDF and uses an offline regression across other prompts to estimate expected improvement. In simulations on a preference-based alignment benchmark, it matched Best-of-$N$ reward at lower generations-equivalent cost, with latency conditioning especially beneficial when latency varies substantially, increases with batch size, or is predictable across batches. These are simulation results; the excerpt does not establish performance in live serving.

### Source excerpt

arXiv:2611.07453v1 Announce Type: cross Abstract: Test-time compute (TTC) improves language model responses by spending additional computation at inference time, and the right amount of compute depends on the prompt. Adaptive TTC methods allocate compute based on signals of response quality such as observed rewards, often by stopping once the signal exceeds a fixed threshold, and generally measure cost by the number of generations or tokens. But in online serving, latency is also a cost: it grows with server load, so a method using a fixed threshold spends too little compute when generation is fast and too much when it is slow. For Best-of-$N$ sampling, a widely used TTC method, we propose a two-stage adaptive algorithm that draws an initial batch, then uses the observed rewards and the batch's latency to decide whether to draw a second batch and at what size by comparing expected improvement against predicted cost: it draws one when the first batch's best reward is below a threshold that decreases with the first batch's latency, without explicitly measuring server load. We map each reward to its value under the cumulative distribution function (CDF) of the prompt's reward distribution, so the same tradeoff parameters can be reused across tasks; a regression fitted offline on other prompts predicts a distribution for the CDF value of the first batch's best reward, giving the expected improvement in closed form. We evaluate the method by simulation on a preference-based alignment benchmark, measuring generations-equivalent cost: generations plus latency, converted to generations. The method reaches Best-of-$N$'s reward at lower generations-equivalent cost, and conditioning on the measured first-batch latency saves at least as much of that cost as fixed-threshold (latency-unaware) baselines, and more when latency is highly variable, when it grows with batch size, and when the second batch's latency is highly predictable from the first's.

## TokenScale: Timely and Accurate Autoscaling for Disaggregated LLM Serving with Token Velocity

DevFeed: [TokenScale: Timely and Accurate Autoscaling for Disaggregated LLM Serving with Token Velocity](<https://devfeed.tech/articles/tokenscale-timely-and-accurate-autoscaling-for-disaggregated-llm-serving-with-token-velocity-119083.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2512.03416>)

Author: Ruiqi Lai, Hongrui Liu, Chengzhi Lu, Zonghao Liu, Siyu Cao, Siyang Shao, Yixin Zhang, Luo Mai, Dmitrii Ustiugov

Published: 2026-11-05T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Distributed Computing](<https://devfeed.tech/sources/arxiv-distributed-computing.md>)

Topics: [LLM optimization](<https://devfeed.tech/topics/llm-optimization.md>), [performance-engineering](<https://devfeed.tech/topics/performance-engineering.md>), [Large Language Model](<https://devfeed.tech/topics/llm.md>), [Latency](<https://devfeed.tech/topics/latency.md>)

Tags: [autoscaling](<https://devfeed.tech/tags/autoscaling.md>), [backpressure](<https://devfeed.tech/tags/backpressure.md>), [costs](<https://devfeed.tech/tags/costs.md>), [cs-dc](<https://devfeed.tech/tags/cs-dc.md>), [disaggregation](<https://devfeed.tech/tags/disaggregation.md>), [evaluation](<https://devfeed.tech/tags/evaluation.md>), [gpu](<https://devfeed.tech/tags/gpu.md>), [latency](<https://devfeed.tech/tags/latency.md>), [llm](<https://devfeed.tech/tags/llm.md>), [llm-serving](<https://devfeed.tech/tags/llm-serving.md>), [performance](<https://devfeed.tech/tags/performance.md>), [prefill](<https://devfeed.tech/tags/prefill.md>), [proactive](<https://devfeed.tech/tags/proactive.md>), [production](<https://devfeed.tech/tags/production.md>), [resource-utilization](<https://devfeed.tech/tags/resource-utilization.md>), [system-design](<https://devfeed.tech/tags/system-design.md>), [token](<https://devfeed.tech/tags/token.md>)

### AI overview

TokenScale aims to reduce SLO violations and over-provisioning in disaggregated LLM serving by pairing a leading indicator with a burst-response mechanism. Its Token Velocity metric measures work rate across prefill, network, and decode stages to signal backpressure early and support proactive scaling. Convertible Decoders let decoder GPUs take on prefill work during traffic spikes, avoiding the initialization delay of new prefillers. In evaluation on a GPU cluster using production traces, TokenScale raised SLO attainment from 50-88% to 80-96% and reduced costs by 4-14% compared with DistServe, BlitzScale, and AIBrix. The supplied text is an abstract and does not detail evaluation conditions or the metric's implementation.

### Source excerpt

arXiv:2512.03416v2 Announce Type: replace Abstract: The architectural shift to prefill/decode (PD) disaggregation in LLM serving improves resource utilization but struggles with the bursty nature of modern workloads. Existing autoscaling policies, often retrofitted from monolithic systems like those in AIBrix and DistServe, rely on lagging indicators such as GPU utilization or coarse-grained request counts. This results in slow reactions to load spikes, leading to significant Time-to First-Token (TTFT) and Time-Per-Output-Token (TPOT) SLO violations and costly over-provisioning. We introduce TokenScale, an autoscaling framework that resolves this performance mismatch through two innovations. First, we propose Token Velocity, a novel metric that unifies the prefill, network, and decode stages by quantifying their rate of work. As a leading indicator of system backpressure, it enables proactive scaling. Second, Convertible Decoders allow decoder GPUs to dynamically execute prefill tasks during traffic spikes, creating a rapid-response buffer that absorbs bursts and eliminates the initialization latency of new prefillers. Our evaluation on a GPU cluster with production traces shows TokenScale improves SLO attainment from 50-88% to 80-96% and reduces costs by 4-14% over state-of-the-art systems, including DistServe, BlitzScale, and AIBrix. By uniting a predictive metric with a flexible system design, TokenScale significantly boosts the performance and efficiency of disaggregated LLM serving infrastructure.

## Anomaly Detection on Resource-Constrained Devices with Reservoir Computing

DevFeed: [Anomaly Detection on Resource-Constrained Devices with Reservoir Computing](<https://devfeed.tech/articles/anomaly-detection-on-resource-constrained-devices-with-reservoir-computing-116989.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.04835>)

Author: Marco Angioli, Mathias Persson, Hiroto Tamura, Hadi Banaee, Amy Loutfi, Gouhei Tanaka, Alberto Giaretta, Denis Kleyko

Published: 2026-11-04T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Machine Learning](<https://devfeed.tech/sources/arxiv-machine-learning.md>)

Topics: [Machine Perception](<https://devfeed.tech/topics/machine-perception.md>), [Latency](<https://devfeed.tech/topics/latency.md>)

Tags: [anomaly-detection](<https://devfeed.tech/tags/anomaly-detection.md>), [autoencoder](<https://devfeed.tech/tags/autoencoder.md>), [connectivity](<https://devfeed.tech/tags/connectivity.md>), [cs-lg](<https://devfeed.tech/tags/cs-lg.md>), [devices](<https://devfeed.tech/tags/devices.md>), [edge](<https://devfeed.tech/tags/edge.md>), [energy](<https://devfeed.tech/tags/energy.md>), [inference](<https://devfeed.tech/tags/inference.md>), [latency](<https://devfeed.tech/tags/latency.md>), [memory](<https://devfeed.tech/tags/memory.md>), [microcontroller](<https://devfeed.tech/tags/microcontroller.md>), [mlperf](<https://devfeed.tech/tags/mlperf.md>), [optimizations](<https://devfeed.tech/tags/optimizations.md>), [reduce](<https://devfeed.tech/tags/reduce.md>), [reservoir-computing](<https://devfeed.tech/tags/reservoir-computing.md>), [roc-auc](<https://devfeed.tech/tags/roc-auc.md>), [series](<https://devfeed.tech/tags/series.md>), [shift](<https://devfeed.tech/tags/shift.md>), [state](<https://devfeed.tech/tags/state.md>), [temporal](<https://devfeed.tech/tags/temporal.md>), [time-series](<https://devfeed.tech/tags/time-series.md>)

### AI overview

The study finds that Mahalanobis Distance of Reservoir States (MDRS) can meet the MLPerf Tiny anomaly detection performance target with fewer operations and a smaller model than the reference autoencoder. Its optimizations replace random reservoir connections with a circular shift, compute a time series' anomaly score from the final reservoir state, and use reservoir memory to reduce preprocessing cost. On a NUCLEO-F401RE microcontroller, optimized MDRS preprocessing yielded ROC AUC 0.869, a 24 KB model, estimated energy use of 24 µJ per inference, and 0.32 ms latency--75.9 times faster than the autoencoder with standard preprocessing. Only MDRS and the autoencoder met the benchmark target among the evaluated methods.

### Source excerpt

arXiv:2611.04835v1 Announce Type: new Abstract: Anomaly detection at the edge requires the deployment of models that achieve competitive performance, while minimizing computational cost and memory footprint. In this article, we investigate the suitability of the Mahalanobis Distance of Reservoir States (MDRS), an unsupervised time-series anomaly detection method, for resource-constrained devices and propose two optimizations to reduce its computational requirements. Specifically, we replace random reservoir connectivity (MDRS-Rand) with a circular shift (MDRS-Circ) and show that the anomaly score for an entire time series can be computed from the final reservoir state alone. Furthermore, we show that exploiting the reservoir's temporal memory enables more efficient input preprocessing, further reducing computational cost. We evaluate the MDRS variants on the MLPerf Tiny anomaly detection benchmark against the reference autoencoder and other anomaly detection methods, and deploy their inference kernels on a NUCLEO-F401RE microcontroller to assess latency, estimate energy consumption, and measure memory footprint. Among the evaluated methods, only MDRS and the autoencoder meet the benchmark's performance target, with MDRS requiring fewer operations and smaller models. With optimized preprocessing, MDRS achieves an ROC AUC of 0.869, a model size of 24 KB, an estimated energy consumption of 24 uJ per inference, and a latency of 0.32 ms, corresponding to a 75.9x speedup over the autoencoder with standard preprocessing.

## Beyond Leaderless: Randomized Consensus Against Mobile Crashes

DevFeed: [Beyond Leaderless: Randomized Consensus Against Mobile Crashes](<https://devfeed.tech/articles/beyond-leaderless-randomized-consensus-against-mobile-crashes-113793.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2611.01323>)

Author: Ittai Abraham, Kaya Alpturer, Pranav Garimidi, Joachim Neu, Tim Roughgarden, Ertem Nusret Tas

Published: 2026-11-03T05:00:00Z

Content type: article

Language: en

Sources: [arXiv: Distributed Computing](<https://devfeed.tech/sources/arxiv-distributed-computing.md>)

Topics: [consensus algorithm](<https://devfeed.tech/topics/consensus-algorithm.md>), [Distributed Systems & Parallel Computing](<https://devfeed.tech/topics/distributed-systems-parallel-computing.md>), [Latency](<https://devfeed.tech/topics/latency.md>)

Tags: [adaptive](<https://devfeed.tech/tags/adaptive.md>), [adversary](<https://devfeed.tech/tags/adversary.md>), [beacon](<https://devfeed.tech/tags/beacon.md>), [broadcast](<https://devfeed.tech/tags/broadcast.md>), [consensus](<https://devfeed.tech/tags/consensus.md>), [crashes](<https://devfeed.tech/tags/crashes.md>), [cs-dc](<https://devfeed.tech/tags/cs-dc.md>), [latency](<https://devfeed.tech/tags/latency.md>), [leader](<https://devfeed.tech/tags/leader.md>), [mobile](<https://devfeed.tech/tags/mobile.md>), [multiple](<https://devfeed.tech/tags/multiple.md>), [network](<https://devfeed.tech/tags/network.md>), [replica](<https://devfeed.tech/tags/replica.md>), [vote](<https://devfeed.tech/tags/vote.md>), [voting](<https://devfeed.tech/tags/voting.md>)

### AI overview

ML-FVS adapts fixed-view Simplex to withstand mobile crashes by allowing multiple leaders to broadcast proposals. At the voting deadline, replicas vote on proposals seen so far; a hidden threshold beacon revealed afterward orders leaders by priority and resolves votes to the highest-priority proposal. The model combines permanent adaptive Byzantine faults with mobile crashes. With a constant crash budget and Byzantine corruption delayed by Δ, the authors report O(Δ) expected latency across repeated views whose leaders start honest; with immediate Byzantine corruption, they report liveness. Their evaluation under random network delays and stochastic faults finds fewer view failures with multiple leaders, even when rational leaders delay proposals to increase rewards.

### Source excerpt

arXiv:2611.01323v1 Announce Type: new Abstract: Many consensus protocols fail to terminate under attack from an adversary that has the power to move between replicas and repeatedly crash one of them, even in synchrony. We present Multi-Leader Fixed-View Simplex (ML-FVS) to defend against such an adversary while keeping the simple structure and good-case latency of Simplex consensus. Our protocol extends fixed-view Simplex with multiple leaders, each of which may broadcast a proposal. At the voting deadline, each replica votes on the set of proposals it has seen so far. A hidden threshold beacon revealed after the votes gives a priority order on the multiple leaders, resolving each vote to its highest-priority proposal. Our model combines permanent adaptive Byzantine faults with mobile crashes. For a constant crash budget and $\Delta$-delayed Byzantine corruption, we obtain $O(\Delta)$ expected latency over repeated views whose leaders start honest. With immediate Byzantine corruptions, the protocol still achieves liveness. We also evaluate the protocol under random network delays and stochastic faults, finding that multiple leaders reduce view failures even when rational leaders delay their proposals to increase rewards.

## How we halved DNS build delay and cut memory use by 85%

DevFeed: [How we halved DNS build delay and cut memory use by 85%](<https://devfeed.tech/articles/how-we-halved-dns-build-delay-and-cut-memory-use-by-85-113832.md>)

Original publisher: [Read original article](<https://blog.cloudflare.com/dns-targeted-zone-builds/>)

Author: Alex Fattouche

Published: 2026-11-02T14:00:00Z

Content type: article

Language: en

Sources: [Cloudflare Blog](<https://devfeed.tech/sources/cloudflare-blog.md>)

Topics: [Domain Name System (DNS)](<https://devfeed.tech/topics/dns-resolution-process.md>), [scheduling](<https://devfeed.tech/topics/scheduling.md>), [Latency](<https://devfeed.tech/topics/latency.md>), [authoritative name server](<https://devfeed.tech/topics/authoritative-name-server.md>), [rate-limiting](<https://devfeed.tech/topics/rate-limiting.md>)

Tags: [2](<https://devfeed.tech/tags/2.md>), [2022](<https://devfeed.tech/tags/2022.md>), [3](<https://devfeed.tech/tags/3.md>), [about](<https://devfeed.tech/tags/about.md>), [account](<https://devfeed.tech/tags/account.md>), [added](<https://devfeed.tech/tags/added.md>), [aggregate](<https://devfeed.tech/tags/aggregate.md>), [build](<https://devfeed.tech/tags/build.md>), [builds](<https://devfeed.tech/tags/builds.md>), [changes](<https://devfeed.tech/tags/changes.md>), [deep-dive](<https://devfeed.tech/tags/deep-dive.md>), [delay](<https://devfeed.tech/tags/delay.md>), [dns](<https://devfeed.tech/tags/dns.md>), [full](<https://devfeed.tech/tags/full.md>), [here](<https://devfeed.tech/tags/here.md>), [kafka](<https://devfeed.tech/tags/kafka.md>), [memory](<https://devfeed.tech/tags/memory.md>), [planner](<https://devfeed.tech/tags/planner.md>), [quicksilver](<https://devfeed.tech/tags/quicksilver.md>), [rebuild](<https://devfeed.tech/tags/rebuild.md>), [records](<https://devfeed.tech/tags/records.md>), [speed-reliability](<https://devfeed.tech/tags/speed-reliability.md>), [used](<https://devfeed.tech/tags/used.md>), [with](<https://devfeed.tech/tags/with.md>)

### AI overview

A targeted build planner and unified scheduler reduced unnecessary full-zone DNS rebuilds while preserving correctness. Events identify changed settings and records; the planner maps them to the smallest safe work--specific names, names plus metadata, apex plus metadata, purge, or full-zone build. Unknown or broad changes still trigger full builds, and large zones avoid escalation to a full build because it cannot finish. Before writing, the builder checks fresh database state, reconstructs shared metadata for zones with the same name, and verifies replica lag. The scheduler merges queued work per zone while retaining Kafka acknowledgments through retries, so uncertain or unfinished work is replayed rather than lost. Equivalence tests and a production-stream mirror compare output before rollout. In a 24-hour measurement, 92.95% of builds avoided full-zone work and full-build share fell about 40%. Peak working-set memory dropped 85% (79 to 12 GiB); the largest p99 delay spike fell from 14.6 to 1.21 minutes. Shorter rebuild intervals and aggregator waits helped halve median completion delay from about 7.7 to 3.3 seconds. This measures writing to Quicksilver, not its subsequent network

### Source excerpt

Zone setting changes used to rebuild entire DNS zones, even ones with tens of millions of records. Here is how we built a planner that picks the smallest safe build, cutting full builds by 40%, peak memory by 85%, and halving build delay.

## One Word Fixed a 60-Second nginx Delay

DevFeed: [One Word Fixed a 60-Second nginx Delay](<https://devfeed.tech/articles/one-word-fixed-a-60-second-nginx-delay-112287.md>)

Original publisher: [Read original article](<https://blog.devops.dev/one-word-fixed-a-60-second-nginx-delay-a25e16b7f210?source=rss----33f8b2d9a328---4>)

Author: DRAVID KUMAR

Published: 2026-11-01T20:01:02Z

Content type: article

Language: en

Sources: [DevOps.dev](<https://devfeed.tech/sources/devops-dev.md>)

Topics: [nginx](<https://devfeed.tech/topics/nginx.md>), [Latency](<https://devfeed.tech/topics/latency.md>), [cURL](<https://devfeed.tech/topics/curl.md>), [Domain Name System (DNS)](<https://devfeed.tech/topics/dns-resolution-process.md>)

Tags: [2](<https://devfeed.tech/tags/2.md>), [3](<https://devfeed.tech/tags/3.md>), [about](<https://devfeed.tech/tags/about.md>), [all](<https://devfeed.tech/tags/all.md>), [assets](<https://devfeed.tech/tags/assets.md>), [auth](<https://devfeed.tech/tags/auth.md>), [bash](<https://devfeed.tech/tags/bash.md>), [browser](<https://devfeed.tech/tags/browser.md>), [but](<https://devfeed.tech/tags/but.md>), [command](<https://devfeed.tech/tags/command.md>), [connect](<https://devfeed.tech/tags/connect.md>), [crashes](<https://devfeed.tech/tags/crashes.md>), [curl](<https://devfeed.tech/tags/curl.md>), [days](<https://devfeed.tech/tags/days.md>), [debugging](<https://devfeed.tech/tags/debugging.md>), [delay](<https://devfeed.tech/tags/delay.md>), [devops](<https://devfeed.tech/tags/devops.md>), [fixed](<https://devfeed.tech/tags/fixed.md>), [linux](<https://devfeed.tech/tags/linux.md>), [nginx](<https://devfeed.tech/tags/nginx.md>), [one](<https://devfeed.tech/tags/one.md>), [productivity](<https://devfeed.tech/tags/productivity.md>), [word](<https://devfeed.tech/tags/word.md>)

### AI overview

An nginx reverse proxy caused intermittent 60-second delays on Google sign-in because `localhost` resolved to both IPv6 and IPv4. On slow requests nginx tried `[::1]:5000`, where no service was listening, waited for its 60-second connection timeout, then retried the working IPv4 address. Direct requests to the app were fast, and the nginx error log recorded upstream connection timeouts. Replacing `proxy_pass http://localhost:5000;` with `proxy_pass http://127.0.0.1:5000;` resolved the stalls in the reported checks: five slow requests in a 15-request sample before, none in nine requests after, plus a successful browser sign-in. The author could not confirm what triggered the issue two days earlier. As prevention, the account recommends using explicit IPv4 addresses for local nginx upstreams, monitoring timeout logs, and considering a shorter `proxy_connect_timeout`; these steps are reported recommendations, not further tested outcomes.

### Source excerpt

Part of my ongoing DevOps series. Previously: Flask on VPS | Grafana Reverse Proxy | Linux LVM NOTE: This is a real production incident. Domain names and service names are changed. Every command below was actually run -- nothing is theoretical. The Symptom: Users Were Waiting a Full Minute to Sign In Every third Google sign-in request to auth.example.com stalled for about 60 seconds. Users clicked "Sign in with Google" and the page just sat there. Some waited. Some retried. Some probably gave up and never came back. Google sign-in had worked normally until about two days earlier. Here's what made it maddening: Nothing visibly failed. No error pages, no crashes. Every sign-in eventually succeeded -- only a minute late. Google wasn't slow. Google's own pages loaded in about 1 second. Most callbacks were fast. All static assets on the main site were under 0.3 seconds. It was intermittent. Exactly every third request. A pattern, not randomness. And the one detail that kept me up at night: email/password login worked perfectly. It went straight to the auth service inside Docker and never touched this nginx route. So the problem was isolated. But hidden. The Investigation: Seven Steps to the Truth Each step ruled something out. Here's exactly how we narrowed it down. Step 1: Browser HAR -- The First Clue Captured during a slow sign-in: MeasurementResultGET /auth/google60,165 msDNS + connect + TLS~40 msGoogle's pages~1.0 sCallback~0.5 s Ruled out: The network and Google. The delay is on our auth server. Step 2: Single Curl -- The Problem Is Intermittentcurl -w "Time: %{time_total}s\n" -o /dev/null -s "https://auth.example.com/auth/google" Result: 302 in 0.34 s. Showed: The problem is intermittent, not constant. One request proves nothing. Step 3: Curl Loop -- The Pattern Emerges bash for i in {1..15}; do curl -w "Request $i: %{time_total}s\n" -o /dev/null -s "https://auth.example.com/auth/google" sleep 20 done Result: RequestTime160.2 s20.8 s31.2 s460.7 s50.4 s60.9 s760.1 s80.3

## Inter-Satellite Link Optimization for Low-Latency Global Networking

DevFeed: [Inter-Satellite Link Optimization for Low-Latency Global Networking](<https://devfeed.tech/articles/inter-satellite-link-optimization-for-low-latency-global-networking-107602.md>)

Original publisher: [Read original article](<https://arxiv.org/abs/2604.15528>)

Author: Arman Mollakhani, Jerayu Tiamraj, Shu-Jie Cao, Dongning Guo

Published: 2026-10-30T04:00:00Z

Content type: article

Language: en

Sources: [arXiv: Networking](<https://devfeed.tech/sources/arxiv-networking.md>)

Topics: [communications](<https://devfeed.tech/topics/communications.md>), [Latency](<https://devfeed.tech/topics/latency.md>), [graph theory](<https://devfeed.tech/topics/graph-theory.md>), [Combinatorial optimization](<https://devfeed.tech/topics/combinatorial-optimization.md>)

Tags: [connectivity](<https://devfeed.tech/tags/connectivity.md>), [convex](<https://devfeed.tech/tags/convex.md>), [cs-ni](<https://devfeed.tech/tags/cs-ni.md>), [delay](<https://devfeed.tech/tags/delay.md>), [diameter](<https://devfeed.tech/tags/diameter.md>), [global](<https://devfeed.tech/tags/global.md>), [latency](<https://devfeed.tech/tags/latency.md>), [link](<https://devfeed.tech/tags/link.md>), [low-latency](<https://devfeed.tech/tags/low-latency.md>), [network](<https://devfeed.tech/tags/network.md>), [networking](<https://devfeed.tech/tags/networking.md>), [optimization](<https://devfeed.tech/tags/optimization.md>), [orbit](<https://devfeed.tech/tags/orbit.md>), [satellite](<https://devfeed.tech/tags/satellite.md>), [selection](<https://devfeed.tech/tags/selection.md>), [shortest-path](<https://devfeed.tech/tags/shortest-path.md>), [simulations](<https://devfeed.tech/tags/simulations.md>), [topology](<https://devfeed.tech/tags/topology.md>)

### AI overview

The proposed two-stage method reduces inter-satellite network diameter by optimizing a convex relaxation that maximizes graph algebraic connectivity, then converting its fractional links into a feasible topology with maximum-weight b-matching. In Walker-Delta simulations, it achieves smaller diameters than local search and standard patterned topologies. For 1,584 satellites with four inter-satellite links each, up to 2,500 km long, the reported diameter is as low as 12 hops and maximum satellite-to-satellite propagation delay is about 90 ms. Including up- and downlinks totaling 8 ms yields up to about 98 ms user-to-user propagation delay, comparable to a great-circle fiber route between antipodal points. These are simulation results.

### Source excerpt

arXiv:2604.15528v2 Announce Type: replace Abstract: Large-scale low-Earth-orbit satellite constellations offer a promising platform for global low-latency networking, in part because signals propagate faster in free space than in fiber and copper. In such systems, end-to-end latency is largely determined by the inter-satellite link (ISL) topology. In particular, the network diameter--the longest shortest path, in hops, between any pair of satellites--serves as a key performance metric for time-sensitive applications. Designing diameter-optimal topologies is challenging due to node degree constraints, orbital dynamics, and the combinatorial nature of link selection. This paper proposes a two-stage optimization framework for ISL topology design. First, a continuous relaxation of the link-selection problem is formulated as a convex program that maximizes the algebraic connectivity of the graph Laplacian, serving as a tractable surrogate for diameter minimization. Second, the resulting fractional solution is mapped to a feasible topology via maximum-weight b-matching, a problem solvable in polynomial time. Simulations on Walker-Delta constellations show that the proposed method consistently achieves smaller network diameters than a local-search baseline and standard patterned topologies. For a 1,584-satellite constellation in which each satellite has four ISLs of up to 2,500 km, the diameter is as small as 12 hops, and the maximum propagation delay between any two satellites is about 90 ms. With up- and downlinks adding up to 8 ms, the user-to-user propagation delay is at most about 98 ms, comparable to the delay over a great-circle fiber route between antipodal points on Earth.

[Next page](<https://devfeed.tech/topics/latency.md?cursor=WyIyMDI2LTEwLTMwVDA0OjAwOjAwKzAwOjAwIiwgImY1ZDYxNmQyLTU0YWQtNGJkMS1iOTBjLTQ5OWFlMDFiZGQ0NyJd>)