How Kubernetes CPU Throttling Causes Latency Spikes Even at Low CPU Usage

·1 views

Kubernetes enforces CPU limits through the Linux kernel's Completely Fair Scheduler (CFS) in 100-millisecond windows, which can freeze containers even when average CPU usage appears low. A container hitting its quota early in a window may remain frozen for the rest of it, causing latency spikes invisible to standard monitoring dashboards. SREs are advised to track the CFS throttling ratio metric rather than raw CPU usage, flagging applications where over 15–25% of scheduler windows are throttled. Runtime misconfigurations in Java and Go can worsen throttling by spawning threads based on host core counts rather than container limits, a problem addressable via UseContainerSupport and Uber's automaxprocs library. For latency-sensitive workloads on bare-metal nodes, assigning pods to the Guaranteed QoS class with integer CPU values allows Kubernetes to bypass CFS quotas entirely by pinning containers to dedicated physical cores.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

AgentGPT 2026 Pricing: $40/Month Pro Plan Comes With a 25-Loop Ceiling

AgentGPT, the open-source autonomous AI agent platform maintained by Reworkd, offers three pricing tiers in 2026: a free trial at $0, a Pro plan at $40 per month, and a custom-priced Enterprise tier. The Pro plan includes 30 agents per day, access to GPT-4, unlimited web search, and a cap of 25 loops per agent. Each loop represents one think-act-observe cycle, meaning agents have a fixed runway of 25 steps to complete any given task before stopping. At $40 per month, AgentGPT sits more than double the $19 median price among productivity software plans, according to ToolFi's 2026 dataset of 469 priced tools. The platform has gained significant community traction, with its GitHub repository accumulating around 36,300 stars and 9,270 forks as of September 2026.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Developer shares five hard-learned lessons building a Cardano-Midnight USDM bridge app

A developer building cross-chain USDM transfer tools between Cardano and Midnight's testnets documented five undocumented pitfalls encountered during the process. One key issue is that USDM sent from Midnight to Cardano lands at an enterprise address not monitored by the Lace wallet, making funds appear lost when they are not. VIA's block explorer also requires different transaction hash formats depending on the source chain, and defaults to Mainnet view, causing completed testnet transfers to appear nonexistent. The developer built two working applications — Tender, a browser-based transfer app, and Privoice, a private invoice contract — completing eight transfers and one deployed contract. The findings are shared as practical guidance for others building user-facing applications on the VIA cross-chain infrastructure.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Researcher Builds Wazuh SIEM Lab to Test Real-World Attack Detection Gaps

A security researcher configured a Wazuh SIEM environment using a Dockerized Damn Vulnerable Web Application (DVWA) to evaluate how well the platform detects real attacks. The setup involved custom container log ingestion, where an initial permission misconfiguration silently blocked the Wazuh agent from reading Apache access logs until directory permissions were corrected. Three live attack scenarios were then executed, including Nmap and Nikto reconnaissance scans, which triggered over 380 events and activated Wazuh rules for scanner activity and web errors. An OS command injection payload submitted through DVWA successfully ran system commands in the browser, revealing a key gap where web-layer logs captured the attack but the host operating system remained blind to the actual commands executed. The experiment highlights that effective SIEM deployment requires deliberate visibility engineering, not just tool activation, to correlate application-level and OS-level telemetry.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How a Love Fortune App Can Enable Sharing Without Storing Personal Data

A Japanese-style love fortune application called 恋みくじ poses a unique privacy challenge: enabling shareable results without exposing users' personal emotions or relationship details. The proposed architecture separates data into three categories — public content, private context, and operational data — ensuring only public fortune identifiers appear in shared URLs. A safe shareable link contains only a result ID tied to the app's public catalog, never the visitor's name, question, or relationship history. Common pitfalls like embedding sensitive details in query strings are highlighted, as such data can leak through browser history, server logs, and link preview services. The article also warns that long, random-looking URLs do not guarantee privacy if they point to database records storing personal user data.

0 comments Read more at DEV Community