비머 코펜하겐

\documentclass[aspectratio=169]{beamer}
\usetheme{Copenhagen}
\usecolortheme[RGB={31,59,112}]{structure}

\title[Taming Retry Storms]{Taming Retry Storms}
\subtitle{Reliability lessons from a global payment platform}
\author[J. Berg]{Jonas Berg}
\institute[Northpay]{Site Reliability Engineering, Northpay}
\date{ScaleConf Europe 2026}

\begin{document}

\begin{frame}
  \titlepage
\end{frame}

\begin{frame}{Agenda}
  \tableofcontents
\end{frame}

\section{Anatomy of an incident}

\begin{frame}{Anatomy of a retry storm}
  \begin{itemize}
    \item A 40 ms latency bump in one region tripped client timeouts set to 50 ms.
    \item Every timed-out call retried three times with no jitter, tripling load.
    \item The extra load pushed p99 latency higher, timing out even more calls.
    \item Within four minutes, healthy capacity was serving 11x normal traffic.
  \end{itemize}
\end{frame}

\section{Defenses that worked}

\begin{frame}{Defenses that worked}
  \begin{block}{Budgeted retries}
    Cap retries at 10\% of successful requests per client, measured over a
    sliding window. Storms starve themselves instead of the service.
  \end{block}
  \begin{block}{Deadline propagation}
    Pass the remaining deadline downstream so a doomed request is dropped
    early rather than retried at every hop.
  \end{block}
\end{frame}

\begin{frame}{Closing}
  \begin{center}
    {\Large Retries are load amplifiers. Budget them.}\\[1em]
    Slides and the incident write-up: northpay.example/talks
  \end{center}
\end{frame}

\end{document}