1
0
Fork 0
forked from Deuxfleurs/garage
garage/doc/talks/2021-04-28_spirals-team/escaping_the_cloud.tex

124 lines
No EOL
2.9 KiB
TeX

\section{Escaping the cloud}
% %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% \begin{frame}{Down to Earth with home-hosting}
% \todo{Stanley Parabole reference?}
% \end{frame}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\begin{frame}{Why?}
\begin{itemize}
\item \textbf{Privacy}: no prying eyes besides your ISP
\item \textbf{Control} of your infrastructure
\item \textbf{Ecology}: reuse old hardware
\end{itemize}
\vfill
\begin{block}{\emph{Tim Berners-Lee} (1994)}
``Now, if someone tries to monopolize the Web, for example pushes proprietary variations on network protocols, then that would make me unhappy.''
\end{block}
\begin{itemize}
\item Make Tim Berners-Lee happy
\end{itemize}
\end{frame}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\begin{frame}{What?}
\centering\Large
A data store for commodity hardware on heterogenous household connections.
\vfill\raggedright\normalsize
\begin{block}{Targetting user-facing services}
\begin{itemize}
\item Static sites
\item E-mails
\item Instant communication
%\item Video streaming % No need for a data store
\item Collaboration
\end{itemize}
\end{block}
\vfill
Nothing fancy like sensors data streams, AI or IoT.
\end{frame}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\begin{frame}{What?}
\begin{block}{Requirements}
\begin{itemize}
\item \textbf{No single point of failure} / flat hierarchy:
Any node can die for extended periods of time.
\item \textbf{Multi-site}: cluster spans regions/countries.
\item \textbf{Acceptable performance}.
\item \textbf{Lightweight}: targets legacy hardware.
\item \textbf{Conceptually simple}: built for low-tech organisations.
Adding/maintaining cluster nodes should be easy.
\end{itemize}
\end{block}
\vfill
\begin{block}{Non-goals}
\begin{itemize}
\item \textbf{Super badass performance}.
\item \textbf{NAT traversal} etc.: we require full-mesh connectivity.
\end{itemize}
\end{block}
\end{frame}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\begin{frame}{How?}
\begin{itemize}
\item Theoretically possible with object storage \& CRDTs.
\vfill
\item Household uplinks are getting decent (optical fibers).
\end{itemize}
\end{frame}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\begin{frame}{Research Questions}
\begin{itemize}
\item Decent performance despite bad inter-node connectivity.
\vfill
\item Tailoring workloads as a function of nodes' capabilities:
\begin{itemize}
\item Make use of low-end nodes (e.g. Raspberry Pis),
\item Avoid impeding global performance because of low-end nodes.
\end{itemize}
\vfill
\item Building CRDTs for target use-cases:
\begin{itemize}
\item Software engineering: DSL or native code?
\item Provide APIs to data store users? Risky?
\end{itemize}
\vfill
\item Cluster management: effortless UX, low perf. overhead.
\end{itemize}
\end{frame}