garage-unix-socket/doc/talks/2021-04-28_spirals-team/sota.tex

323 lines
No EOL
8.3 KiB
TeX

\section{State of the art}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\begin{frame}{The CAP theorem}{Consistency vs. Availability}
\begin{block}{Eric Brewer's theorem}
``A shared-state system can have \textbf{at most two} of the following properties at any given time:
\begin{itemize}
\item \textbf{C}onsistency
\item \textbf{A}vailability
\item \textbf{P}artition tolerance''
\end{itemize}
\end{block}
\begin{center}
\Large
Under network partitions, a distributed data store has to sacrifice either availability or consistency.
\end{center}
\vfill
\begin{itemize}
\item \textbf{Consistency-first}: Abort incoming queries;
\item \textbf{Availability-first}: Return possibly stale data.
\end{itemize}
\end{frame}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\begin{frame}{Consistency-first: the ACID model}{Consistency vs. Availability}
\textbf{Transaction}: unit of work within an ACID data store.
%Comprises multiple operations.
%E.g. bank transfer.
%E.g. a bank transfer from A to B is a transaction involving two operations: withdraw money from A & credit B with the same money amount.
\vfill
\begin{itemize}
\item \textbf{\underline{A}tomicity}: Transactions either complete entirely or fail.
No transaction ever seen as in-progress.
\item \textbf{\underline{C}onsistency}: Transactions always generate a valid state.
The database maintains its invariants across transactions.
\item \textbf{\underline{I}solation}: Concurrent transactions are seen as sequential.
Transactions are serializable, or sequentially consistent.
\item \textbf{\underline{D}urability}: Committed transactions are never forgotten.
\end{itemize}
\vfill\centering
Reads are fast, writes are slow.
\vfill\raggedright
Example: relational databases.
\end{frame}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\begin{frame}[fragile]{Concurrent writes in ACID}{Consistency vs. Availability}
\begin{columns}
\column{.5\columnwidth}
\begin{block}{}
\begin{lstlisting}
transaction AcqDoses(y):
x <- SELECT #vaccines;
UPDATE #vaccines = (x + y);
\end{lstlisting}
\end{block}
\vspace{5ex}
Supports compound operations.
\column{.5\columnwidth}
\centering
\includegraphics[width=\columnwidth]{figures/conflict_acid.pdf}
\end{columns}
\end{frame}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\begin{frame}{Availability-first: the BASE model}{Consistency vs. Availability}
Some apps prefer availability, e.g. Amazon products' reviews.
\vfill
The BASE model trades Consistency \& Isolation for Availability.
%Some applications do not care about strong consistency and prefer being highly available (e.g. Amazon's product reviews).
%In order to achieve higher availability, the BASE model relaxes consistency constraints of the ACID model: "eventual consistency".
\vfill
\begin{itemize}
\item \textbf{\underline{B}asic \underline{A}vailability}:
The data store thrives to be available.
\item \textbf{\underline{S}oft-state}:
Replicas can disagree on the valid state.
\item \textbf{\underline{E}ventual consistency}:
In the absence of write queries,
the data store will eventually converge to a single valid state.
\end{itemize}
\vfill\centering
Writes are fast, reads are slow.
\vfill\raggedright
Examples: key-value \& object stores.
\end{frame}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\begin{frame}{Concurrent writes in BASE}{Consistency vs. Availability}
\begin{columns}
\column{.5\columnwidth}
\begin{block}{Object}
\begin{itemize}
\item Unique key
\item Arbitrary value
\item Metadata
\end{itemize}
\end{block}
\vspace{5ex}
Conflict resolution = client's job!
\vspace{5ex}
No compound operations.
\column{.5\columnwidth}
\centering
\includegraphics[width=\columnwidth]{figures/conflict_base.pdf}
\end{columns}
% KV storage is another example, distinction is minor here
% Object = unique key, arbitrary value, metadata.
% Object storage only provides semantics to investigate causal order of queries *for individual objects*. No compound operations, no transactions.
% Much easier to distribute, and "scale-out".
% Write is fast, read is slow (gotta collect all object versions).
% \todo{vaccines example with BASE model}
\end{frame}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\begin{frame}{Strong Eventual Consistency w/ CRDTs}{Consistency vs. Availability}
\centering\small
\fullcite{defago_conflict-free_2011}
\vfill\raggedright\normalsize
\begin{block}{Strong Eventual Consistency (SEC)}
\begin{itemize}
\item CRDTs specify distributed operations
\item Conflicts will be solved according to specification
\item Proven \& bound eventual convergence
\end{itemize}
\end{block}
\vfill\centering
\includegraphics[width=.5\columnwidth]{figures/crdt.pdf}
\end{frame}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\begin{frame}[fragile]{Concurrent writes with CRDTs}{Consistency vs. Availability}
\begin{columns}
\column{.5\columnwidth}
\begin{block}{}
\begin{lstlisting}
CRDT Counter(x0):
history = {}
op. incr(y):
history U= {(UUID(), y)}
op. decr(y):
history U= {(UUID(), -y)}
op. read():
x = x0
for (_, y) in history:
x += y
return x
\end{lstlisting}
\end{block}
\vspace{2ex}
Operations commute?
$\implies$ screw total order!
\column{.5\columnwidth}
\centering
\includegraphics[width=\columnwidth]{figures/conflict_crdt.pdf}
\end{columns}
\end{frame}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\begin{frame}{A complex CRDT: the DAG}{Consistency vs. Availability}
\centering
\only<1>{\includegraphics[height=\textheight]{figures/dag_crdt.png}}%
\only<2>{
Just to say I swept a lot under the rug.
\vfill
For details, go read:
\fullcite{defago_conflict-free_2011}
\vfill
For an implementation, check \textbf{AntidoteDB}.
}
\end{frame}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\begin{frame}{State of the practice}{Path dependency to the ``cloud''}
\begin{block}{The BASE model is fashionable because}
\centering
``\emph{High-performance} object storage for \emph{AI analytics} with PBs of \emph{IoT data streams} at the \emph{edge}, using \emph{5G}.''
% \begin{itemize}
% \item Highest performance
% \item IoT data streams are inherently distributed
% \end{itemize}
\end{block}
\vfill\centering
\includegraphics[width=.9\columnwidth]{figures/minio_edge.png}
\vfill\raggedright
%\begin{block}{}
\begin{itemize}
\item Always backed by cloud: high performance network links.
\item Edge nodes always seen as clients or data sources, not peers.
\end{itemize}
%\end{block}
% There is \textbf{always a central cloud cluster} in these use-cases.
% Hidden constraint: \textbf{high performance inter-node connectivity}.
\end{frame}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% \begin{frame}{A brief history of storage}
% We keep it short because we'll follow chronological order in the next section too.
% \end{frame}
% \begin{frame}{In the beginning, there were \emph{monoliths}}
% \includegraphics[width=.5\columnwidth]{figures/stonehenge.jpg}
% Web applications used to be monolithic:
% \begin{itemize}
% \item One or two servers;
% \item Availability was not an obsession;
% \item Latency was acceptable.
% \end{itemize}
% Relational databases were queens.
% \end{frame}
% \begin{frame}{Then came \emph{expectations}}
% Then, the whole world went online, and suddenly: expectations!
% \begin{itemize}
% \item ``Milliseconds matter.'' (Algolia slogan)
% \item Critical networked services (healthcare, logistics) need 100\% availability
% \end{itemize}
% $\implies$ Microservices \& horizontal scalability.
% \todo{Develop on the `herd not sheep' paradigm a bit.}
% \end{frame}
% \begin{frame}{Distributing state/storage: the remaining unknown}
% The microservices orchestration game works well for \emph{stateless} services.
% However, any application requires \emph{state}, persistent data.
% And this is tough. As we will now see.
% (Not that it's not well studied: distributed storage has always been fashionable.)
% \end{frame}