VLDB

\documentclass{vldb}
\usepackage{graphicx}
\usepackage{balance}
\usepackage{amsmath,amssymb}
\usepackage{booktabs}
\usepackage[T1]{fontenc}

\vldbTitle{Kairos: Efficient Time-Travel Queries over Versioned Data Lakes}
\vldbAuthors{First Last, Jane Doe, John Smith}
\vldbDOI{https://doi.org/XX.XX/XXX.XX}
\vldbVolume{18}
\vldbNumber{5}
\vldbYear{2026}
\vldbAvailabilityUrl{https://github.com/example/kairos}

\begin{document}

\title{Kairos: Efficient Time-Travel Queries over\\Versioned Data Lakes}

\numberofauthors{3}
\author{
\alignauthor First Last\\ \affaddr{University of Example}\\ \email{[email protected]}
\alignauthor Jane Doe\\ \affaddr{Example Research Labs}\\ \email{[email protected]}
\alignauthor John Smith\\ \affaddr{University of Example}\\ \email{[email protected]}
}
\maketitle

\begin{abstract}
Versioned data lakes (Iceberg, Delta Lake, Hudi) support time-travel
queries, but performance on historical snapshots degrades quadratically
with version depth. Kairos is a new query engine extension that
pre-computes sparse materialized deltas and routes point-in-time queries
through the optimal subset. On an Iceberg deployment with 10k snapshots,
Kairos achieves a 17$\times$ p50 speedup at 8\% additional storage.
\end{abstract}

\section{Introduction}
Time-travel queries are essential for audit, compliance, and ML training
reproducibility. Existing engines effectively re-derive the snapshot on
every historical read.

\section{Background}
Apache Iceberg snapshots, Delta Lake checkpoints, bitemporal databases.

\section{Design}
Kairos builds sparse delta materializations across the version chain.
Queries are routed through the optimal subset of deltas.

\subsection{Cost Model}
The optimizer minimizes
\begin{equation}
  C(\mathcal{D}) = \alpha |\mathcal{D}| + \beta \sum_{q \in Q} \text{readCost}(q, \mathcal{D}).
\end{equation}

\section{Implementation}
Kairos is 11{,}800 lines of Java extending Apache Iceberg 1.5.

\section{Evaluation}
\begin{table}[t]
\centering\small
\begin{tabular}{lrrr}
\toprule
System & p50 (ms) & p99 (ms) & Storage OH \\
\midrule
Iceberg stock        & 2{,}840 & 16{,}200 & 0\% \\
Periodic checkpoint  & 1{,}250 &  8{,}400 & 24\% \\
\textbf{Kairos}      & \textbf{168} & \textbf{640} & \textbf{8\%} \\
\bottomrule
\end{tabular}
\end{table}

\section{Related Work}
Bitemporal databases, delta encoding, learned indexing on versioned data.

\section{Conclusion}
Principled sparse delta materialization makes time-travel on versioned
data lakes fast enough for interactive use.

\balance
\bibliographystyle{abbrv}
\bibliography{refs}
\end{document}