Reading List Table Version Management & Version Search

Formulates six versioning problems trading off storage and recreation cost in different ways. Proves most are intractable and proposes heuristics drawing from delay-constrained scheduling and minimum spanning tree literature, including the LMG heuristic that has become the standard baseline. Prototype built as the foundation for DataHub. The reference framing for the entire area.

A platform for collaborative data science with first-class support for branching and merging dataset versions, a version graph DAG, and a query language (VQL) for traversing both data and version history. Goes beyond linear temporal-database histories (e.g., Oracle Flashback) by treating derived data products as first-class citizens.

A "bolt-on" approach that adds versioning to an existing relational DBMS, inheriting its query and analytics capabilities for free. Develops multiple data models for representing versioned data and a lightweight partitioning scheme (LyreSplit) that optimizes query latency. A pragmatic contrast to DataHub's clean-slate design.

Shifts the problem from "what bytes differ between two versions?" to "what semantic transformation explains the change?" Generates verifiable, human-readable explanations covering both syntactic changes (column rename, type cast) and semantic ones (value transformations, derivations). Explanations let analysts decide whether to accept changes, propagate them, or roll back — a capability missing from delta-based versioning systems.

Given a table, can we findall other versions of it in a data lake?

Proposes that data change itself deserves to be a first-class analytic dimension. Introduces the change-cube data model that records insertions, deletions, updates, and schema modifications over time. Demonstrates analytic queries that surface controversy, vandalism, data quality issues, and hidden processes — questions impossible to answer with snapshot-only views.

The system that turns the change-cube into an interactive demo: a web front-end over an underlying change-cube database, demonstrated on the full histories of IMDB and Wikipedia infoboxes. Shows what exploratory analytics over change history can feel like in practice.

Extracts, matches, and analyzes the entire history of 3.5M tables on English Wikipedia — 53.8M total table versions. Shows that web tables have rich life cycles: they're created, change shape, move, grow, shrink, and disappear. An empirical foundation for any work on table version search at web scale.

Establishes that the LMG heuristic introduced by Bhattacherjee et al. can perform arbitrarily badly in the worst case, and shows hardness of o(n)-approximation on general graphs even with relaxed storage. Develops polynomial-time approximation schemes for tree-like graphs — motivated by the observation that real-world version graphs arising from typical edit operations have low treewidth.

An early principled study of archiving versioned scientific data, with formal treatment of hierarchical data, keys, and time-stamped storage. Pre-dates the data-lake era but establishes much of the conceptual vocabulary later versioning systems inherit.

Provenance and versioning are closely linked: a version is what you get when you track derivations through time. ProvDB tracks the full graph of datasets, transformations, and analyses across users and over time. Also referenced in the Model Lakes reading list — useful for seeing how the same techniques apply to both tabular data and trained models.