[prrao87]

Note

The lance-csr-catalog crate is now retired. We only need to depend on lance-csr instead.

Why

Upcoming work on frontend and ingestion UIs will let applications that store graphs as plain Lance tables share one graph contract from lance-csr, instead of each keeping its own copy of the mapping types, its own mapping checks and its own code for opening tables at a pinned version. This PR prepares lance-csr for that, ahead of a 0.8.0 release.

Two things stand in the way:

• lance-csr-catalog doesn't really require a crate of its own. It is about 240 lines, lance-csr is its only dependent, and publishing two crates in order adds release steps and a recovery procedure for when only the first one publishes.
• Querying through a namespace always opened the latest version of each table and dropped the storage options the namespace returned, so a namespace could not pin a snapshot.

What changed

• The catalog lives in lance-csr. GraphSourceCatalog, InMemoryCatalog, SimpleTableSource and DirNamespace move into the lance_csr::source_catalog and lance_csr::namespace modules. The root re-exports are unchanged, so Python users and code that imports lance_csr::DirNamespace see no difference. lance-csr-catalog is no longer built, version-bumped or published, and the release guide drops its two-crate recovery steps.
• Namespace reads honour pinned versions. CypherQuery::execute_with_namespace accepts any LanceNamespace, such as Lance's own DirectoryNamespace, and opens each table at the version and with the storage options describe_table returns. With no version it opens the latest, as before, which is always the case for DirNamespace.
• A shared graph mapping. A new mapping module adds:
  • GraphMapping, the version 1 JSON graph descriptor (label to table, key, display column, and relationship endpoints), with validate() for the rules that need no table data: at least one node label, at most 64 labels, labels that are ASCII identifiers unique ignoring case, one graph role per table, and endpoints that name a mapped node label.
  • GraphMapping::graph_config, which builds a GraphConfig from the tables' Arrow schemas and checks that key, display and endpoint columns exist, keys are scalar, and endpoint types match the node keys.
  • GraphSnapshot, a list of tables pinned to a path and Lance version plus the mapping. It ignores unknown fields, so it reads a publication manifest that also carries row counts or schemas. open() opens every mapped table at its pinned version and returns what CypherEngine::new takes.
• open_lance_dataset is public, as the one helper for opening a table at a pinned version.

Breaking (Rust only)

• The lance-csr-catalog crate is gone; depend on lance-csr instead.
• execute_with_namespace takes impl LanceNamespace and execute_with_namespace_arc takes Arc<dyn LanceNamespace>. Existing calls with a DirNamespace or Arc<DirNamespace> still compile.

The version bump to 0.8.0 follows in its own PR, as RELEASE.md describes. The crates.io page for lance-csr-catalog will keep its 0.7.1 README until a final version pointing to lance-csr is published by hand.

Validation

• cargo test --workspace: all pass, including new tests for the mapping checks, reading a descriptor and a publication manifest, opening a snapshot at a pinned version, and a namespace that pins version 1 after the table was overwritten.
• cargo clippy --workspace --all-targets -- -D warnings and cargo fmt --all --check: clean.
• Python: make install, make test (1706 passed) and make lint: clean.
• uv run --project docs --locked zensical build --strict and uvx typos@1.26.0 .: clean.
• CI pending.

🤖 Generated with Claude Code

[cloudflare-workers-and-pages]

Deploying lance-csr with    Cloudflare Pages

Latest commit:	06a0b30
Status:	⚡️  Build in progress...

View logs