Beacon 2.0.0-rc3 pre-release

[robinskil]

Released. v2.0.0-rc3 came out on 2026-08-19.
The open work moves to #413, the rc4 pre-release. The rc4 target date is 2026-08-21.

The publish is partial. The tag started three publish workflows. Only one passed.
The container image reached the registry. @beacon/client did not reach npm, because npm publish
returned ENEEDAUTH. beacon-datalake-cli did not reach PyPI, because ruff returned E501 on
line 66 of repl/reader.py. Both jobs get a fix in #413.

These tasks stay open. Each one moves to #413.

• CHANGELOG.md holds no 2.0.0-rc.2 section, and no 2.0.0-rc.3 section. Steps 2 and 3 below.
• beacon-db/beacon-db-py/pyproject.toml still reads 2.0.0rc2. Step 7 below.
• scripts/bump-version.py and scripts/check-version.py both skip beacon-db-py. That is the cause.
• The tag reads v2.0.0-rc3, and Cargo.toml reads 2.0.0-rc.3. The two forms still disagree.
• The release notes hold the auto-generated pull request list. Step 22 below.
• The ODV and BBF projection bug (ODV and BBF scans apply only part of a pushed-down projection #382) is open. Pull request Apply the whole pushed-down projection in the ODV and BBF scans (#382) #398 holds the fix.
• Stage 1 of Release tests: Python tests for the HTTP API and BeaconDB #399 gave integration-tests/formats/, with 13 files and 158 tests. The driver layer is open.

---

Parent issue: #389

Publish the third pre-release of 2.0.0 this week. The target date is 2026-08-21.

The last pre-release is v2.0.0-rc2
of 2026-08-03. main holds 27 commits after that tag.

Contents after v2.0.0-rc2

Added

• Read Apache Iceberg tables, through an external table or read_iceberg (Read Apache Iceberg tables: external tables and read_iceberg #357)
• Read Icechunk repositories, through the Zarr reader (Read Icechunk repositories through the Zarr reader #356)
• A pure-Rust netCDF reader behind a flag (feat(netcdf): add the oxcdf pure-Rust reader behind a flag #337)
• A pure-Rust HDF5 reader behind a flag (Optional pure-Rust HDF5 reader via oxcdf-hdf5 #358)
• A persistent statistics store per file, plus file pruning at scan time (Persistent per-file statistics: a store, a background collector, and scan-time file pruning #347)
• Column ranges from netCDF and HDF5 files (Prune HDF5 and netCDF files with their recorded statistics #360)
• Column ranges from Zarr stores (File statistics: Zarr supplies column ranges #362)
• A statistics collector at boot, and a view of its state (See what file statistics are doing, and collect them at boot #370)
• One provider for every file scan (One provider for every file scan, pruning before it reads #364)
• An nd scan across every partition, through a work queue (Divide an nd scan across every partition with a work queue #371)
• A schema cache, so a query reads each file once (Keep each file's schema, so a query does not read the file twice #375)
• 123 spatial functions with PostGIS names (Use datafusion-spatial for 123 spatial functions with PostGIS names #379)
• An Arrow decoder and streams in the TypeScript client (Expose Arrow decoder and streaming capabilities in TypeScript SDK #346)
• Bounding box pushdown for GeoParquet (Fix the GeoParquet scan, and add bounding box pushdown #383)
• A pull request template with a disclosure for generative AI (Add a pull request template #384)

Changed

• The admin UI renders a result from the Arrow columns (Render admin UI results from Arrow columns #381)
• The minimum Rust version is 1.94, up from 1.91
• A build from source needs PROJ 9.6.2 and pkg-config
• Fewer dependencies, and a shorter build (Features/deps cleanup #380)
• The docs state one product, and the README is new (Reposition Beacon as one product: a query engine and server for scientific data #334, Rewrite the README #335)

Fixed

• The GeoParquet scan applied only part of a projection (Fix the GeoParquet scan, and add bounding box pushdown #383)
• The admin UI ignored the server URL prefix (Admin web UI respects the server URL prefix #369)
• File statistics pruned no netCDF or HDF5 file (Prune HDF5 and netCDF files with their recorded statistics #360)
• A netCDF time variable kept its _FillValue cells as dates (fix(netcdf): null the _FillValue cells of a CF time variable #343)
• A netCDF backend reported no chunk shape (fix(netcdf): report the chunk shape of the file #341)
• The nd chunk shape function cut only the first axis (fix(nd): fill the chunk shape from the last axis #344)
• An nd batch lost rows after a coalesce (fix(nd): decode every row of a coalesced nd batch #336)
• The nd worker test failed at random (Assert the nd worker wait directly, not by a worker count #376)

Merge before the tag

• Close the OIDC claim bypass, and update the dependencies (Close an OIDC claim bypass and update dependencies for security #385). This is a security fix.
• Update js-yaml and @redocly/openapi-core (Bump js-yaml and @redocly/openapi-core in /beacon-clients #388)
• Update postcss (Bump postcss from 8.5.6 to 8.5.26 in /docs #387)
• Update nanoid (Bump nanoid from 3.3.15 to 3.3.18 in /beacon-clients #386)

Decide before the tag

• ODV and BBF scans apply only part of a projection (ODV and BBF scans apply only part of a pushed-down projection #382). Fix it in rc3, or hold it for 2.0.0.
• A schema merge depends on the file the disk answers first (A schema merge in five formats depends on which file the disk answers first #377). Same choice.

Both bugs block 2.0.0. A pre-release is a good place to test the fix.

Tests, stage 1 of #399

#399 plans a Python test suite above 1000 tests. Stage 1 lands in rc3.

• Add the driver layer. One test body runs on the HTTP API and on BeaconDB.
• Move the 180 API tests onto the driver.
• Add the generators for netCDF, Zarr, Parquet, CSV and GeoParquet.
• Add 48 file shapes for those 5 formats. A netCDF file is gridded or ragged.
• Add the 8 core operations over each shape. About 768 tests.
• Write the what, why and how table in integration-tests/README.md.

Steps

1. Merge the four pull requests above.
2. Rename the [Unreleased] section in CHANGELOG.md to 2.0.0-rc.3.
3. Add the section that 2.0.0-rc.2 never got, or state that rc2 holds only doc changes.
4. Set the version to 2.0.0-rc.3 in Cargo.toml.
5. Set the version to 2.0.0-rc.3 in beacon-clients/beacon-ts/package.json.
6. Set the version to 2.0.0rc3 in beacon-clients/beacon-datalake-cli/pyproject.toml.
7. Set the version to 2.0.0rc3 in beacon-db/beacon-db-py/pyproject.toml.
8. Rename docs/docs/2.0.0-rc2/ to docs/docs/2.0.0-rc3/.
9. Update every 2.0.0-rc2 link in docs/.vitepress/config.mts.
10. Leave LATEST_VERSION at 1.8.0 in docs/.vitepress/theme/version.js. A pre-release is not the latest.
11. Run cargo +1.94 fmt --check and cargo +1.94 clippy --workspace.
12. Run cargo +1.94 test --workspace.
13. Run the integration tests with integration-tests/run.sh.
14. Run stage 1 of the test matrix (Release tests: Python tests for the HTTP API and BeaconDB #399). Record the result here.
15. Run cargo audit. Record every advisory that stays open.
16. Tag the release, and push the tag.
17. Check the three publish workflows.
18. Check the container image for linux/amd64 and linux/arm64.
19. Check @beacon/client on the npm rc tag, not on latest.
20. Check beacon-datalake-cli on PyPI. pip install --pre finds it.
21. Mark the GitHub release as a pre-release.
22. Write the release notes. rc2 got only a compare link.

Open question: the tag name

The tags read v2.0.0-rc1 and v2.0.0-rc2. Cargo.toml reads 2.0.0-rc.2. The workflow comments
read v2.0.0-rc.1. The three forms do not agree.

The container tag comes from the git tag, and the crate version must match what CI verifies. Pick one
form for rc3, and use the same form until 2.0.0. v2.0.0-rc.3 matches SemVer and the crate version.

Reference

• The publish workflows accept any SemVer pre-release suffix. See the comments in
  .github/workflows/publish-beacon-server.yml.
• CI verifies the version. It does not write the version.

[robinskil]

Released as v2.0.0-rc3 on 2026-08-19. The remaining work moves to #413, the rc4 pre-release, with the target date 2026-08-21.