Skip to content

Releases

Release v0.1.0

Choose a tag to compare

@homer6 homer6 released this 16 Nov 07:25
454f519

Released 2026-11-16. The first release. The sandbox runtime for a Lake workspace's own methodologies: the interface a methodology implements, the methodology command the stage, run and collect pods use, the checks that run before a methodology's code is imported, and model serving. Because it is the first release, this page lists everything in it.

What it includes

The methodology interface (interface version 1)

  • orama_methodology: a methodology implements fit, predict, score and, optionally, explain, and receives its data as a Data object (interface.md).
  • The manifest methodology.json and its checks: every problem is reported at once, not only the first.
  • The lock file: every requirement pinned with == and at least one --hash=sha256:; anything else is refused.

The methodology command (cli.md)

Command Where it runs What it does
stage stage pod pages the run's data from Lake (methodology_run_input) into the run volume
run run pod (the sandbox) checks the sandbox, then runs the methodology on the staged data; always writes result.json and log.txt
collect collect pod posts the result once (methodology_run_output); an artifact over 24 MiB, up to 1 GiB, goes first through methodology_run_artifact in 8 MiB chunks, named by size and SHA-256
stage-model model deployment pages a deployment's artifact (methodology_deployment_input), checking its size and SHA-256 before it is moved into place
serve model deployment (the sandbox) the same sandbox checks, then GET /healthz, POST /predict and POST /explain; one call into the code at a time, rows never logged
validate-manifest local checks a methodology directory
install-locked image builds installs a hash-locked lock file (wheels only) into /methodology/.packages

methodology --version reports methodology-runtime 0.1.0, and the Lake calls carry User-Agent: orama-methodology-runtime/0.1.0. Calls to Lake retry on 429, 5xx and connection errors; a conflict "already recorded" on collect counts as success.

The sandbox (sandbox.md)

  • run and serve refuse to start (exit 3) before importing any of the methodology's code when they find a credential-looking environment variable, a mounted service account or run credential file, or network the run was not granted. A refusal names the variable or path, never its value.
  • Only stage and collect talk to Lake, and only to the calls above. run never opens a connection.
  • Security fixes (#6, for #4 and #5):
    • stage and collect never import from /work, the volume the sandboxed run writes: bin/methodology starts Python with -P, the image sets PYTHONSAFEPATH=1, and the package removes the working directory and anything under /work from sys.path before any other import.
    • collect reads no symbolic link: the output directory and each output file are opened with O_NOFOLLOW and used only when they are regular files, and a large artifact is hashed and sent from the descriptor that was checked.

The images

  • Dockerfile: the image oramasearch/methodology-runtime, on python:3.12.12-slim-bookworm pinned by its multi-platform index digest, non-root uid 10001, the standard library only.
  • templates/methodology.Dockerfile: the image Lake builds for each methodology, FROM this one.
  • examples/: a working standard-library logistic regression and a broken example.

Licence and project files

  • MIT LICENSE and COPYRIGHT, an SPDX-License-Identifier: MIT line in every source file, THIRD_PARTY_LICENSES, CONTRIBUTING.md, CLA.md and CODE_OF_CONDUCT.md (#10).
  • SECURITY.md: how to report a vulnerability privately (#8).
  • The public tree carries nothing internal (#9).
  • A software bill of materials, sbom.spdx.json and sbom.cdx.json, regenerated by tools/sbom.py.

Pull requests

PR Change
#2 The runtime: interface, harness, sandbox checks, stage and collect, the image (#1)
#3 Serve a trained model; artifacts over 24 MiB in chunks
#6 Security: stage and collect never import from /work; collect reads no symbolic link (#4, #5)
#8 SECURITY.md
#9 Internal material kept out of the public tree
#10 MIT licence, SPDX headers, third-party notices and contributor documents
this release the release notes, a test that the version has its notes, the SBOM regenerated

What it does not include

  • No pull request was open at the cut, so nothing was deferred.
  • No image with a licence-clean base (below).

Known issues

  • The image is not published (#11, a release blocker for the open-source release). It is built on python:3.12-slim (Debian), which bundles GPL and LGPL userland (glibc, bash, coreutils, dpkg, readline). It moves to scratch or another static base with static binaries before the image is made public. templates/methodology.Dockerfile follows it.

Verification

Proof Result
python3 -m unittest, Python 3.15 63 tests, all pass
python3.12 -m unittest, Python 3.12 (the image's version) 63 tests, all pass
tools/spdx-headers.py --check passes
tools/sbom.py --check passes, regenerated after the last change
Image build from the tag, on the cluster BuildKit (linux/amd64): oramasearch/methodology-runtime:v0.1.0 (also :12209b3) = sha256:8fa7a37d724cadf8bc6f166d63d75de1caa287344351181788d0c0ca81cb9ea5, in the private Docker Hub repository only

Not run: tools/sandbox-smoke.sh (it needs a local container engine).