CI: py3.9 test job fails — unbounded torch constraint no longer resolves on Python 3.9

[jhamon]

Problem

The Run tests (ubuntu-latest, 3.9) matrix cell fails at poetry install with:

So, because pinecone-text depends on torch (>=1.13.1), version solving failed.
##[error]Process completed with exit code 1.

All other matrix cells (3.10/3.11/3.12 on ubuntu/macos/windows) pass. This is a pre-existing environmental failure, not introduced by any recent PR — it surfaces on every PR that runs CI.

Root cause

pyproject.toml:

python = ">=3.9,<4.0"
torch  = { version = ">=1.13.1", optional = true }

The torch >=1.13.1 constraint is unbounded. Newer torch releases (2.3+) dropped Python 3.9 support (they require python >=3.10), so Poetry cannot find a single torch version satisfying the full dependency graph on 3.9 → resolution fails.

Options (needs a support-matrix decision)

1. Drop Python 3.9 from the support matrix and python constraint (3.9 is EOL Oct 2025). Simplest; aligns with upstream ecosystem.
2. Cap torch per Python version, e.g. torch = [{version=">=1.13.1,<2.3", python="3.9"}, {version=">=1.13.1", python=">=3.10"}], to keep 3.9 working.

Recommendation: option 1 (drop 3.9) unless there is a downstream consumer still on 3.9.

Context

Surfaced while clearing CI feedback on the least-privilege-permissions PR (#99), which only edits workflow permissions: blocks and cannot cause this failure. Deferred here because the fix is a dependency/support-matrix decision, out of scope for that PR.

Failing job: https://github.com/pinecone-io/pinecone-text/actions/runs/29167061936/job/86581892240

[jhamon]

Folding in the exact transitive chain from the duplicate #102 (now closed as dup of this issue), captured from a fresh failing run on 2026-07-15:

torch >=1.13.1  →  triton 3.3.1  →  setuptools >=40.8.0
setuptools 83.0.0 requires Python >=3.10  →  forbidden on 3.9  →  torch forbidden  →  solve fails

So the resolver now escalates the transitive setuptools floor to 83.x (via torch → triton), and 83.x drops Python 3.9. This confirms the two live options already listed here:

1. Drop Python 3.9 (EOL Oct 2025) — bump python = ">=3.10,<4.0" and remove the 3.9 matrix leg. Public support-matrix change; recommended.
2. Keep 3.9 — cap torch (or setuptools) with a python="3.9" marker so the older, 3.9-compatible chain resolves.

Both change a published package's constraints, so this needs a maintainer/support-policy call before shipping — it is not a CI-job-local pin like ts-client #398. Blocks #99 and every pinecone-text PR that runs the 3.9 leg.

[jhamon]

Cap-torch approach (PR #104) is proven insufficient — recommend Option 1 (drop py3.9)

PR #104 capped torch <2.3 on Python <3.10. It fixes Poetry resolution on ubuntu-3.9, but CI still fails on macos-3.9 and windows-3.9 (16 test failures), for two intrinsic reasons:

1. torch 2.2.2 is built against NumPy 1.x → RuntimeError: Numpy is not available under NumPy 2.x.
2. transformers refuses torch.load unless torch ≥ 2.6 (CVE-2025-32434) → all SPLADE tests fail.

This is a hard contradiction: linux-3.9 needs torch < 2.3 to resolve (triton→setuptools py≥3.10 chain), but transformers needs torch ≥ 2.6. No torch version can make the py3.9 matrix cell pass.

Recommendation: Option 1 — drop Python 3.9. It reached EOL 2025-10-31; the current ML stack no longer supports it. This is a user-facing support-matrix change and needs maintainer sign-off. PR #104 is now in draft pending that decision.

Note: PR #99 (least-privilege CI permissions) is also red solely on ubuntu-3.9 for the same resolve failure — it will go green automatically once this is resolved.