test: test compatible integrations with Python 3.15; update pyproject by julian-risch · Pull Request #4104 · deepset-ai/haystack-core-integrations

julian-risch · 2026-10-15T07:13:46Z

Related Issues

Move remaining integrations to Python 3.15 once upstream dependencies support it #4110 tracks the integrations that stay on 3.14.

Proposed Changes:

Python 3.15.0 was released on 2026-10-09, and haystack-ai supports it since deepset-ai/haystack#13054. This PR follows #3001, which did the same for 3.14:

79 integrations whose dependencies install and import on 3.15: TEST_MATRIX_PYTHON now tests 3.15 instead of 3.14 as the maximum version, and pyproject.toml declares the Python :: 3.15 classifier. This includes pinecone's inline matrix (index name unchanged), docling_serve's static matrix and llama_stack (["3.12", "3.15"]).
cohere and ollama: add pydantic>=2.14.0; python_version >= '3.15'.
- Their pydantic>=2.12.0 floor made the lowest-direct job on 3.15 resolve pydantic-core 2.41.1, which uses PyO3 0.26 and doesn't build on 3.15.
- pydantic 2.14.0 is the first release whose pydantic-core has cp315 wheels.
- The floor on 3.11–3.14 is unchanged.
Templates: scripts/utils/templates/workflow.yml and pyproject.toml use 3.15 for new integrations.

These integrations stay on 3.14, without the 3.15 classifier, until upstream ships:

Blocker on 3.15	Integrations
onnxruntime has no cp315 wheels (microsoft/onnxruntime#33271)	chroma, fastembed, markitdown, optimum
torch has no cp315 wheels on PyPI. The ones on download.pytorch.org segfault in non-main threads on macOS and Windows (pytorch/pytorch#200779)	docling, hanlp, sentence_transformers, transformers
torchaudio has no cp315 wheels	funasr
Rust extensions built with PyO3 < 0.29 don't build on 3.15 (fastuuid, jiter, ormsgpack, pydantic-core, tiktoken)	litellm, dspy, ragas, cognee
ddtrace declares `requires-python <3.15`	datadog
`spacy.load()` fails on 3.15 (explosion/spaCy#14047)	spacy, presidio, unstructured
ibm-watsonx-ai declares `requires-python <3.15`	watsonx
weaviate-client 4.23.1 fails on import: `class BatchCollectionProtocol(Generic[Properties], Protocol[Properties])` raises "Cannot inherit from Generic[...] multiple times"	weaviate
No CPU build of vLLM for 3.15	vllm
jq 1.12.0 has no cp315 Windows wheel, and its sdist doesn't build on Windows	mirage
valkey-glide has no cp315 wheels, and its sdist needs `protoc`	valkey
The openai-whisper sdist doesn't build on 3.15, and it needs torch	whisper
Our own `oracledb<3.0.0` pin has no cp315 wheels. oracledb 26.x has them, so relaxing the pin is a separate change	oracle
Already tested below 3.14	chonkie, langfuse, paddleocr, nvidia

How did you test it?

Resolved every integration's dependencies for CPython 3.15 on Linux, macOS and Windows with uv pip compile, and checked every pinned version on PyPI for cp315, abi3 or pure-Python wheels. The only gap that affects every integration is PyYAML 6.0.3, which has no cp315 wheels yet and installs from its sdist as pure Python, the same way haystack-ai does.
Locally on CPython 3.15.0 (macOS arm64), I ran hatch run test:unit, then the nightly path with haystack-ai from main:
- anthropic: 147 passed
- qdrant: 140 passed
- weaviate: weaviate-client fails on import, so weaviate is excluded
This PR's CI runs every changed workflow on 3.11 and 3.15 on all configured OSes.
- The first run covered 83 integrations. vllm, mirage, valkey and whisper failed on 3.15 and are back on 3.14, and the cohere and ollama lowest-direct jobs are fixed as described above.
- The remaining failures are integration tests on 3.11:
  - togetherai, perplexity, huggingface_api, astra, openrouter and hetzner: these same tests also fail on main.
  - ollama: one test failed because the local model refused to answer (assert '42' in "I'm sorry, ...").

Notes for the reviewer

Based on test: use Python 3.11 in tests instead of 3.10 #4103, which moves the minimum tested version to 3.11. GitHub retargets this PR to main once test: use Python 3.11 in tests instead of 3.10 #4103 is merged.

Checklist

I have read the contributors guidelines and the code of conduct
I have updated the related issue with new insights and changes
I added unit tests and updated the docstrings — n/a, CI and metadata only
I've used one of the conventional commit types for my PR title: fix:, feat:, build:, chore:, ci:, docs:, style:, refactor:, perf:, test:.

🤖 Generated with Claude Code

haystack-ai main (3.5.0rc0) and the upcoming 3.4.0 require Python >=3.11, so the nightly "run unit tests with Haystack main branch" step fails on every 3.10 leg. Move the minimum tested version to 3.11 in all workflows and the workflow template. llama_stack's push default and coverage gates use 3.12, its minimum version. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

Python 3.15.0 was released on 2026-10-09. Integrations whose dependencies install on 3.15 now test 3.15 instead of 3.14 as their maximum version and declare the 3.15 classifier. Integrations blocked by missing cp315 wheels upstream (torch, torchaudio, onnxruntime, PyO3 < 0.29 extensions, packages capped below 3.15) stay on 3.14. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

github-actions · 2026-10-15T09:58:31Z