Hi community,

An issue was recently reported in #921 regarding the strict PyArrow version 
constraints in the TsFile Python package. These constraints may cause 
dependency conflicts when users install newer versions of PyArrow:


https://github.com/apache/tsfile/issues/921


We reviewed the current implementation and did not find any code-level reason 
that requires an upper bound on the PyArrow version. Although the 
implementation uses PyArrow's `_import_from_c` and `_export_to_c` APIs, we 
tested them with multiple PyArrow versions and found no compatibility issues.


Therefore, we propose adjusting the Python dependency policy as follows:


1. Do not set upper version bounds for NumPy, pandas, or PyArrow.
2. Set only the minimum supported dependency versions according to the Python 
version.
3. If a future dependency release introduces an actual compatibility issue, we 
should first try to update the TsFile implementation to support it. A version 
restriction can be added later if necessary, based on the specific issue.


The proposed minimum PyArrow versions are:


- Python < 3.13: PyArrow >= 16
- Python 3.13: PyArrow >= 18
- Python >= 3.14: PyArrow >= 22


We have also adjusted the minimum NumPy and pandas versions according to their 
Python version support and aligned the build-time and runtime dependency 
constraints.


For verification, we tested both the minimum supported dependency combinations 
and the latest available dependency combinations. All related DataFrame tests 
passed (27/27) for every tested combination.


The complete remote test and build results are available here:


- Python unit-test matrix, 9/9 passed:
  https://github.com/ColinLeeo/tsfile/actions/runs/32955466173


- Python wheel-build matrix, 11/11 passed:
  https://github.com/ColinLeeo/tsfile/actions/runs/32955469256


The tests cover Linux, macOS, and Windows with multiple supported Python 
versions. The wheel builds cover Linux x86_64/aarch64, macOS x86_64/arm64, and 
Windows with Python 3.9–3.14.


To support the minimum NumPy version, 1.26, we also adjusted several DataFrame 
test assertions. Before calling `np.allclose`, pandas nullable floating-point 
Series are now explicitly converted to NumPy arrays. This change only affects 
test-side data conversion. It does not change the production implementation, 
actual read/write behavior, or assertion coverage.


Comments and suggestions on this dependency policy and the proposed version 
ranges are welcome. If no additional compatibility issues are identified, we 
plan to update the dependency configuration of the TsFile Python package 
accordingly.


Thanks,
Colin

Reply via email to