[scikit-learn] version warning - do I have to fix the minor version when unpickling?

[Guillaume Lemaître]

On Thu, 15 Jun 2023 at 08:47, Martin Gütlein <guetlein at posteo.de> wrote:

> Hi Guillaume,
>
> thanks for your swift reply. So that means we should even fix the patch
> version, right?
>

Indeed, this is the safest.

Be aware that you can always try a newer version but it would require a good
stress test to be sure that the underlying private API did not change. I can
understand that it might be difficult to do so.

Regarding versioning, when releasing a new patch version, we only include
bug fixes. So it should be limited changes and localized change but we do
not constrain ourselves to not change the private functions if it is
required.


> And, this goes into both directions? So the service executing the code
> should not have a newer version, not even a patch, then the version the
> pickeled model was built with?
>

Indeed, once you unpickled the model and interact with it, you should be in
an environment that has the same version as the pickle to be safe.


>
> Cheers, Martin
>
>
> Am 14.06.2023 18:24 schrieb Guillaume Lemaître:
> > Hi Martin,
> >
> > The public API is stable but the internal can change which can affect
> > the pickle.
> > For instance, calling a missing private function that does not exist
> > can happen.
> > Since that we don’t guarantee any support in this regard, this is
> > the reason why a warning is raised even between minor or patch
> > version.
> >
> > Cheers,
> >
> > On Wed, 14 Jun 2023 at 09:34, Martin Gütlein <guetlein at posteo.de>
> > wrote:
> >
> >> Hi,
> >>
> >> I just got the warning:
> >>
> >> "site-packages/sklearn/base.py:318: UserWarning: Trying to unpickle
> >>
> >> estimator DecisionTreeRegressor from version 1.1.2 when using
> >> version
> >> 1.2.2. This might lead to breaking code or invalid results. Use at
> >> your
> >> own risk".
> >>
> >> because I only fixed the major version (i.e, my dependency is
> >> "scikit-learn>=1,<2"). I assumed that code version 1.2 is
> >> compatible to
> >> pickeled models from version 1.1. Is this not true?
> >>
> >> Do I have to change to fix the minor version instead (i.e., use
> >> "scikit-learn>=1.2,<1.3")?
> >>
> >> Kind regards,
> >> Martin
> >>
> >> _______________________________________________
> >> scikit-learn mailing list
> >> scikit-learn at python.org
> >> https://mail.python.org/mailman/listinfo/scikit-learn [1]
> >
> > --
> >
> > Guillaume Lemaitre
> >
> > Scikit-learn @ Inria Foundation
> > https://glemaitre.github.io/ [2]
> >
> > Links:
> > ------
> > [1] https://mail.python.org/mailman/listinfo/scikit-learn
> > [2] https://glemaitre.github.io/
> >
> > _______________________________________________
> > scikit-learn mailing list
> > scikit-learn at python.org
> > https://mail.python.org/mailman/listinfo/scikit-learn
>


-- 
Guillaume Lemaitre
Scikit-learn @ Inria Foundation
https://glemaitre.github.io/
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://mail.python.org/pipermail/scikit-learn/attachments/20230615/35ffbddb/attachment-0001.html>