[RFC][DISCUSS] Python3 Timeline of 2019

[tqchen]

It has been half a year since our last discussion #1602

Many projects officially deprecate python2 starting 2019. I would like to open this RFC again to see the community's opinion on this.

As per the previous discussion, embracing python3 allows us to make use of new features such as type annotation that enables better DSL, and makes cleaner code. It also makes us consistent with the rest of the python eco-system. Let us discuss two things:

When should the move happen

Such a big change will likely need to happen in a major release

• Do it now in the 0.6 release cycle since 0.5 has been released in 2019
  • We need to send out early notifications to the user that this is happening and have clear error messages
• Do it in the 0.7 release cycle after 0.6 release

Which minimum version of python3 to depend on

• python3.5 This is the safest choice as suggested by many in the previous discussion
• python3.6 More features, type class support(annotate class members), although not strictly necessary in all places

[zhiics]

When is the plan for 0.6? I think we can probably do it in 0.6 if it is several months/half a year away. For transitioning, python 3.6?

A side problem, do we have any plan to enable/support c++ 14/17?

[antinucleon]

0.6: + 1
C++ 14/17: Not hurry because we have most of features in dmlc-core. Most production system is still running at GCC-5

[FrozenGene]

I think we can do it in 0.6 in dev module but make sure runtime be compatibility with Python 2 in 0.6

And in 0.7, we make dev module / runtime module be Python 3.

I would recommend Python 3.5, we should do the safest choice.

[were]

0.6 +1.
pip for Python2 will be fully deprecated next year.
I believe spurning Python2 would be the earlier the better.
Many Python libraries have compatibility issues between 2 and 3.
I often suffer from handling those corner cases.
I often fail CI because of not handling those corner cases.
This is annoying.

I have no strong preference, and I am leaning to 3.6 more.

[junrushao]

0.6 + 1
Deprecation won't be a significant influence, because Python is merely just a frontend, by disabling which won't hurt deployment.

[kevinthesun]

0.6 +1
3.5 is a safe choice.

[yongwww]

0.6 +1

[kazum]

0.6 +1
Since we don't have enough python2 tests in our CI, some frontends already doesn't work well with python2. Rather than fixing the problem, dropping the python2 support and refining codes with python3 features for 0.6 is much more constructive.

I prefer python 3.6, but 3.5 as a safe choice is also fine.

[tqchen]

Seems we are converging toward drop python2 in 0.6 the cycle, in terms of versions perhaps we can start by requiring 3.5