[Roadmap] 1.2.0 Roadmap

[hcho3]

@dmlc/xgboost-committer Please add your items here by editing this post. Let's ensure that

• Each item has to be associated with a ticket
• Major design/refactoring are associated with a RFC before committing the code
• Blocking issue must be marked as blocking
• Breaking change must be marked as breaking

For other contributors who have no permission to edit the post, please comment here about what you think should be in 1.2.0.

• GPU port of Accelerated Failure Time objective for survival analysis (GPU implementation of AFT survival objective and metric #5714)
• Optimized JSON IO (Implement fast numeric printing routines. #5772).
• Callback function for device DMatrix (Use DMatrix Proxy for implementing data callback. #5629).
• Device DMatrix for dask interface ([FEA][Python] Multi-GPU training with DeviceQuantileDMatrix  #5583, Dask device dmatrix #5901).
• Integrate dask interface with Python asyncio. (dask async operation with XGBoost failed (with Rapids on GPU) #5639, [BLOCKING] [Dask] Asyncio support. #5862)
• Upgrade rabit. (Update rabit. #5680)

Blocking pull requests

• [BLOCKING] Remove to_string. #5934
• [BLOCKING] Handle empty rows in data iterators correctly #5929
• [BLOCKING] [jvm-packages] add gpu_hist and enable gpu scheduling #5171
• [BLOCKING] Thread-safe prediction by making the prediction cache thread-local. #5853
• [BLOCKING] [R] Add a compatibility layer to load Booster object from an old RDS file #5940
• [BLOCKING] [Dask] Asyncio support. #5862
• [Breaking] Fix custom metric for multi output. #5954

Breaking change:

• From now on R and Python custom evaluation metric will receive raw prediction. [Breaking] Fix custom metric for multi output. #5954

[thvasilo]

Regarding distributed learning, Databricks has this big warning on their XGBoost docs: a cluster running Spark has to be re-started if an XGBoost job fails.

This is bad user experience, as a shared cluster might be running a number of other things and it would have to be shut down when a job fails, killing everything else.

The related issue #4826 hasn't had an update in a long time, so not sure if that's still the case with 1.1.0.

[ldesreumaux]

Is there any plan on making the Json serialization method no-longer "experimental" in version 1.2.0? What is the current status on this? Is the Json format currently stable or should we expect big changes in future releases?

[thvasilo]

I'll also bring up re-visiting the default parameters issue, #4986 and the linked issue there, 1.2.0 might be a good point to at least change the min_child_weight.

[hcho3]

@ldesreumaux We would like to remove the experimental label after we make JSON serialization robust. #5772 adds a whole suite of tests to ensure that floating-point values are preserved by JSON serialization.

[trivialfis]

I'll also bring up re-visiting the default parameters issue, #4986 and the linked issue there, 1.2.0 might be a good point to at least change the min_child_weight.

I think we will have to do it in next release. Which is not a bad thing, I'm targeting cat data support so we might unify a lots of internal code and will be able to support a unified set of parameters.

[thvasilo]

Is categorical data support planned for 1.2.0 @trivialfis ?

[hcho3]

@dmlc/xgboost-committer It's been more than 2 months since we release 1.1.0, and we made 130 commits since then. Let's think about releasing 1.2.0 soon. Consider which pull requests are blocking.

[CodingCat]

I think #5929 is definitely something BLOCKING

#5171 is something good to have as the first version supporting Spark 3.0 with a GPU capability

[trivialfis]

• New callback functions ([POC] Implement callback for dask. #5612)
• Expose rabit context for dask interface. (Host IP resolution problem w/ dask on kubernetes or dask-gateway #5765, [dask] Accept user defined tracker address. #5408)

Removed from the roadmap.

[hcho3]

Added blocking PRs

[wbo4958]

I can add gpu_hist unit test today, @CodingCat @hcho3

[hcho3]

@wbo4958 I marked gpu_hist as "blocking". We'll make sure it's included in 1.2.0.

[hcho3]

@thvasilo Categorical data support won't be included in 1.2.0. cc @trivialfis

[trivialfis]

@hcho3 Can we branch out?

[hcho3]

Sure, let me do that this Friday.

[hcho3]

Now that all blocking PRs have been merged, let me prepare a new release branch.

[hcho3]

1.2.0 is now released.