Tracking deprecated features.

| Feature | PR(or time) | Replacement | Removed |
|----------------------------------------------------------|----------------|-----------------------------------------------------------------------|----------------|
| GPU Objectives. | #3643 | Now all objectives having GPU port will use GPU by default. | #4690 |
| `grow_fast_histmaker` | #3836 | Renamed to `grow_quantile_histmaker`. | ✔ |
| Some of R parameters documented in `R/xgb.importance.R`. | #1964 | ?? | ?? |
| `num_pbuffer_deprecated` | ?? | ?? | ✔ |
| `python/training.py: train::learning_rates` | #1797 | Now use use callback API. | #5155 |
| `python/sklearn.py: seed` | ?? | Now use random_state. | #4929 |
| `python/sklearn.py: nthread` | ?? | Now use n_jobs. | #4929 |
| `XGDMatrixCreateFromCSC` | #1600 | `XGDMatrixCreateFromCSCEx` | ✔ |
| `XGDMatrixCreateFromCSR` | #1600 | `XGDMatrixCreateFromCSR_omp` | ✔ |
| `silent` | #3982 | Now use `verbosity` | #5476 |
| `gpu_exact` | #4527 | Users are encouraged to use `hist` or `approx` on GPU. | #4742 |
| `n_gpus` | #4749 | Use distributed frameworks like `dask` or `spark` | #6821 |
| `reg:linear` | #4267 | Renamed to `reg:squarederror` | #4267 |
| `XGDMatrixSetGroup` | #4864 | Use `XGDMatrixSetUIntInfo` instead. | ✘ |
| `TreeParam.max_depth` | #5101 | This variable is never used (TrainParam.max_depth is still supported) | #5101 |
| Old Python callback functions | #6199 | A new callback interface is designed. | #7280 |
| Label encoder in `XGBClassifier` (`use_label_encoder`) | #6269 | User should perform label encoding manually. | #7357 |
| `ntree_limit` in Python | #6668 | Use `iteration_range` or model slicing instead. | #8345 (2.0.0) |
| positional arguments in Python package | #6365 | Use keyword arguments instead. | ✘ |
| `use_gpu` in PySpark | #9390 (2.0.0) | Use `device` instead. | #11554 (3.1.0) |
| `gpu_hist` tree method | #9385 (2.0.0) | Use `device` instead. | #11549 (3.1.0) |
| `gpu_coord_descent` updater | #9507 (2.0.0) | Use `device` instead. | #11395 (3.1.0) |
| `predictor` (`cpu_predictor`, `gpu_predictor`) | #9129 (2.0.0) | Use `device` instead. | #9129 (2.0.0) |
| `gpu_id` | #9385 (2.0.0) | Use `device` instead. | #11543 (3.1.0) |
| `fit` parameters in the sklearn interface | #6751 (1.6.0) | Use `set_params` instead. | #9986 (2.1.0) |
| Command line interface | #9485 (2.1.0) | Use other language bindings instead | #11720 (3.2.0) |
| Reading files from URLs like S3 | #9504 (2.1.0) | Use third-party libraries instead. | #9504 (2.1.0) |
| MPI support in RABIT | #9525 (2.1.0) | Use RABIT or NCCL instead (along with grpc for federated) | #9525 (2.1.0) |
| `XGDMatrixSetDenseInfo` | #10139 (2.1.0) | Use `XGDMatrixSetInfoFromInterface` | ✘ |
| `manylinux2014` Python wheels | #10475 (2.1.0) | Use a recent Linux distribution and `manylinux2_28` wheels instead | #11673 (3.1.0) |
| `feature_weights` in the sklearn `fit` method | #9506 (3.0.0) | Use `set_params` instead. | #12479 (3.5.0) |
| Dask `scheduler_address` | #10983 (3.0.0) | Use the `Config` instead. | #12480 (3.5.0) |
| Text file inputs. | #11590 (3.1.0) | Use third-party data structures like numpy. | ✘ |
| `gblinear` | #12030 (3.3.0) | XGBoost will focus on tree model. | ✘ |
| RMM plugin | #12415 (3.5.0) | Use the CUDA async pool instead. | ✘ |
| `binary:logitraw` | #12477 (3.5.0) | Use `binary:logistic` with `output_margin=true` | ✘ |