Unify logging facilities.

[trivialfis]

• Enhance ConsoleLogger to handle different verbosity.
• Override macros from dmlc.
• Refactor cli_main to share common variables.

This is an attempt to unify some parameters, specifically debug_verbose and silent picked up from various places. If the method is approved, we can move on to some more dangerous and messy parameters like n_gpus and gpu_id.

[trivialfis]

One motivation is to honor the parameter silent and debug_verbose automatically, preventing un-guarded logging.

[codecov-io]

Codecov Report

Merging #3982 into master will increase coverage by 0.2%.
The diff coverage is 71.07%.

@@             Coverage Diff             @@
##             master    #3982     +/-   ##
===========================================
+ Coverage     56.35%   56.55%   +0.2%     
  Complexity      205      205             
===========================================
  Files           186      187      +1     
  Lines         14771    14824     +53     
  Branches        498      498             
===========================================
+ Hits           8324     8384     +60     
+ Misses         6208     6201      -7     
  Partials        239      239

Impacted Files	Coverage Δ	Complexity Δ
src/tree/param.h	86.34% <ø> (-0.38%)	0 <0> (ø)
src/common/math.h	100% <ø> (ø)	0 <0> (ø)	⬇️
src/data/data.cc	75.13% <ø> (ø)	0 <0> (ø)	⬇️
src/gbm/gblinear.cc	12.34% <0%> (-2.12%)	0 <0> (ø)
src/tree/updater_quantile_hist.cc	33.82% <0%> (+0.08%)	0 <0> (ø)	⬇️
src/gbm/gbtree.cc	17.05% <0%> (-1.62%)	0 <0> (ø)
src/learner.cc	26.43% <100%> (-0.58%)	0 <0> (ø)
tests/cpp/test_logging.cc	100% <100%> (ø)	0 <0> (?)
src/linear/updater_coordinate.cc	100% <100%> (ø)	0 <0> (ø)	⬇️
include/xgboost/logging.h	85% <100%> (+10%)	0 <0> (ø)	⬇️
... and 5 more

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 93f9ce9...1f491eb. Read the comment docs.

[RAMitchell]

Are the changes in CLI main necessary? It will be very difficult to review and prove its correctness as I don't think we even have tests for this.

Please check the monitor class log statement is the correct type.

I think there is a change in behaviour here, before if I set debug_verbose = 1 I would get the timing information for gpu_hist and hist. Now I would have to set this at 2 to get the same information, is this correct?

[trivialfis]

@RAMitchell Thanks for the review. The debug_verbose actually is an un-documented internal parameter, somehow people started using it, which reminds me that I should update the doc.

Yes, separated test for cli should be added.

[trivialfis]

@RAMitchell

Are the changes in CLI main necessary? It will be very difficult to review and prove its correctness as I don't think we even have tests for this.

I stripped the changes in CLI into minimum. Tests for CLI will be postponed for another PR.

[hcho3]

Thank you for improving the logging system. I like how you can set the warning level uniformly across all the program. Good job.

[hcho3]

This is perfectly fine for now, but in the long term, we should port this improvement to DMLC-Core, so that we can get fine-grained warning level for all projects using DMLC-Core.

[trivialfis]

I have some fears of changing dmlc-core. I looked at tvm and treelite a few times, but that's all. So when the port is needed, please let me know, I will need some help. :)

[hcho3]

It would be useful to begin discussion for better logging system in TVM and MXNet. Already, there is a talk of better error handling: apache/tvm#2279.

[trivialfis]

@hetong007 @khotilov Could you please help taking a look and deprecating silent in R package?

[trivialfis]

@CodingCat I'm deprecating the silent parameter, currently still working, but could you help looking at the JVM package? I will take care of python.