## What changes were proposed in this pull request?
It is reported that there is performance downgrade when applying ML pipeline for dataset with many columns but few rows.
A big part of the performance downgrade comes from some operations (e.g., `select`) on DataFrame/Dataset which re-create new DataFrame/Dataset with a new `LogicalPlan`. The cost can be ignored in the usage of SQL, normally.
However, it's not rare to chain dozens of pipeline stages in ML. When the query plan grows incrementally during running those stages, the total cost spent on re-creation of DataFrame grows too. In particular, the `Analyzer` will go through the big query plan even most part of it is analyzed.
By eliminating part of the cost, the time to run the example code locally is reduced from about 1min to about 30 secs.
In particular, the time applying the pipeline locally is mostly spent on calling transform of the 137 `Bucketizer`s. Before the change, each call of `Bucketizer`'s transform can cost about 0.4 sec. So the total time spent on all `Bucketizer`s' transform is about 50 secs. After the change, each call only costs about 0.1 sec.
<del>We also make `boundEnc` as lazy variable to reduce unnecessary running time.</del>
### Performance improvement
The codes and datasets provided by Barry Becker to re-produce this issue and benchmark can be found on the JIRA.
Before this patch: about 1 min
After this patch: about 20 secs
## How was this patch tested?
Existing tests.
Please review http://spark.apache.org/contributing.html before opening a pull request.
Author: Liang-Chi Hsieh <viirya@gmail.com>
Closes#17770 from viirya/SPARK-20392.
8ce0d8ffb6