Streaming Conformance too slow when just 20 rules are used

[yruslan]

Describe the bug

Streaming Conformance takes way too much time to warm up (20 minutes) and to process. a single micro-batch (8 minutes).

This is way too slow.

This happens due to the catalyst issue reported to Spark: https://issues.apache.org/jira/browse/SPARK-28090

And we have a workaround for batch: #190, #413

The job hangs completely if no workarounds are used: #1306

To Reproduce

Steps to reproduce the behavior OR commands run:

1. Create a schema with nested arrays of structs.
2. Create a dataset with 20 conformance rules or so. Some of the rules should operate inside an array.
3. Run streaming conformance.

Expected behaviour

Since all transformations are just projections the performance of a conformance job should be good.

Additional context

Ideas of a solution:

• [ ] Tweak Catalyst workaround
• [ ] Rewrite conformance interpreter so that it uses only one single traversal for all conformance rules. We need to take in to account that rules are dependent. All dependencies need to be resolved in order to be applicable in a single traversal.
• [ ] Rewrite conformance interpreter using RDDs of Row + schema. Use a mapping lambda function to do conformance transformations in an imperative way. This way we use Spark only as a computation engine. Since no Catalyst is used, the Catalyst bug won't affect the computation.

[yruslan]

Seems this is not so critical since only startup time is affected.