Implement key-based sampling in the planner

[kaikalur]

We have many ad hoc queries working on large datasets that keep failing/running long time/timing out. So for cases when the user just wants to get a sense of the results, we do smart sampling based on mapping hashes of keys to a percent. When the feature is enabled, we traverse the plan for the query and sample the first integer or string key found (in that order)

• both sides of an equi-join
• group by
• partition by of a window/rownum/topnrownum
• left-side (source) of semijoin

We apply the sampling predicate only once in every branch of graph so that eventually all qualifying scans will be sampled.

Test plan - added tests

Please make sure your submission complies with our Development, Formatting, and Commit Message guidelines. Don't forget to follow our attribution guidelines for any code copied from other projects.

Fill in the release notes towards the bottom of the PR description.
See Release Notes Guidelines for details.

== RELEASE NOTES ==

General Changes
* Added Smart Sampling feature that samples the tables at query (planning) time. This helps in cases when the user is just exploring the data, testing a pipeline or looking for trends. We sample all the tables (and only tables) in such a way that all "key-based" operations like JOIN/GROUP BY/window functions etc will be complete for the sampled keys. 

[kaikalur]

Fixed it to not create new plan nodes, instead use the idiomatic replaceChildren call.

[kaikalur]

Cleaned up more - renamed the function to key_sampling_percent so it's generally useful and also documented it.

[rongrong]

There's merge conflict. Please rebase.

[rongrong]

Do you also want to introduce a version of this function for varbinary?

[kaikalur]

Hmm interesting. I will open an issue for later. For now are using only varchar version.

[rongrong]

This is still not really a description of what this feature does. If you think there's no way to describe it clearly here, add proper documentation of this feature and refer it so people can learn more about it.

[kaikalur]

Yeah - where do I add it? Add a chapter to the user doc?

[kaikalur]

Actually the PR message has pretty much the documentation.

[rongrong]

Users won't read PR messages though. This is ok for now. But we probably want to document recommended usage, behavior and limitations once we polish the feature off.

[rongrong]

static import

[kaikalur]

Done. Irritating that Intellij is inconsistent with this.

[rongrong]

static import

[kaikalur]

Done

[rongrong]

nits:

keys.stream()
.filter(x -> TypeUtils.isIntegralType(x.getType().getTypeSignature(), functionAndTypeManager))
.findFirst();

[kaikalur]

Done

[rongrong]

nits: break into multiple lines.

[kaikalur]

Done

[rongrong]

This can throw if the function doesn't exist. You probably want to handle that.

[kaikalur]

Hmm - doesn't it throw PrestoException? Then we can just propagate it?

[rongrong]

The error message might not be clear to users. We'd probably throw a function not found error, but the function is not used in user query. Ideally we'd probably want to catch here and rewrite the error message to something like sampling method not found.

[kaikalur]

Done

[rongrong]

I don't think these are necessary.

[kaikalur]

It looked idiomatic - all planner classes use it. So copied lol. Removed now.

[rongrong]

Not all. Many didn't I think. 😛