[SPARK-13263] [SQL] SQL Generation Support for Tablesample

[gatorsmile]

In the parser, tableSample clause is part of tableSource.

tableSource
@init { gParent.pushMsg("table source", state); }
@after { gParent.popMsg(state); }
    : tabname=tableName 
    ((tableProperties) => props=tableProperties)?
    ((tableSample) => ts=tableSample)? 
    ((KW_AS) => (KW_AS alias=Identifier) 
    |
    (Identifier) => (alias=Identifier))?
    -> ^(TOK_TABREF $tabname $props? $ts? $alias?)
    ;

Two typical query samples using TABLESAMPLE are:

    "SELECT s.id FROM t0 TABLESAMPLE(10 PERCENT) s"
    "SELECT * FROM t0 TABLESAMPLE(0.1 PERCENT)"

FYI, the logical plan of a TABLESAMPLE query:

sql("SELECT * FROM t0 TABLESAMPLE(0.1 PERCENT)").explain(true)

== Analyzed Logical Plan ==
id: bigint
Project [id#16L]
+- Sample 0.0, 0.001, false, 381
   +- Subquery t0
      +- Relation[id#16L] ParquetRelation

Thanks! cc @liancheng

[SparkQA]

Test build #51028 has finished for PR 11148 at commit 58b6038.

• This patch passes all tests.
• This patch merges cleanly.
• This patch adds no public classes.

[rxin]

this seems pretty brittle. i wonder if we can change sample to make it very clear this is either a table sample, or a normal sample.

[liancheng]

Agree with @rxin. Especially, I don't think MetastoreRelation is wrapped by a Subquery in this case.

[gatorsmile]

Thank you! Will follow your suggestions!

[gatorsmile]

@rxin After searching the code base, all the samples are table sample. Could you give me more hints? I also agree it looks brittle. I did my best to make it more robust/general.

Since TABLESAMPLE is part of tableSource clause, I am facing two options:

1. Add the extra handling of TABLESAMPLE into MetastoreRelation and Subquery
2. Add a new case for TABLESAMPLE.

Let me know which option you like. The current solution is based on option 2.

[rxin]

You can for example in the Sample operator add a method that returns true if it is created from table sample in the parser.

[rxin]

Basically it'd be better to have something very explicit, rather than relying on some conditions that might change in the future.

[gatorsmile]

Unable to create a test case for this branch. I do not know if we should throw an exception here or building a SQL.

[gatorsmile]

@rxin I see. Will do it. Thanks!

[SparkQA]

Test build #51236 has finished for PR 11148 at commit 12be2c3.

• This patch passes all tests.
• This patch merges cleanly.
• This patch adds no public classes.

[SparkQA]

Test build #51243 has finished for PR 11148 at commit fc9a156.

• This patch passes all tests.
• This patch merges cleanly.
• This patch adds no public classes.

[rxin]

Looks more robust this way. I will let @liancheng do the final reviews and merge.

[gatorsmile]

Thanks for your reviews! @rxin

[SparkQA]

Test build #51326 has finished for PR 11148 at commit e4e1b72.

• This patch passes all tests.
• This patch merges cleanly.
• This patch adds no public classes.

[SparkQA]

Test build #51679 has finished for PR 11148 at commit aae08b8.

• This patch passes all tests.
• This patch merges cleanly.
• This patch adds no public classes.

[liancheng]

LGTM, thanks!

Merged to master. There was a minor conflict. Resolved it manually while merging.

[gatorsmile]

Thank you! @liancheng @rxin