[SPARK-18389][SQL] Disallow cyclic view reference

[jiangxb1987]

What changes were proposed in this pull request?

Disallow cyclic view references, a cyclic view reference may be created by the following queries:

CREATE VIEW testView AS SELECT id FROM tbl
CREATE VIEW testView2 AS SELECT id FROM testView
ALTER VIEW testView AS SELECT * FROM testView2

In the above example, a reference cycle (testView -> testView2 -> testView) exsits.

We disallow cyclic view references by checking that in ALTER VIEW command, when the analyzedPlan contains the same View node with the altered view, we should prevent the behavior and throw an AnalysisException.

How was this patch tested?

Test by SQLViewSuite.test("correctly handle a cyclic view reference").

[jiangxb1987]

cc @gatorsmile @hvanhovell @cloud-fan @nsyca

[SparkQA]

Test build #73854 has finished for PR 17152 at commit 1bd0260.

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

[gatorsmile]

Based on our current impl of view, createOrReplaceTempView can still trigger the cyclic view reference. Right?

[jiangxb1987]

@gatorsmile Currently we don't perform recursive resolution over a temporary view, so perhaps that won't trigger a cyclic view reference. For example:

scala> spark.sql("CREATE TEMPORARY VIEW v1 AS SELECT * FROM tab")
res3: org.apache.spark.sql.DataFrame = []

scala> spark.sql("CREATE TEMPORARY VIEW v2 AS SELECT * FROM v1")
res4: org.apache.spark.sql.DataFrame = []

scala> spark.sql("ALTER VIEW v1 AS SELECT * FROM v2")
res5: org.apache.spark.sql.DataFrame = []

scala> spark.sql("SELECT * FROM v1")
res6: org.apache.spark.sql.DataFrame = [a: int, b: string]

[gatorsmile]

Yeah. The temporary view does not have such an issue, because we did not change it. My typo.

What I mean is CREATE OR REPLACE VIEW. AlterViewAsCommand does not cover that code path. : )

[cloud-fan]

When do other databases report this error? During view creating/alter or during view resolution?

[jiangxb1987]

Hive report the error during alter view:

hive> CREATE VIEW v1 AS SELECT * FROM t1;
OK
Time taken: 0.556 seconds
hive> CREATE VIEW v2 AS SELECT * FROM v1;
OK
Time taken: 0.099 seconds
hive> ALTER VIEW v1 AS SELECT * FROM v2;
FAILED: SemanticException Recursive view default.v1 detected (cycle: default.v1 -> default.v2 -> default.v1).
hive> CREATE VIEW v3 AS SELECT * FROM v2;
OK
Time taken: 0.354 seconds
hive> ALTER VIEW v1 AS SELECT * FROM v3 JOIN v2;
FAILED: SemanticException Recursive view default.v1 detected (cycle: default.v1 -> default.v3 -> default.v2 -> default.v1).

[nsyca]

How about this test case?

sql("alter view v1 as select * from jt where exists (select 1 from v2)")

Should we get the same exception as well?

[nsyca]

What is missing in your code is whenever you hit a SubqueryExpression, you need to traverse the plan of that expression to detect cyclic references? See an example of the code in #16493.

[nsyca]

Going back to @gatorsmile 's question, does this fix cover the scenario below?

sql("create or replace view v1 as select * from v2")

If this is an existing problem and your PR does not cover it, would you intend to address it in this PR?

[jiangxb1987]

@gatorsmile @nsyca Thank you for your comments! I've added the coverage for both CREATE OR REPLACE VIEW and SubqueryExpressions.

[SparkQA]

Test build #74088 has finished for PR 17152 at commit f487af3.

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

[nsyca]

Shall we use the pattern matching instead of isInstanceOf-asInstanceOf? The logic in the code looks good to me.

[cloud-fan]

+1

[SparkQA]

Test build #74169 has finished for PR 17152 at commit 46da41e.

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

[cloud-fan]

thanks, merging to master!