Status

Current state: [Under Discussion"]

Discussion thread: here (http://apache-flink-mailing-list-archive.1008284.n3.nabble.com/DISCUSS-FLIP-84-Improve-amp-Refactor-API-of-Table-Module-td34537.html)

JIRA: Unable to render Jira issues macro, execution error.

Released: <Flink Version>

Please keep the discussion on the mailing list rather than commenting on the wiki (wiki discussions get unwieldy fast).

Motivation

In Flink 1.9, TableEnvironment introduces `void execute(String jobName)` interface to trigger the Flink table program execution, and extends `void sqlUpdate(String sql)` interface to evaluates not only a INSERT statement but also a DDL statement and a USE statement. But with more use cases coming up, there are some fatal shortcomings in current API design.

1. Inconsistent execution semantics for `sqlUpdate()`. For now, one DDL statement passed to this method will be executed immediately while one `INSERT INTO` statement actually gets executed when we call the `execute()` method, which confuses users a lot.
2. Don’t support obtaining the returned value from sql execute. The FLIP-69[1] introduces a lot of common DDLs such as `SHOW TABLES`, which require that TableEnvironment can have an interface to obtain the executed result of one DDL. SQL CLI also has a strong demand for this feature so that we can easily unify the execute way of SQL CLI and TableEnvironemt. Besides, the method name `sqlUpdate` is not consistent with doing things like `SHOW TABLES`.
3. Unclear and buggy support buffering SQLs/Tables execution[2]. Blink planner has provided the ability to optimize multiple sinks, but we don’t have a clear mechanism through TableEnvironment API to control the whole flow.
4. Unclear Flink table program trigger point. Both `TableEnvironment.execute()` and `StreamExecutionEnvironment.execute` can trigger a Flink table program execution. However if you use TableEnvironment to build a Flink table program, you must use `TableEnvironment.execute()` to trigger execution, because you can’t get the StreamExecutionEnvironment instance. If you use StreamTableEnvironment to build a Flink table program, you can use both to trigger execution. If you convert a table program to a DataStream program (using StreamExecutionEnvironment.toAppendStream/toRetractStream), you also can use both to trigger execution. So it’s hard to explain which `execute` method should be used.
5. Don’t support executing Flink table jobs in an asynchronous way. In streaming mode, one `INSERT INTO xxx` statement may never end. It’s also possible that one ETL task takes too much time to be done in a batch environment. So it’s very natural and necessary to support executing Flink table jobs in an asynchronous way. 

Let’s give an example to explain the buffering SQLs/Tables execution problem:

1. Users are confused by what kinds of sql are executed at once and what are buffered and what kinds of sql are buffered until triggered by the execute method.
2. Buffering SQLs/Tables will cause behavior undefined. We may want to insert data into the `test` table with the `/tmp1` path but get the wrong result of `/tmp2`.

Public Interfaces

1. We propose to deprecate the following methods in TableEnvironment:

• void sqlUpdate(String sql)

• void insertInto(String targetPath, Table table)

• void execute(String jobName)

• String explain(boolean extended)

• Table fromTableSource(TableSource<?> source)
  
  

2. meanwhile, we propose to introduce the following new methods in TableEnvironment:
   • ResultTable executeStatement(String statement) 
     synchronously execute the given single statement immediately, and return the execution result.
     
     public interface ResultTable {
         TableSchema getResultSchema();
         Iterable<Row> getResultRows();
     }
     
     
   • DmlBatch createDmlBatch()
     create a DmlBatch instance which can add dml statements or Tables to the batch and explain or execute them as a batch.
     
     interface DmlBatch {
         void addInsert(String insert);
         void addInsert(String targetPath, Table table);
         ResultTable execute() throws Exception ;
         String explain(boolean extended);
     }
     
     
3. For current messy Flink table program trigger point, we propose that: for TableEnvironment and StreamTableEnvironment, you must use `TableEnvironment.execute()` to trigger table program execution, once you convert the table program to a DataStream program (through `toAppendStream` or `toRetractStream` method), you must use `StreamExecutionEnvironment.execute` to trigger the DataStream program.

Proposed Changes

`void sqlUpdate(String sql)`

Now `void sqlUpdate(String sql)` method will execute DDLs right now while DMLs must be triggered by `TableEnvironment.execute()`. Both behaviors should be kept consistent. This method will buffer the `INSERT` statement which causes the above problem. So this method will be deprecated. We propose a new blocking method with execution result:

This method only supports executing a single statement which can be DDL, DML, SHOW, DESCRIBE, EXPLAIN and USE statement. This method will be executed synchronously and return a ResultTable which is the representation of the execution result, and contains the result data and the result schema. If an error occurs, this method will throw an exception.

The following table describes the result for each kind of statement:

`insertInto(String, Table)` 

Like the `INSERT` statement, the Tables passed to this method will also be buffered and will cause the buffer problem. So we advise deprecating this method.

 `Table.insertInto` will use this deprecated method now. Once this method is removed in the future, we will change the behavior of `Table.insertInto` method from lazy execution (triggered by `TableEnvironment.execute` method) to immediate execution (like `executeStatement` method). 

`execute(String jobName)` & `explain(boolean)`

Since we will disable buffering SQLs/Tables and plans, it’s meaningless to provide `execute(String jobName)` as the trigger entry point and explain(boolean) method should also not be used anymore. So we advise deprecating those two methods. Instead, we introduce a new method named `createDmlBatch` and a new class named `DmlBatch` to support multiple SQLs/Tables optimization.

From the class name of  `DmlBatch`, we know that only DML statements can be added to `DmlBatch`. In `DmlBatch`, only INSERT is supported now, DELETE and UPDATE can also be supported in the future.

 `DmlBatch` supports adding a list of SQLs and Tables through the `addXX` methods, getting the plan of all statements through the `explain` method, optimizing the whole statements and submitting the job through the `execute` method.  The added statements and Tables will be cleared when calling the `execute` method.

Each statement or Table has a return value which is the affected row count of a statement or a Table. So the ResultTable has multiple columns. All column types are BIGINT, and the column name is "affected_rowcount_" plus the index of the statement or Table. e.g. 

The schema and data in ResultTable: 

`Table fromTableSource(TableSource<?> source)`

Since Flip-64 has provided `ConnectTableDescriptor#createTemporaryTable` to register TableSource in TableEnvironment. This method should be deprecated too, it’s an omission in that flip.

How to correct the execution behavior?

First, let’s discuss the buffer problem in depth. Actually there are two levels of buffer, TableEnvironment will buffer SQLs/Tables and StreamExecutionEnvironment will buffer transformations to generate StreamGraph. Each TableEnvironment instance holds a StreamExecutionEnvironment instance. Currently, when translating a FlinkRelNode into a Flink operator, the generated transformations will be added to StreamExecutionEnvironment’s buffer. The bug[2] is caused by this behavior. Let’s give another simple example to explain the problem of StreamExecutionEnvironment’s buffer.

The job submitted by each execute method contains the topology of both queries. Users are confused about the behavior. As suggested in "Public Interfaces"，`StreamExecutionEnvironment.execute` only triggers DataStream program execution, and `TableEnvironment.execute` only triggers table program execution. So the expected behavior for the above example is `env.execute("job name")` submits the second query, and `tEnv.execute("job name") ` submits the first query. 

To meet the requirement, we will change the current behavior of TableEnvironment: TableEnvironment instance buffers the SQLs/Tables and does not add generated transformations to the StreamExecutionEnvironment instance when translating to execution plan. (The solution is similar to DummyStreamExecutionEnvironment. We can use StreamGraphGenerator to generate StreamGraph based on the transformations. This requires the StreamTableSink always returns DataStream, and the StreamTableSink#emitDataStream method should be removed since it’s deprecated in Flink 1.9) StreamExecutionEnvironment instance only buffers the transformation translated from DataStream.

Now, we introduce `DmlBatch` to require users to explicitly buffer SQLs/Tables to support multiple sinks optimization. Although the `insertInto`, `sqlUpdate` and `execute` methods are deprecated, they will not be immediately deleted, so the deprecated methods and new methods must work together in one or more versions. The TableEnvironment’s buffer will be removed once the deprecated methods are deleted. 

After we correct the behavior of the `execute` method, users can easily and correctly write the table program even if the deprecated methods, the new methods and the `to DataStream` methods are mixed used.

Examples:

We will list some examples using old API and proposed API to have a straightforward comparison in this section.

`sqlUpdate` vs `executeStatement`:

`execute & explain` & vs `createDmlBatch`:

Deprecated methods and new methods work together

TableEnvironment’s methods and DataStream work together

Summary:

Methods of TableEnvironment

New methods for single statement & multiple statements

Compatibility, Deprecation, and Migration Plan

1. Methods of TableEnvironment to be deprecated:
• void sqlUpdate(String sql)
• void insertInto(String targetPath, Table table)
• JobExecutionResult execute(String jobName)
• String explain(boolean extended)
• Table fromTableSource(TableSource tableSource)
2. You need to change to the program a little if you use `StreamExecutionEnvironment.execute` to trigger a table program execution or use `StreamTableEnvironment.execute()` to trigger a DataStream program execution.

Test Plan

The `DmlBatch#explain` method can be tested with unit tests, and other new methods can be tested with integration tests. We will also add some integration tests to verify the new methods can work with the deprecated methods correctly.

Rejected Alternatives

TableEnvironment#executeBatch(String... statement)

This method is consistent with the style of other methods in TableEnvironment, however It does not support Table API and can not explain the plan.

References

[1] FLIP-69 Flink SQL DDL Enhancement

[2] discuss planner buffer execute 

[3] FLIP-64: Support for Temporary Objects in Table module

[4] JDBC statement addBatch interface

[5] multiple statements in SQL CLI

[6] multiple statements in TableEnvironment

[7] flip-73 Introducing Executors for job submission

[8] flip-74 Flink JobClient API

[9] Sqline deal with batch execute

Appendix - Future Plan:

Notice: depends on FLIP-73/FLIP-74, not the scope of this flip.

To support execute time-cost batch sql or no-end streaming sql, it’s needed to provide an asynchronous execution way. 

Provide async execute method for executeStatement and DmlBatch.execute

Similarly as above, suggested methods:

We also should support executing batch sql asynchronously. Suggested method:

Add an async execute method to org.apache.flink.table.delegation.Executor

Since we already have flips[7][8] to provide asynchronous management, it’s convenient and natural to provide such a method.

Add an async execute method to org.apache.flink.streaming.api.environment#StreamExecutionEnvironment

SQL CLI integrates with new API

1. How SQL CLI leverage the DmlBatch class to obtain optimization?

   We can reference other system design like Sqlline Batch Command[9] and introduce similarly command but we should notice that the sql in batch can only be `insert into`.
   
   

2. How SQL CLI parse and execute multiple statements?

   We don’t want to support multiple statements parsing in the TableEnvironment but this feature is needed in the SQL CLI for it’s natural to execute an external script. I have thought provided a parse method like `List<String> parse(String stmt)`, but it’s not intuitive to understand and this method shouldn’t belong to the TableEnvironment API. As the discussion in the pull-request [5][6], calcite has provided the `SqlNodeList parseSqlStmtList()` method to parse a list of SQL statements separated by a semicolon and constructs a parse tree. I think the SQL CLI can use this method to parse multiple statements and execute every single statement one by one through TableEnvironmet#executeStatement(String statement). Here is one thing we should take care of is that there are some special commands like `help/set/quit` in SQL CLI to control the environment’s lifecycle and change the variables of the context. IMO, there are some ways to deal with these commands in the multiple statements:
   1. Support these special control commands in flink-sql-parser and the shortcoming will be that TableEnvironment should take care of those noisy commands and flink-sql-parser will lose it’s more widely expansibility to other external systems. For example, SQL CLI may need to support `source xx` that execute an external script, it’s not proper to make TableEnvironment parser to see such syntax.
      1.  pro’s: 
         • unified parser
         • can handle corner case, e.g. https://github.com/apache/flink/pull/8738
      2.  con’s: 
         • many commands are only used for sql-client, e.g. help, quit, source
         • how to meet the requirements of non-builtin commands, e.g. commands from flink-sql-gateway
         • not easy to extend, it’s more difficult to implement a client-specific command in sql-parser than in specific client 
   2. SQL CLI parses those control commands on its own and should pre-split the multiple statements according to the control command. Then SQL CLI can pass the part of multiple statements to SqlParser and obtain a SqlNodeList. 
      1. pro’s:
         • sql-parser is more clean
         • more easy to extend for sql-client
      2.  con’s: 
         • many parsers: SqlCommandParser(in sql client)，sql-parser
         • may meet the corner case, e.g.  https://github.com/apache/flink/pull/8738
   3. Flink already introduces a `Parser` interface which is exposed by `Planner`. We can add one more method to `Parser` like: List<String> splitStatement(String) and then we can borrow calcite to achieve this functionality. Special client commands (e.g. help, quit, source) are not supported in sql-parser now. Because the SqlParser#parseStmtList return SqlNodeList, not a string list, those special commands are not defined in SqlNode. So I think this approach is only a complement to the first one.
   4. Support a utility class to parse a statement separated by semicolon into multiple statements.
      1. pro’s:
         • more easy to extend for sql-client
         • can handle corner case in a unified place
      2. con’s:
         • many parsers: sql-parser,  a utility parser
           
   5. -
      
      I think option d is better. Looking forward to more people's opinions. (we can open an another flip to discuss this)
      
      
3. Other open question?