Status

Current state: ["Under Discussion"]

Discussion thread: here (<- link to https://lists.apache.org/thread/mc0lv4gptm7som02hpob1hdp3hb1ps1v)

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

Released: 1.16

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

Motivation

The current syntax/features of Flink SQL is very perfect in both stream mode and batch mode. But there are still some usability to improve. For example, If the user wants to insert data into a new table, 2 steps are required:

• First, prepare the DDL statement of the table named t1;
• Second, insert the data into t1;

These two steps seem to be normal, but if there are many fields, spelling DDL statements can be difficult, and write out these columns in the following insert statement. Therefore, we can support CTAS (CREATE TABLE AS SELECT) like MySQL, Oracle, Microsoft SQL Server, Hive, Spark, etc ... It will be more user friendly. In addition, the Hive dialect already has some support for CTAS. My suggestion would be to support a variation of an optional Feature T172, “AS subquery clause in table definition”, of SQL standard.

Public API Changes

Through the appendix research summary and analysis, the current status of CREAE TABLE AS SELECT(CTAS) in the field of big data is:

• Flink: Flink dialect does not support CTAS. ==> LEVEL-1
• Flink: Hive dialect already supports CTAS but does not guarantee atomic(can not roll back). ==> LEVEL-2
• Spark DataSource v1: is atomic (can roll back), but is not isolated. ==> LEVEL-3
• Spark DataSource v2: Guaranteed atomicity and isolation. ==> LEVEL-4
• Hive MR: Guaranteed atomicity and isolation. ==> LEVEL-4

Hive SQL and Spark SQL are mainly used in offline(batch mode) scenarios; Flink SQL is suitable for both real-time(streaming mode) and offline(batch mode) scenarios. In a real-time scenario, we believe that the job is always running and does not stop, and the data is written in real time and visible in real time, so it is no need to provide atomicity.

To ensure that Flink SQL is semantically consistent in Streaming mode and Batch mode, combining the current situation of Flink and the needs of our business, choosing LEVEL-2 atomicity as the default behavior for Flink streaming and batch mode. If the user requires LEVEL-3 atomicity,  this ability can be achieved by enabling an atomicity option. In general, batch mode usually requires LEVEL-3 atomicity. In a nutshell, Flink provide two level atomicity guarantee, LEVEL-2 as the default behavior.

Syntax

We proposing the CREATE TABLE AS SELECT(CTAS) clause as following:

CREATE TABLE [ IF NOT EXISTS ] table_name 
[ WITH ( table_properties ) ]
[ AS query_expression ]

The optional AS query_expression is we new propose change which is the SELECT query expression.

Assuming the following CTAS query:

CREATE TABLE ctas_hudi
 WITH ('connector.type' = 'hudi')
 AS SELECT id, name, age FROM hive_catalog.default.test WHERE mod(id, 10) = 0;

Resulting table equivalent to:

CREATE TABLE ctas_hudi
 (
 	id BIGINT,
 	name STRING,
 	age INT
 )
 WITH ('connector.type' = 'hudi');

INSERT INTO ctas_hudi SELECT id, name, age FROM hive_catalog.default.test WHERE mod(id, 10) = 0;

Public Interface

Table

Providing method that are used to execute CTAS(CREATE TABLE AS SELECT) for Table API user.

@PublicEvolving
public interface Table extends Explainable<Table>, Executable {

    /**
     * Declare the pipeline defined by the given {@link Table} object
     * to create the table at the specified path.
     */
     CreateOrRepalceTable saveAs(String tablePath);
}

CreateOrReplaceTable

Proposing a public interface CreateOrReplaceTable used by CTAS(CREATE TABLE AS SELECT) for table API user.

/**
 * The table is used for the {@code Create Table As Select} syntax in Table API level. This interface 
 * is created by {@code Table#save} method, and provides some methods such as {@code option} and {@code create} 
 * method help to fill the table options and produce {@link TablePipeline}. Currently we only support 
 * {@code Create Table As Select} syntax, but the {@code Replace Table AS SELECT} and {@code Create Or Replace Table AS SELECT} 
 * syntax may be supported in the future.
 */
@PublicEvolving
public interface CreateOrReplaceTable {

    /**
     * Adding table options to CreateOrReplaceTable.
     *
     * <p>Example:
     *
     * <pre>{@code
     * tab.option("connector", "filesystem");
     * }</pre>
     */
    CreateOrReplaceTable option(String key, String value);

    /**
     * Create the table in the specified path and write the pipeline data to this table.
     *
     * <p>Example:
     *
     * <pre>{@code
     * table.saveAs("my_ctas_table")
     *         .option("connector", "filesystem")
     *         .option("format", "testcsv")
     *         .option("path", "/tmp/my_ctas_table/")
     *         .create();
     * }</pre>
     */
    TablePipeline create();

    /**
     * Create the table under the specified path if not exist and write the pipeline data to this table.
     *
     * <p>Example:
     *
     * <pre>{@code
     * table.saveAs("my_ctas_table")
     *         .option("connector", "filesystem")
     *         .option("format", "testcsv")
     *         .option("path", "/tmp/my_ctas_table/")
     *         .createIfNotExist();
     * }</pre>
     */
    TablePipeline createIfNotExist();
}

The CreateOrReplaceTable interface is introduced newly because if we add the create/createIfNotExist API in the Table interface, the user must call the saveAs API before calling these API, which will cause additional usage costs to the user. This API only support Create Table As Select syntax currently, but in the future, we maybe support Replace Table As Select and  Create Or Replace As Table syntax which is also supported by some other batch compute engine.

The recommended way to use CreateOrReplaceTable as following:

TablePipeline tablePipeline = table.saveAs("my_ctas_table")
        .option("connector", "filesystem")
        .option("format", "testcsv")
        .option("path", "/tmp/my_ctas_table/")
        .create();
tablePipeline.execute();

We save the properties set through the option API and set them in the CatalogBaseTable when executing the create/createIfNotExist API, so as to generate the TableSink.

Catalog

Regarding Catalog interface, to support Create Table As Select syntax(atomicity guarantee, if don't need atomicity, the second option is no need), two changes are needed here:

1. Providing a new method inferTableOptions that is used to infer the options of CatalogBaseTable, these options will be used to compile the sql to JobGraph successfully. This method throw UnsupportedOperationException default.
2. This interface should extands the java Serializable interface, then it can be serialized as a part of JobGraph and pass to JM side, so this require the essential options to construct a catalog object can be serialized.

/**
 * The CREATE TABLE AS SELECT(CTAS) syntax requires CREATE/DROP TABLE operations via Catalog on the JM side,
 * so this interface should extends Serializable interface, then it can be serialized as a part of {@link JobGraph}.
 */
@PublicEvolving
public interface Catalog extends Serializable {

    /**
     * This method is used to infer the default options for {@link CatalogBaseTable} through {@link Catalog} options to compile 
     * the sql successfully by planner when using the {@code Create Table As Select} syntax. 
     *
     * Assuming an user want to select data from a kafka table and then insert the result to mysql table, if the mysql table is not existed in 
     * in physical mysql storage, user also doesn't want to create the table manually in mysql side because of complex type mapping.
     * User can create the {@link JdbcCatalog} firstly which connect to the msyql instance, then use 
     * {@code CREATE TABLE `mysql`.`user_db`.`order_cnt` AS SELECT * FROM `KafkaTable`} syntax, it is convenient to load data from kafka
     * to msyql. Due to the {@link JdbcCatalog} has provides user, password, url and other options, so user doesn't need to fill the 
     * option in the query. If user doesn't specify the target table required options, planner will call this method to fill the options to
     * {@link CatalogBaseTable} which are need to compile sql by planner.
     *
     * <p>{@link JdbcCatalog} example:
     * <pre>{@code
     *     // If the user does not set any Table's options,
     *     // then the implementation of JdbcCatalog#inferTableOptions
     *     // can be like this to avoid the execution failure.
     *     public CatalogBaseTable inferTableOptions(ObjectPath tablePath, CatalogTable table) {
     *         Map<String, String> tableOptions = table.getOptions();
     *         tableOptions.put("connector", "jdbc");
     *         tableOptions.put("url", getBaseUrl());
     *         tableOptions.put("table-name", tablePath.getObjectName());
     *         tableOptions.put("username", getUsername());
     *         tableOptions.put("password", getPassword());
     *         return table.copy(tableOptions);
     *     }
     * }</pre>
     */
    default CatalogBaseTable inferTableOptions(ObjectPath tablePath, CatalogTable table) {
        throw new UnsupportedOperationException();
    }

}

TableConfigOptions

Add configuration options to allow users to enable atomicity.

Implementation Plan

We provide two semantics for Flink CTAS: Non-atomic and Atomic. Non-atomic implementations are the default behavior of Streaming and Batch modes. 

Non-atomic (default)

The overall execution process is shown in the following figure.

The non-atomic implementation is basically the same as the existing Insert data process, except that the sink table is first created on the Client side via Catalog before performing the insert.

Compile the SQL, parse the schema of the sink table based on the query, then create the table, and finally submit the job to write data to the sink table. No need for too much introduction.

Atomic

The atomicity implementation of Flink CTAS requires two parts:

• Enabling the atomicity option.
• Catalog can be serialized, ensuring atomicity by performing created/dropped table on the JM side.

 The following describes the process when the user enables the atomicity support option. The overall execution process is shown in the following figure.

Due to the client process may exit soon, such as detached mode, and the job execution maybe take a longer time, so the table creation(before job start) and drop(when job failed or cancelled) should be executed in JM side instead of client. In addition, the creation and drop of table is executed through Catalog which is consistent with existing behavior. Therefore, a new Hook mechanism should be introduced which is needed by JM to execute the corresponding action,  the hook depend on Catalog to complete the function. In summary, the overall execution process of CTAS job is as following:

1. Flink Client compiles SQL and generates an execution plan, In this process, the Hook that needs to be executed on the JM side is generated, the Catalog and CatalogBaseTable are the member variable of Hook which should be serializable .
2. Submit the job to the cluster through the client.
3. When the job starts, construct the hook object,  deserialize the Catalog and CatalogBaseTable in hook. Then call the Catalog#createTable method through the hook to create the CatalogBaseTable.
4. Task start to execute.
5. If the final status of the job is failed or canceled, the created CatalogBaseTable needs to be dropped by the hook to call the Catalog#dropTable method.

The next describes the details of the implementation.

Planner

Providing method for planner to register JobStatusHook with StreamGraph.

public class StreamGraph implements Pipeline {

    

    private final List<JobStatusHook> jobStatusHooks = new ArrayList<>();

    ... ...

    /** Registers the JobStatusHook. */
    void addJobStatusHook(JobStatusHook hook) {
        ...
    }
}

The final tasks of the job are all generated by Planner. We want to complete the create/drop table action through Hook on the JM side, so we need an API to register the Hook on the JM side.

Introducing the process of CTAS in Planner:

step1:

Compile SQL to generate CatalogBaseTable (The sink table to be created) and CreateTableASOperation.

step2:

Use Catalog#inferTableOptions method to filling the required options of CatalogBaseTable. The specific implementation is determined by the Catalog.

For example, when using JdbcCatalog, if the user does not fill any table options, JdbcCatalog will set connector to 'jdbc' and fill username, password and base-url.  When using HiveCatalog, if the user does not fill in any table options, HiveCatalog will set connector to 'hive'; User-implemented catalogs can also use this mechanism to fill some required options.

It should be noted that the InMemoryCatalog case, the tables saved in it all exist in the external system, so the table options have to be filled in manually by the user, the Catalog cannot infer it automatically. If the Catalog does not support ManagedTable and the user does not set the connector information, the execution will fail.

step3:

Using CatalogBaseTable and Catalog objects to construct JobStatusHook. Due to  the JobStatusHook is finally executed on the JM side, and the CatalogBaseTable needs to be created/dropped through the Catalog in hook, so Catalog and CatalogBaseTable are member variables of hook, which also need to be serialized and can be passed to JM.

step4:

Planner registers JobStatusHook with StreamGraph, then the JobStatusHook is also serialized and passed to JM through the serialization of JobGraph. 

For CatalogBaseTable, we use CatalogPropertiesUtil to serialize/deserialize it , it's the tools that Flink already provides.

For Catalog, we need Catalog to extends Serializable so that it can be serialized directly. 

The purpose of doing so is:

1. By this way, the Catalog registered via DDL and Table API both can be serialized.
2. Reduce the cost of user-defined catalogs without much consideration for serialization (If the CREATE TABLE AS SELECT (CTAS) syntax is supported, the catalog must be serializable).

Key Points for Catalog Support Serializability:

Built-in Catalog:

• InMemoryCatalog: Due to the CatalogDatabase and CatalogBaseTable etc can't be serialized by java serialization mechanism directly, so the InMemoryCatalog doesn't support to serialize.
• JdbcCatalog: The required member variables to construct Catalog object are directly serializable, such as username, password, base url, etc. The JdbcDialectTypeMapper interface need extends the serializable.
• HiveCatalog: All member variables can be serialized directly, except for the HiveConf object, which cannot be serialized directly. We can refer to JobConfWrapper to solve the serialization problem of HiveConf.

User-defined Catalog:

• User-defined catalog that require support for CREATE TABLE AS SELECT (CTAS) syntax should ensure that the catalog can be serialized.

Create Table As Select(CTAS) feature depends on the serializability of the catalog. To quickly see if the catalog supports CTAS, we need to try to serialize the catalog when compile SQL in planner and if it fails, an exception will be thrown to indicate to user that the catalog does not support CTAS because it cannot be serialized.

Note: The metadata in InMemoryCatalog is only stored in Flink's memory and does not modify metadata in external systems, so InMemoryCatalog does not need to support atomicity. That is, InMemoryCatalog does not need to be adapted to support serialization.

Runtime

Provide JM side, job status change hook mechanism.

/**
 * Hooks provided by users on job status changing.
 */
@Internal
public interface JobStatusHook extends Serializable {

    /** When Job become CREATED status. It would only be called one time. */
    void onCreated(JobID jobId);
 
    /** When job finished successfully. */
    void onFinished(JobID jobId);

    /** When job failed finally. */
    void onFailed(JobID jobId, Throwable throwable);

    /** When job get canceled by users. */
    void onCanceled(JobID jobId);
}

Flink's current Hook design cannot meet the needs of CTAS. For example, the JobListener is executed on the client side; JobStatusListener is on the JM side, but it cannot be serialized and it function cannot meets CTAS. Thus we tend to propose a new interface which is called JobStatusHook, that could be attached to the JobGraph and executed in the JobMaster. The interface will also be marked as Internal. 

The process of CTAS in runtime

1. When the job starting, the JobGraph will be deserialized, and then the JobStatusHook can also be deserialized.
2. When deserializing JobStatusHook, Catalog and CatalogBaseTable will also be deserialized.
   • Deserialize CatalogBaseTable using CatalogPropertiesUtil#deserializeCatalogTable method.
   • After deserializing the Catalog, call Catalog#open to ensure that the Catalog can be used.
3. When the job is start and the job status changes, the JobStatusHook method will be called by JM:

For example, our JobStatusHook implementation is called CTASJobStatusHook, and use JdbcCatalog, JdbcCatalog is serialized by Planner has been covered in the previous section and will not be repeated.

Then when the job status changes, the CTASJobStatusHook method can be called:

• When the job status is CREATED, the runtime module will call the CTASJobStatusHook#onCreated method, which will call the JdbcCatalog#createTable method to create a table.
• When the final status of the job is FAILED, the runtime module will call the CTASJobStatusHook#onFailed method, which will call the JdbcCatalog#dropTable method to drop table.
• When the final status of the job is CANCELED, the runtime module will call the CTASJobStatusHook#onCanceled method, which will call the JdbcCatalog#dropTable method to drop table.
• When the final status of the job is FINISH, the runtime module will call the CTASJobStatusHook#onFinished method, and we do not need to do any additional operations.

Data Visibility

Regarding data visibility, it is determined by the TableSink and runtime-mode:

Stream mode:

If the external storage system supports transactions or two-phase commit, then data visibility is bound to the Checkpoint cycle. Otherwise, data is visible immediately after writing, which is consistent with the current flink behavior.

Batch mode:

• FileSystem Sink: Data should be written to the temporary directory first, visible after the final job is successful(final visibility).
• Two-phase commit Sink:  Data visible after the final job is successful(final visibility).
• Supports transaction Sink:  Commit transactions after the final job is successful(final visibility). Commit transactions periodically or with a fixed number of records(incremental visibility).
• Other Sink:  Data is visible immediately after writing(write-visible).

Managed Table

For Managed Table, please refer to FLIP-188 . Table options that do not contain the ‘connector’ key and value represent a managed table. CTAS also follows this principle. For details, please refer to the Table Store docs: https://nightlies.apache.org/flink/flink-table-store-docs-master/docs/development/create-table.

CTAS supports Managed Table and Non-Managed Table, user need to be clear about their business needs and set the table options correctly. The Catalog#inferTableOptions API can also automatically infer whether to add the connector attribute based on whether the Catalog supports ManagedTable.

Compatibility, Deprecation, and Migration Plan

It is a new feature with no implication for backwards compatibility.

Test Plan

changes will be verified by UT

Rejected Alternatives

Catalog serialize

Option 1: Add serialize/deserialize API to catalog

If we added serialize and deserialize APIs, Catalog must implement serialization and deserialization itself. We save the class name of Catalog together with the serialized content, like this:

Since the Catalog class may not have a parameterless constructor, so we can't use Class#newInstance to initialize an object, we can use the framework objenesis to solve. After using objenesis to get the Catalog object (an empty Catalog instance), get the real Catalog instance through the Catalog#deserialize API. This solves the serialization/deserialization problem of Catalog.

For example, JdbcCatalog#serialize can save catalogName, defaultDatabase, username, pwd, baseUrl, and JdbcCatalog#deserialize can re-initialize a JdbcCatalog object through these parameters; HiveCatalog#serialize can save catalogName, defaultDatabase, hiveConf, hiveVersion, and HiveCatalog#deserialize can re-initialize a HiveCatalog object through these parameters; InMemoryCatalog#serialize only needs to save the catalogName and defaultDatabase, and InMemoryCatalog#deserialize can re-initialize an InMemoryCatalog object through these two parameters.

The tables in the InMemoryCatalog already exist in the external system. The metadata information held in the InMemoryCatalog is only used by the job itself, and is held only in memory. Therefore, all metadata information in the InMemoryCatalog does not need to be serialized and passed to JM. In JM, only need to initialize a new InMemoryCatalog.

The solution serialization tool is more complex to implement, and the user-defined Catalog is more expensive to implement, so it is abandoned.

Option 2: Serialize the options in the Create Catalog DDL

We need to serialize catalog name and the options which are used in catalog DDL, then JM side can use these options to re-initialize the catalog by flink ServiceLoader mechanism(Using FactoryUtil#createCatalog to get catalog).  To InMemoryCatalog, here are some special case. Due to the tables in InMemoryCatalog already exist in the external system, metadata information in InMemoryCatalog is only used by the job itself and is only stored in memory. The database related information in InMemoryCatalog needs to be serialized and then passed to JM, otherwise the database may not exist when JM creates the table. Other objects do not need to be serialized.

Here we give an example about catalog serializable process that catalog is created by DDL way.

CREATE CATALOG my_catalog WITH(
    'type' = 'jdbc',
    'default-database' = '...',
    'username' = '...',
    'password' = '...',
    'base-url' = '...'
);

1) The Planner registers the catalog to the CatalogManager, it also registers the properties in the with keyword to the CatalogManager.

2) When serializing the catalog, only need to serialize and save the catalog name(my_catalog) and properties, like this:

The advantages of this solution are simple design, ease of compatibility and reduced complexity of implementation for the user, and does not require complex serialization and deserialization tools.

The disadvantage of this solution is that it does not cover the TableEnvironment#registerCatalog usage scenario, and database related information is missing in the InMemoryCatalog scenario.

Note:  This solution only works if we create the Catalog using DDL, because we can only get the Catalog properties with the with keyword. If we use a Catalog registered by TableEnvironment#registerCatalog method, we cannot get these properties. Therefore, CTAS does not currently support jobs that use TableEnvironment#registerCatalog.

The following issues require attention:

1. HiveCatalog:  
   1. If hive-conf-dir is specified, since the configuration of hive-conf-dir is a local path, please make sure that all nodes in the cluster put the hive configuration file under the same path, otherwise JM will not find the file and FAILED. This problem also exists in the current application mode of Flink.
   2. If hive-conf-dir is not specified, then HiveCatalog will look for hive-site.xml from Java Classpath, then we have to solve the hive-site.xml upload problem and make sure that all modes in Flink Client and JM Classpath can find Otherwise the job will fail.

References

1. Support SELECT clause in CREATE TABLE(CTAS)
2. MySQL CTAS syntax
3. Microsoft Azure Synapse CTAS
4. LanguageManual DDL#Create/Drop/ReloadFunction
5. Spark Create Table Syntax

Appendix

Program research

I have investigated other big data engine implementations such as hive, spark:

Hive(MR) ：atomic

Hive MR is client mode, the client is responsible for parsing, compiling, optimizing, executing, and finally cleaning up.

Hive executes the CTAS command as follows:

1. Execute query first, and write the query result to the temporary directory.
2. If all MR tasks are executed successfully, then create a table and load the data.
3. If the execution fails, the table will not be created.

Spark(DataSource v1) : non-atomic

There is a role called driver in Spark, the driver is responsible for compiling tasks, applying for resources, scheduling task execution, tracking task operation, etc.

Spark executes CTAS steps as follows:

1. Create a sink table based on the schema of the query result.
2. Execute the spark task and write the result to a temporary directory.
3. If all Spark tasks are executed successfully, use the Hive API to load data into the sink table created in the first step.
4. If the execution fails, driver will drop the sink table created in the first step.

Spark(DataSource v2, Not yet completed, Hive Catalog is not supported yet) : optional atomic

Non-atomic

 Non-atomic implementation is consistent with DataSource v1 logic. For details, see CreateTableAsSelectExec .

Atomic

Atomic implementation( for details, see AtomicCreateTableAsSelectExec), supported by StagingTableCatalog and StagedTable .

StagedTable supports commit and abort. 

StagingTableCatalog is in memory, when executes CTAS steps as follows:

1. Create a StagedTable based on the schema of the query result, but it is not visible in the catalog.
2. Execute the spark task and write the result into StagedTable.
3. If all Spark tasks are executed successfully, call StagedTable#commitStagedChanges(), then it is visible in the catalog.
4. If the execution fails, call StagedTable#abortStagedChanges().

Research summary

We want to unify the semantics and implementation of Streaming and Batch, we finally decided to use the implementation of Spark DataSource v1.

Reasons:

• Streaming mode requires the table to be created first(metadata sharing), downstream jobs can consume in real time.
• In most cases, Streaming jobs do not need to be cleaned up even if the job fails(Such as Redis, cannot be cleaned unless all keys written are recorded).
• Batch jobs try to ensure final atomicity(The job is successful and the data is visible; otherwise, drop the metadata and delete the temporary data).