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


The whole conception and architecture of SQL Client are proposed in FLIP-24 which mainly focuses on embedded mode. The goal of this FLIP is to extend FLIP-24 to support gateway mode and expose the Gateway with pluggable endpoints. The reason why we introduce the gateway with pluggable endpoints is that many users has their preferences. For example, the HiveServer2 users prefer to use the gateway with HiveServer2-style API, which has numerous tools. However, some filnk-native users may prefer to use the REST API. Therefore, we hope to learn from the Kyuubi's design that expose multiple endpoints with different API that allow the user to use.

The goal of the FLIP:

  • Introduce the Gateway with REST endpints
  • Design the pluggable endpoint API.
  • Allows the SQL Client to submit the statement to the SQL Gateway.

Core Concepts

Like many big data platforms, Flink SQL Gateway also has the following concepts.


Session represents the users who visit the Gateway in the peiord. Flink SQL Gateway uses the SessionHandle as the index to identify the Session. In addition to uniquely identifying the user being accessed, it also acts as an isolation of resources, including jar resources, configuration information and meta information.


Every user request is transformed to Operation.

  • User uses the OperationHandle to identify the user request.
  • User can send the request to the Gateway to manage the Operation:
    • Get OperationStatus that describes whether the Operation is running or meets error.
    • Cancel the running Operation:
    • Close the finished Operation;


SessionManager is responsible to manage the Session, including:

  • Every user should register a Session before sending its request to the Gateway;
  • Every user should close its corresponding Session when it exists. If a Session is inactive for a period of time, the SessionManager will clean up the Session automatically;
  • SessionManager should limit the number of the active Session to avoid OOM;


OperationManager is responsible to manage the Operation . The Operation's initialization, execution and clean up should be controlled by the OperationManager. When the Session closes, the OperationManager also needs to clean up all alive Operation.

We organize all the concepts in the following graph.


The architecture of the Gateway is in the following graph.

Flink SQL Gateway is composed of the Endpoints, GatewayService and MetricSystem.

  • GatewayService: It's responsible to manage the active Sessions and submitted Operations. When user tries to get the results, the GatewayService should return the corresponding results.
    • GatewayService exposes API about the management of the Session,Operation
    • GatewayService executes the Operation async: when submit the operation to the OperationManager, return the OperationHandle. It doesn't wait the Operation execution finishes ;
    • GatewayService is shared between all the loaded endpoints. 
  • Endpoint: It's the main entry point for the users to visit.
    • Considering the different needs of the different users, we perfer the pluggable endpoint API and user can extends with different mode, e.g. HivServer2, PrestoCoordinator. Referring to the implementation of connector in Flink, we prefer to use a loading mechanism similar to SPI to dynamically load the corresponding Endpoint according to the user's configuration items.

    • In many cases, user Client's version is different from the Endpoint's version. Therefore, we need to confirm the version of the communication during the open session. It determines the interface how the returned results are serialized in the server side. Considering that different endpoints have their own version management, we propose to let each Endpoint manage its own Endpoint version. Every time the Endpoint needs to return the result to the Client, it serialize the results according to the version determined during the session registeration.

    • Different Endpoints expose different execution modes, e.g. HiveServer2 supports asynchronous/synchronous job submission (DML, DQL), while origin REST API currently only supports asynchronous job submission. In order to simplify the corresponding execution mode, we only support asynchronous submission in GatewayService, and Endpoint controls the corresponding execution mode. In the synchronous execution mode, the Endpoint monitors the status of the Operation, and returns the corresponding result to the user after the job is completed.

    • Operation state machine:
      • Considering all Operation has its status, we propose the state machine like HiveServer2:
        • INITIALIZED: Operation is created ;
        • PENDING: Status during the status switch;
        • RUNNING: Operation starts running;
        • FINISHED: Operation finishes;
        • CANCELED: User cancels the Operation;
        • CLOSED: User closes the Operation;
        • ERROR: Operation execution meet errors;
        • TIMEDOUT:Execution timeout

The trigger events for the state machine.

Statement Type





Start to prepare the Operation resourc, e.g. LOG. 

Get the resources to execution(The worker thread starts running)

TableEnvironment#executeSql finishes

Get any exception.

User requests to cancel/close the operation.The operation's execution timeout.

Utils statement including 

  • SHOW
  • USE
  • ...
QUERYCan not fetch any results from the CollectSink.
SET/RESETset/reset session config

Setting/resetting session config finsihes

We can summarize the state machine as follows

  • MetricSystem: it's responsible to report the metric to the specified destination.
    • Gateway is the main entry for the user to submit SQL jobs. In many cases, some metrics are needed to measure the state of the entire system, so as to locate some problems. Or use indicators to facilitate the management of the gateway by the peripheral system, e.g. load balancing.

Public Interfaces

REST Endpoint

In this section, we propose to introduce the REST Endpoint, which is different from the REST endpoint we proposed before.

Session-related API

Using the API in the section, users can register a Session in the Gateway, which maintains the user-level configuration and resources. 



Verb: POST

Response code: 200 OK

Create a new session with the specific configuraion. In the release-1.16, we only supports to load the jar in the local file system. But we can extend this feature to load the jar in the remote filesystem, e.g. HDFS.

Request body


"session_name": "", # optional

"libs": [], # optional. 

"jars": [], # optional

"properties": { # optional, properties for current session

  "key": "value"



Response body


"session_handle": "", # if session is created successfully





Response code: 200 OK

Close a session, release related resources including operations and properties

Request body


Response body


"status": "CLOSED" # if cancel successfully


Get Session Config


Verb: GET

Response code: 200 OK

Get the session config with the specified session handle.

Request body


Response body


"properties": {

  "key": "value"



Configure Session


Verb: POST

Response code: 200 OK

Configures the session with the statement which could be SET/RESET/CREATE/ DROP/LOAD/UNLOAD/USE/ALTER/ADD JAR/REMOVE JAR. It can be used to initialize the Session:

  • register Catalog 
  • register user defined function
  • create predefined table

Note: The statement must be a single command, otherwise the server will throw an exception.

Request body


"statement": "", # required

"execution_timeout": "" # execution time limit in milliseconds, optional


Response body


Trigger Session Heartbeat


Verb: POST

Response code: 200 OK

Trigger heartbeat to tell the server that the client is active, and to keep the session alive as long as configured timeout value.

If a session does not receive a heartbeat or any other operations, the session will be destroyed when the timeout is reached.

Request body


Response body


Because the Gateway allows async execution mode, the API in the section can manipulate the runnning jobs.

Get Operation Status


Verb: GET

Response code: 200 OK

Get the status of a running job.

If the session is expired, the server will throw "session not found" exception.

If the job is finished, the server will throw "job not found" exception.

Request body


Response body


"status": "" # refer to OperationStatus


Cancel Operation


Verb: PUT

Response code: 200 OK

Cancel the running operation and update the opeartion status.

Request body


Response body


"status": "CANCELED" # if cancel successfully


Close Operation



Response code: 200 OK

Remove the specified Operation.

If the user invokes closeOperation twice, the later invocation will get exception.

Request body


Response body


"status": "CLOSED" # if close successfully


The API in the section is used to submit the statement to the Gateway and get the results.

Execute a statement


Verb: POST

Response code: 200 OK

Execute a statement which could be all Flink supports SQL statement.

The SET xx=yy statement will override/update the TableConfig held by current session, and the RESET statement will reset all properties set by SET xx=yy statement.

The USE MODULE/CATALOG/DATABASE xx statement will update the default module/catalog/database in TableEnvironment held by current session.

The statement must be a single command, otherwise the server will throw an exception.

For BEGIN STATEMENT SET, it will open a buffer in the Session and allows the users to submit the insert statement into the Session later. When the Session receives the END statement, the Gateway will submit the buffered statements.

For ADD JAR/REMOVE JAR, if the jar is in the local environment, we will just add it into the class path or remove it from the class path. If the jar is the remote jar, we will create a session level directory and download the jar into the directory. When the session closes, it should also clean up all the resources in the session-level directory.

Request body


"statement": "", # required

"execution_timeout": "" # execution time limit in milliseconds, optional


Response body


"operation_handle": "",

"operation_type": "EXECUTE_STATEMNT",

"has_result": true/false # determine whether needs to fetch results later


Fetch results


Verb: GET

Response code: 200 OK

Fetch a part of result for a flink job execution. If the result data is too large or the result is streaming, we can use this API to get a part of the result at a time. The initialized value of token is 0. The token in the next request must be the same as the token in the current request or must be equal to token (in the current request) + 1, otherwise the client will get an exception from the server. If multiple requests are executed with the same token, the result is the same. This design makes sure the client could get the result even if some errors occur in client. The client can get the next part of result using /v1/sessions/:session_id/jobs/:job_id/result/:{token+1} (which is the value of next_result_uri in the response data). If next_result_uri is empty, no more data is remaining in the server.

The server could drop the old data before current token. (The client successfully obtains those data)

We will introduce fetch_size or max_wait_time (to reach the fetch_size) for optimization in future.

The returned result has the same schema as the TableResult#getResolvedSchema.

Please refer to the Appendix about the transormation between the ResultSet and JSON.

Request body


Response body


"result_type": "PAYLOAD",

"results": [ # currently, there is only one result now. If multiple queries is executed in a single job, there are many results.


"columns": [ # if the execution is successful


"name": "",

"type": {"type":  "BOOLEAN", "nullable": true}# string value of LogicalType



"data": [

["value", ], # a row data




"next_result_uri": /v1/sessions/:session_id/jobs/:job_id/result/:{token+1} # if not empty, uses this uri to fetch next part of result, else there is no more result.

"exception": {

      "root_cause": "....",

     "exception_stack": "..." 



Statement Completement


Verb: GET

Response code: 200 OK

Complete the statements. For example, users input SELE in the terminal and press the tab, the terminal will use the API to complete the statement and return the SELECT.

Request body


"statement": "",



Response body


"candidates": []



Get Info


Verb: GET

Response code: 200 OK

Get meta data for this cluster

Request body


Response body


"product_name": "Apache Flink",

"version": "1.16" # Flink version


Get Version


Verb: GET

Response code: 200 OK

Get the current avaliable versions for the Rest Endpoint. The client can choose one of the return version as the protocol for later communicate.

Request body


Response body


"versions": ["v1", "v2"] # The rest endpoint support version.



Please using the following options to configure the REST endpint.

Option name

Default Value(Required)



rest (Yes)

REST endpoint should use 'rest'.


REST endpoint port. (No)

The address that the SqlServer binds itself.

Pluggable Endpoint Discovery

The pluggable endpoint discovery will load the endpoint dynamically. It enables the users to load the endpoint they want to use. Here We use the SPI mechanism to discover the Endpoint.

/** Interface for Endpoint. */
public interface SqlGatewayEndpoint {

    void start() throws Exception;
    void stop() throws Exception;

 * A factory for creating Endpoint from Configuration. This
 * factory is used with Java's Service Provider Interfaces (SPI) for discovery.
public interface SqlGatewayEndpointFactory extends Factory {

    SqlGatewayEndpoint createSqlGatewayEndpoint(Context context);

    interface Context {
        SqlGatewayService getSqlGatewayService();

        MetricGroup getMetricGroup();

        /** Gives read-only access to the configuration of the Endpoint. */
        ReadableConfig getConfiguration();

We also expose the option sql-gateway.endpoint.type to allow user to specify the endpoints. Considering that the different endpoints may have the same settings, e.g. port, users should add the endpoint identifier as the prefix to specify the option, For simplicity, we don't plan to introduce another yaml for SQL Gateway and users can specify the gateway options in the flink-conf.yaml.

For example, users can add the following options in the flink-conf.yaml.

sql-gateway.endpoint.type: rest, hiveserver2 localhost 9001
sql-gateway.endpoint.hiveserver2.address: localhost
sql-gateway.endpoint.hiveserver2.port: 9002

GatewayService API

Object API

The API is used to describe the Session and Operation.

public class HandleIdentifier {
    UUID publicId;
    UUID secretId;

public class SessionHandle {
    HandleIdentifier identifier;

 * Every Endpoint should expose its version and extend the interface.
interface EndpointVersion {}

enum RestEndpointVersion implements EndpointVersion {

 * It's equal to the HiveServer2 TProtocolVersion. It should belong to the
 * hive module.
enum HiveServer2EndpointVersion implements EndpointVersion {

  // V2 adds support for asynchronous execution

  // V3 add varchar type, primitive type qualifiers

  // V4 add decimal precision/scale, char type

  // V5 adds error details when GetOperationStatus returns in error state

  // V6 uses binary type for binary payload (was string) and uses columnar result set

  // V7 adds support for delegation token based connection

  // V8 adds support for interval types

  // V9 adds support for serializing ResultSets in SerDe

  // V10 adds support for in place updates via GetOperationStatus

  // V11 adds timestamp with local time zone type


enum OperationType { 

public class OperationHandle {
    HandleIdentifier identifier;

SqlGatewayService API

interface SqlGatewayService {
    // -------------------------------------------------------------------------------------------
    // Session Management
    // -------------------------------------------------------------------------------------------

    SessionHandle openSession(SessionEnvironment environment) throws SqlGatewayException;
    void closeSession(SessionHandle sessionHandle) throws SqlGatewayException;
    Map<String, String> getSessionConfig(SessionHandle sessionHandle) throws SqlGatewayException;
    // -------------------------------------------------------------------------------------------
    // Operation Management
    // -------------------------------------------------------------------------------------------
     * Get operation info to describe the Operation.
    OperationInfo getOperationInfo(SessionHandle sessionHandle, OperationHandle operationHandle) throws SqlGatewayException;
    /** Get the result schema for the specified Operation. */
    ResolvedSchema getOperationResultSchema(SessionHandle sessionHandle, OperationHandle opreationHandle) throws SqlGatewayException;
    void cancelOperation(SessionHandle sessionHandle, OperationHandle operationHandle) throws SqlGatewayException;
    void closeOperation(SessionHandle sessionHandle, OperationHandle operationHandle) throws SqlGatewayException;
    // -------------------------------------------------------------------------------------------
    // Statements 
    // -------------------------------------------------------------------------------------------
     * Using the statement to initialize the Session. It's only allowed to 
    void configureSession(SessionHandle sessionHandle, String statement, long executionTimeoutMs) throws SqlGatewayException;
    /** Execute the statement with the specified Session. It allows to execute with Operation-level configuration.*/
    OperationHandle executeStatement(
        SessionHandle sessionHandle, 
        String statement, 
        long executionTimeoutMs, 
        Configuration executionConfig) throws SqlGatewayException;
	 * Fetch the results with token id. 
    ResultSet fetchResults(SessionHandle sessionHandle, OperationHandle operationHandle, int token, int maxRows) throws SqlGatewayException;

     * Fetch the Operation-level log from the GatewayService. For some endpoint, it allows to fetch the log at the operation level.
    ResultSet fetchLog(SessionHandle sessionHandle, OperationHandle operationHandle, FetchOrientation orientation, int maxRows) throws SqlGatewayException;
    * Only supports to fetch results in FORWARD/BACKWARD orientation.
    * - Users can only BACKWARD from the current offset once.
    * - The Gateway doesn't materialize the changelog.
    ResultSet fetchResult(SessionHandle sessionHandle, OperationHandle operationHandle, FetchOrientation orientation, int maxRows) throws SqlGatewayException;          

     * For the same functionality, every endpoint has its result schema. Therefore, 
     * the endpoint submit the callable executor to the OperationManager that manages
     * lifecycle of the Operaiton. The callable executor organizes the results
     * as the Endpoint requires.
    OperationHandle submitOperation(OperationType type, Callable<ResultSet> executor, ResolvedSchema resultSchema) throws SqlGatewayException;
    // -------------------------------------------------------------------------------------------
    // Utils 
    // -------------------------------------------------------------------------------------------
     * Describe the cluster info.
    Map<String, String> getGatewayInfo();
    void heartbeat(SessionHandle sessionHandle) throws SqlGatewayException;
     * Endpoint is status-less. All the session configs are memorized in the GatewayService side.
	EndpointVersion getSessionEndpointVersion(SessionHandle sessionHandle) throws SqlGatewayException;
    /** Returns a list of completion hints for the given statement at the given position. */
    List<String> completeStatement(String sessionId, String statement, int position) throws SqlGatewayException;

    // -------------------------------------------------------------------------------------------
    // Catalog API 
    // -------------------------------------------------------------------------------------------
    String getCurrentCatalog(SessionHandle sessionHandle) throws SqlGatewayException;

    String getCurrentDatabase(SessionHandle sessionHandle) throws SqlGatewayException;
    List<String> listCatalogs(SessionHandle sessionHandle) throws SqlGatewayException;
    List<String> listDatabases(SessionHandle sessionHandle, String catalogName) throws SqlGatewayException;   
    List<TableInfo> listTables(SessionHandle sessionHandle, String catalogName, String databaseName, TableKind tableKind) throws SqlGatewayException;
    List<FunctionInfo> listFunctions(SessionHandle sessionHandle, String catalogName, String databaseName, FunctionScope scope) throws SqlGatewayException;
    ContextResolvedTable getTable(SessionHandle sessionHandle, ObjectIdentifier tableIdentifier) throws SqlGatewayException;
    ContextResolvedFunction getFunction(SessionHandle sessionHandle, ObjectIdentifier functionIdentifier) throws SqlGatewayException;

class TableInfo {
    boolean isTemporary;
    ObjectIdentifier identifier;
    TableKind tableKind;

class FunctionInfo {
    boolean isTemporary;
    ObjectIdentifier identifier;
    FunctionKind kind;

class SessionEnvironment {
    private String sessionName;
    private EndpointVersion sessionEndpointVersion;
    private List<URL> libs;
    private List<URL> jars;
    private Map<String, String> sessionConfig

public class OperationInfo {
    OperationStatus status;
    OperationType type;
    boolean hasResult;

public class ResultSet {
    int nextToken;
    ResultType resultType; 

    ResolvedSchema resultSchema;
    List<RowData> results;
    Exception exception;

public enum ResultType {


Please use the following API to configure the gateway.



Default value




5 min

Session will be closed when it's not accessed for this duration, which can be disabled by setting to zero or negative value.



1 min

The check interval for session timeout, which can be disabled by setting to zero or negative value.




The number of the active sessions.




Maximum number of worker threads for the gateway workers.




Minimum number of worker threads for the gateway workers.



5 min

Keepalive time for an idle worker thread. When the number of workers exceeds the min workers, excessive threads are killed after this time interval.

SQL Gateway Script

Considering many users prefer to use the SqlGateway only, we propose to add a script named the `` in the bin directory. Users can use the command

./ (start|stop|stop-all) [args]

to manipulate the sql gateway.


Start the gateway and write the pid of the startted sql gateway into the pid file. 

Users can specify the -Dkey=value to specify the parameters. For example, users can specify `` 

stop(none)Stop the last in the pid file.
stop-all(none)Stop all the server in the running pid file.

Then users can start the sql client to communicate with the SQL Gateway in the local or remote environment.

SQL Client API

When lanuch the SQL Client in the gateway mode, user should specify the address and port of the Gateway they want to commuincate. Therefore, we should expose the following startup options.




-e, --endpoint


The gateway address:port

User can start up the SQL Client with the command in the bash.

./ -e


With the SQL Gateway, users can just use the SQL to do everything. 

Configure parameters

 Users can use the SET statement to configure the parameters, including 

  • execution parameters, streaming or batch
  • optimization configuration
  • job parameters, job name 

If users want to reset the configuration, users can use the RESET statement to rollback all the settings.

Gateway also allows users to add the jar dynamically. Users can just use the ADD JAR statement to specify the jar path in the sql. 

Manage the metadata

Users can use the DDL to register the required catalog in the Gateway. With the catalog, users can CREATE/DROP/ALTER all tables in the catalog. 

Manage Jobs

Users can submit the DML in the Gateway.  In the future we may also support to use sql to manage the submitted jobs. 


Here is an example to using the SQL client to submit the statement to the SQL Gateway.

>  'type' = 'hive',
>  'hive-conf-dir' = '/opt/hive-conf'
[INFO] Execute statement succeed.

Flink SQL> SET 'execution.runtime-mode' = 'batch';
[INFO] Execute statement succeed.

Flink SQL> CREATE TABLE pageviews (
>   user_id BIGINT,
>   page_id BIGINT,
>   viewtime TIMESTAMP,
>   proctime AS PROCTIME()
> ) WITH (
>   'connector' = 'kafka',
>   'topic' = 'pageviews',
>   'properties.bootstrap.servers' = '...',
>   'format' = 'avro'
> );
[INFO] Execute statement succeed.

Flink SQL> INSERT INTO hive.default_db.kafka_table SELECT * FROM pageviews;
[INFO] Submitting SQL update statement to the cluster...
[INFO] SQL update statement has been successfully submitted to the cluster:
Job ID: 6b1af540c0c0bb3fcfcad50ac037c862


SQL Client Overview 

SQL Client has different modes. The architecture of the SQL Client in the embedded mode as follows.

In the Gateway mode, the SQL Client uses the Rest Client to communicate with the GatewayService.

Actually the architecture in the different mode are almost the same. The only difference is how to communicate with the GatewayService. Therefore, we focus on the Gateway mode in this section. The process logic in the Gateway mode should also works for the embeded mode.

The SQL Client is composed of the CliClientand Executor.

  • The CliClientis responsible to receives the statement from the terminal and print the results;
  • The Exectuor is responsible to execute the statement from the CliClient and return the results.
    • It has a client-level parser, which determines the statement whether is the client-level command, e.g. HELP, QUIT.
    • It can submit the statement to the GatewayService with the REST Client.

Gateway Implementation Details

  • Introduce the package flink-table/flink-sql-gateway-common, which includes the gateway API. When users wants to implement its own endpoint, they only needs to rely on this.
  • Introduce the package flink-table/flink-sql-gateway. It includes rest endpoint and SQL GatewayService.
  • flink-table/flink-sql-client should relies on the the package flink-table/flink-sql-gateway.

Compatibility, Deprecation, and Migration Plan

Because we introduce the Gateway modules there are no compatibility problems. For the SQL Client, we use the Gateway to submit jobs. We will keep all the functionality but change the presentation. For example,  when executing explain, the SQL Client will print the results in the now but currently we takes the results from the table now.

Future work

The Gateway itself has many functionlities. We only list part of the work here:

  1. Metric System
  2. Authentication module
  3. Multiple Endpoint, e.g. HiveServer2 endpoint
  4. Persistent Gateway
  5. ...

We will collect more feedbacks to determine which features is more important to users.

Rejected Alternatives

TableInfo and FunctionInfo VS CatalogTable and CatalogFunction

The CatalogTable and CatalogFunction are much heavier than the TableInfo and FunctionInfo. The CatalogManager requires reading from the Catalog to get the schema. But in the listTables only care about the table name, which is much lighter. Therefore, we propose to use the TableInfo with required fields. 

Support the multi-version Flink in the Gateway VS Support the multi-version in the external Service

Currently many big data tools, e.g. Zeppelin, Livy[1] support the multiple version engine. They both have the similar architecture. The Flink on Zeppelin is like[2]:

The Interpreter is a process, which means that every Flink is in the separate JVM[3].

The current Gateway is responsible for compiling the user SQL and submitting the job to the cluster, which is almost the same as the Interpreter in the graph. 

If we support the multi-version in the Gateway(GatewayService also has a version), it means all the Flink are in the same JVM. It requires the Gateway to solve the shim, classloader for different flink versions, deployments. It will mess up all the codes with refelection. It's better we can follow the design as other tools.




Merge Gateway into the Flink code base VS Support Gateway in the another repo

The main reason to move the Gateway is to support the multiple Flink versions in the Gateway. However, in the discussion above we think the Gateway is bound to the specific Flink and uses the external service to manage the Gateway instances. Considering the Gateway itself is bound to the Flink, it's better we can merge the Gateway into the Flink repo.

It also brings the following benefits:

  1. Reduce the cost, e.g. CI test, releases, maintain cost;
  2. SQL Client has the ability to submit the SQL to the SQLGateway and we can reuse most of the codes;'
  3. Gateway inside the Flink repo can ensure the highest degree of version compatibility
  4. Gateway is indispensable for a SQL engine (think of Trino/Presto, Spark, Hive). Otherwise, Flink will always be a processing system. With Gateway inside the Flink repo, Flink can provide an out-of-box experience as a SQL query engine. Users can try out the gateway for the latest version when a new version is released.

Therefore, we prefer to merge the Gateway into the Flink repo.

Result retrieval for interactive sessions VS result retrieval for fault tolerant exactly-once results

Because the current Gateway doesn't materialize anything to the stroage, we can't promise the exactly-once semantic. It means users can't retrieval the results if the Gateway has been notified the results is been taken away. 

OperationStatus VS JobStaus

The main reason we don't use the JobStatus in the state in the machine:

  • Operation includes DDL, DML and so on, which is much larger than the Job. We can't use a small concept to replace large concept.
  • Some status in the JobStats is meaningless in the Operation. For example, DDL Operation don't need RESTARTING/SUSPENDED/RECONCILING.
  • the Gateway allows to submit job(DML) in sync/async mode. The running status in the Operation Status in the different mode has different meaning:

    • In the async mode, when the gateway submits the job, the state comes to the FINISHED state
    • In the sync mode, the running status in the Operation status includes submitting the job, running job. Even if a failover occurs, we still think that this Operation is in the RUNNING state. Unless the job is unrecoverable, we change the Operation status to ERROR.

Therefore, we propose a new state machine in the Gateway side.


Serialize and deserialize the ResultSet 

The serialization of the ResultSet mainly takes into 3 parts:

  • Serialize the LogicalType
  • Seralize the RowData
  • Serialize the Exception

Serialize the LogicalType

Considering that not all LogicalType are serializable, our plan is much like how the LogicalTypeJsonSerializer does. Currently the LogicalType has 3 kinds:

  • Basic Type 

Type Name


{"type": "CHAR", "nullable": true/false, "length": <LENGTH>}
{"type": "VARCHAR", "nullable": true/false, "length": <LENGTH>}
{"type": "VARCHAR", "nullable": true/false, "length": <LENGTH>}
{"type": "BOOLEAN", "nullable": true/false} 
{"type": "BINARY", "nullable": true/false, "length": <LENGTH>}
{"type": "VARBINARY", "nullable": true/false, "length": <LENGTH>}
{"type": "VARBINARY", "nullable": true/false, "length": <LENGTH>}
{"type": "DECIMAL", "nullable": true/false, "precision": <LENGTH>, "scale": <SCALE>}
{"type": "TINYINT", "nullable": true/false}
{"type": "SMALLINT", "nullable": true/false}
{"type": "INTEGER", "nullable": true/false}
{"type": "BIGINT", "nullable": true/false}
{"type": "FLOAT", "nullable": true/false}
{"type": "DOUBLE", "nullable": true/false}
{"type": "DATE", "nullable": true/false}
{"type": "TIME", "nullable": true/false, "precision": <PRECISION>}
{"type": "TIMESTAMP", "nullable": true/false, "precision": <PRECISION>}
{"type": "TIMESTAMP_LTZ", "nullable": true/false, "precision": <PRECISION>}
{"type": "RAW", "nullable": true/false, "class": <CLASS>, "specialSerializer": <SERIALIZER>} 


{"type": "RAW", "nullable": true/false, "class": <CLASS>, "externalDataType": ...}

  • Collecction Type 

For the collection type, it recursively serialize the element type.

Type Name


{"type": "MAP", "nullable": true/false, "keyType": ..., "valueType": ...}
{"type": "ARRAY", "nullable": true/false, "elementType": ...}
{"type": "MULTISET", "nullable": true/false, "elementType": ...}
  "type": "ROW", 
  "nullable": true/false, 
  "fields": [
      "name": <FILED_NAME>,
      "fieldType": ...
  • User-defined type

Don't support serializing the user defined type. Because the user can't create a new type with SQL.

Serialize the RowData

The RowData contains the two parts: RowKind and Fields.

Therefore, we propose to serialize the RowData as follows.

  "fields": [<COLUMN_VALUE>, ...]

The `<COLUMN_VALUE>` is as same as using the `JSON_STRING` to serialize the value.

The type mapping from Flink type to JSON type is as follows.

Flink TypeJSON Type
BINARY / VARBINARYstring with encoding: base64
DATEstring with format: date
TIMEstring with format: time
TIMESTAMPstring with format: date-time
TIMESTAMP_WITH_LOCAL_TIME_ZONEstring with format: date-time (with UTC time zone)
RAWstring with encoding: base64 

Serialize the Exception

Considering some clients may also care about the root cause, the serialized json object should contain the root cause and the stack. 

"exception": {
  "root_cause": "...", 
  "exception_stack": "..."


	"result_type": "PAYLOAD",
	"results": {
		"columns": [
				"name": "id",
				"type": {"type":  "BIGINT", "nullable": false}
				"name": "name",
				"type": {"type":  "VARCHAR", "nullable": true, "length": 300}
				"name": "birthday",
				"type": {"type":  "TIMESTAMP", "nullable": true, "precision": 3}
		"data": [
				"kind": "INSERT",
				"fields": [101, "Jay", "1990-01-12T12:00.12"], # a row data
				"kind": "DELETE",
				"fields": [102, "Jimmy", null]
  "next_result_uri": /v1/sessions/:session_id/jobs/:job_id/result/:{token+1} # if not empty, uses this uri to fetch next part of result, else there is no more result.
	"exception": {
    	"root_cause": "....",
    	"exception_stack": "..." 

The design of the origin Gateway is in the