This page is meant as a template for writing a FLIP. To create a FLIP choose Tools->Copy on this page and modify with your content and replace the heading with the next FLIP number and a description of your issue. Replace anything in italics with your own description.
Status
Current state: Accepted
Discussion thread: https://lists.apache.org/thread/n6nsvbwhs5kwlj5kjgv24by2tk5mh9xd
VOTE thread: https://lists.apache.org/thread/8c0wlp72kq0dhcbpy08nl1kb28q17kv3
JIRA:
-
FLINK-32580Getting issue details...
STATUS
Released: 1.18
Please keep the discussion on the mailing list rather than commenting on the wiki (wiki discussions get unwieldy fast).
Motivation
CREATE TABLE AS SELECT(CTAS) statement has been support by FLIP-218, but it's not atomic. It will create the table first before job running. If the job execution fails, or is cancelled, the table will not be dropped.
We want Flink to support atomic CTAS, where only the table is created when the Job succeeds.
we refer to FLIP-218: Support SELECT clause in CREATE TABLE(CTAS) , use the existing JobStatusHook mechanism and extend Catalog's new API to implement atomic CTAS capabilities.
Public Interfaces
Introduce interface SupportsStaging , which provided applyStaging API. If DynamicTableSink implements the interface SupportsStaging, it indicates that it supports atomic operations.
/** * Enables different staged operations to ensure atomicity in a {@link DynamicTableSink}. * * <p>By default, if this interface is not implemented, indicating that atomic operations are not * supported, then a non-atomic implementation is used. */ @PublicEvolving public interface SupportsStaging { /** * Provides a {@link StagedTable} that provided transaction abstraction. StagedTable will be * combined with {@link JobStatusHook} to achieve atomicity support in the Flink framework. Call * the relevant API of StagedTable when the Job state is switched. * * <p>This method will be called at the compile stage. * * @param StagingContext Tell DynamicTableSink, the operation type of this StagedTable, * expandable * @return {@link StagedTable} that can be serialized and provides atomic operations */ StagedTable applyStaging(StagingContext context); /** * The context is intended to tell DynamicTableSink the type of this operation. In this way, * DynamicTableSink can return the corresponding implementation of StagedTable according to the * specific operation. More types of operations can be extended in the future. */ interface StagingContext { StagingPurpose getStagingPurpose(); } enum StagingPurpose { CREATE_TABLE_AS, CREATE_TABLE_AS_IF_NOT_EXISTS } }
Introduce StagedTable interface that support atomic operations.
/** * A {@link StagedTable} for atomic semantics using a two-phase commit protocol, combined with * {@link JobStatusHook} for atomic CTAS. {@link StagedTable} will be a member variable of * CtasJobStatusHook and can be serialized; * * <p>CtasJobStatusHook#onCreated will call the begin method of StagedTable; * CtasJobStatusHook#onFinished will call the commit method of StagedTable; * CtasJobStatusHook#onFailed and CtasJobStatusHook#onCanceled will call the abort method of StagedTable; */ @PublicEvolving public interface StagedTable extends Serializable { /** * This method will be called when the job is started. Similar to what it means to open a * transaction in a relational database; In Flink's atomic CTAS scenario, it is used to do some * initialization work; For example, initializing the client of the underlying service, the tmp * path of the underlying storage, or even call the start transaction API of the underlying * service, etc. */ void begin(); /** * This method will be called when the job is succeeds. Similar to what it means to commit the * transaction in a relational database; In Flink's atomic CTAS scenario, it is used to do some * data visibility related work; For example, moving the underlying data to the target * directory, writing buffer data to the underlying storage service, or even call the commit * transaction API of the underlying service, etc. */ void commit(); /** * This method will be called when the job is failed or canceled. Similar to what it means to * rollback the transaction in a relational database; In Flink's atomic CTAS scenario, it is * used to do some data cleaning; For example, delete the data in tmp directory, delete the * temporary data in the underlying storage service, or even call the rollback transaction API * of the underlying service, etc. */ void abort(); }
TableConfigOptions
Add table.ctas.atomicity-enabled option to allow users to enable atomicity when using create table as select syntax.
@PublicEvolving public class TableConfigOptions { @Documentation.TableOption(execMode = Documentation.ExecMode.BATCH_STREAMING) public static final ConfigOption<Boolean> TABLE_CTAS_ATOMICITY_ENABLED = key("table.ctas.atomicity-enabled") .booleanType() .defaultValue(false) .withDescription( "Specifies if the create table as select operation is executed atomically. " + "By default, the operation is non-atomic. The target table is created in Client side, and it will not be dropped even though the job fails or is cancelled. " + "If set this option to true and DynamicTableSink implements the SupportsStaging interface, the create table as select operation is expected to be executed atomically, " + "the behavior of which depends on the actual DynamicTableSink."); }
Proposed Changes
First we need to have a Table interface that can be combined with the abstract transaction capability, so we introduce StagedTable, which can perform start transaction, commit transaction, and abort transaction operations.
The three APIs corresponding to StagedTable:
begin : Similar to open transactions, we can do some prep work, such as initializing the client, initializing the data, initializing the directory, etc.
commit : Similar to commit transactions, we can do some data writing, data visibility, table creation, etc.
abort : Similar to abort transactions, we can do some data cleaning, data restoration, etc.
Note: StagedTable must be serializable, because it used on JM.
Then we need somewhere to create the StagedTable, because different TableSink implement atomic CTAS and need to perform different operations,
for example, HiveTableSink needs to access the Hive Metastore and write to HDFS(OSS etc); JDBCTableSink needs to access the back-end database;
Therefore, we introduce the interface SupportsStaging, which, if implemented by DynamicTableSink, indicates that it supports atomic operations, otherwise it does not support atomic operations.
Flink framework can determine whether DynamicTableSink supports atomicity CTAS by whether it implements the interface SupportsStaging, and if it does, get the StagedTable object through the applyStaging API, otherwise use the non-atomic CTAS implementation.
Identification of atomic CTAS
Normally, in stream mode, we consider the job to be LONG RUNNING, and even if it fails, it needs to resume afterwards, so atomic CTAS semantics are usually not needed.
In addition, there are probably many flink jobs that already use non-atomic CTAS functions, especially Stream jobs, in order to ensure the consistency of flink behavior, and to give the user maximum flexibility, in time DynamicTableSink implements the SupportsStaging interface, users can still choose non-atomic implementation according to business needs.
So, we can infer in the TableEnvironmentImpl whether atomic CTAS is used based on whether the user has enabled it and whether DynamicTableSink implements the SupportsStaging interface, like the following:
boolean isAtomicCtas = tableConfig.get(TableConfigOptions.TABLE_CTAS_ATOMICITY_ENABLED) && dynamicTableSink instanceof SupportsStaging;
Integrate atomicity CTAS
Introduce CtasJobStatusHook (implements JobStatusHook interface), StagedTable is its member variable;
The implementation of the API related to the call to StagedTable is as follows:
/** * This Hook is used to implement the Flink CTAS atomicity semantics, calling the corresponding API * of {@link StagedTable} at different stages of the job. */ public class CtasJobStatusHook implements JobStatusHook { private final StagedTable stagedTable; public CtasJobStatusHook(StagedTable stagedTable) { this.stagedTable = stagedTable; } @Override public void onCreated(JobID jobId) { stagedTable.begin(); } @Override public void onFinished(JobID jobId) { stagedTable.commit(); } @Override public void onFailed(JobID jobId, Throwable throwable) { stagedTable.abort(); } @Override public void onCanceled(JobID jobId) { stagedTable.abort(); } }
Compatibility with existing non-atomic CTAS
We can infer atomicity CTAS support by whether DynamicTableSink implements the interface SupportsStaging or not:
not : it means that atomicity semantics are not supported and the existing code logic is used;
yes : it means that atomicity semantics are supported, then create a CtasJobStatusHook and use the JobStatusHook mechanism to implement atomicity semantics, as described in the code in the previous section.
Optional<DynamicTableSink> dynamicTableSinkOptional = getDynamicTableSink( catalogTable, tableIdentifier, createTableOperation.isTemporary(), catalogManager.getCatalog(catalogName)); if (tableConfig.get(TableConfigOptions.TABLE_CTAS_ATOMICITY_ENABLED) && dynamicTableSinkOptional.isPresent() && dynamicTableSinkOptional.get() instanceof SupportsStaging) { DynamicTableSink dynamicTableSink = dynamicTableSinkOptional.get(); StagedTable stagedTable = ((SupportsStaging) dynamicTableSink) .applyStaging( new SupportsStaging.StagingContext() { @Override public SupportsStaging.StagingPurpose getStagingPurpose() { if (createTableOperation .isIgnoreIfExists()) { return SupportsStaging.StagingPurpose .CREATE_TABLE_AS_IF_NOT_EXISTS; } return SupportsStaging.StagingPurpose .CREATE_TABLE_AS; } }); CtasJobStatusHook ctasJobStatusHook = new CtasJobStatusHook(stagedTable); mapOperations.add( ctasOperation.toStagedSinkModifyOperation( createTableOperation.getTableIdentifier(), catalogTable, ctasCatalog, dynamicTableSink)); jobStatusHookList.add(ctasJobStatusHook); } else { // execute CREATE TABLE first for non-atomic CTAS statements executeInternal(ctasOperation.getCreateTableOperation()); mapOperations.add(ctasOperation.toSinkModifyOperation(catalogManager)); }
To avoid secondary generation of DynamicTableSink, we need to construct a StagedSinkModifyOperation that inherits from SinkModifyOperation and then add the DynamicTableSink member variable.
Current non-atomic CTAS implementations
Current Flink supports non-atomic CTAS operations, when it is CreateTableASOperation, we will create the target table first, and then compile and execute the insert operation.
The current program has the following shortcomings:
First: If the insert operation fails, whether it is a compile failure or a job execution failure, flink will not drop the created target table;
Second: Before the job is executed, because the target table already exists, but no data can be read.
Atomic CTAS demo
Then implementation of the atomic CTAS operation requires only two steps :
1: DynamicTableSink implements the interface SupportsStaging;
2: Introduce the implementation class of the StagedTable interface.
Hive demo
HiveTableSink implements the applyStaging API:
@Override public StagedTable applyStaging(StagingContext context) { Table hiveTable = HiveTableUtil.instantiateHiveTable( identifier.toObjectPath(), catalogTable, HiveConfUtils.create(jobConf), false); hiveStagedTable = new HiveStagedTable( hiveVersion, new JobConfWrapper(jobConf), hiveTable, context.getStagingPurpose() == SupportsStaging.StagingPurpose.CREATE_TABLE_AS_IF_NOT_EXISTS); return hiveStagedTable; }
HiveStagedTable implements the core logic
/** An implementation of {@link StagedTable} for Hive to support atomic ctas. */ public class HiveStagedTable implements StagedTable { private static final long serialVersionUID = 1L; @Nullable private final String hiveVersion; private final JobConfWrapper jobConfWrapper; private final Table table; private final boolean ignoreIfExists; private transient HiveMetastoreClientWrapper client; private FileSystemFactory fsFactory; private TableMetaStoreFactory msFactory; private boolean overwrite; private Path tmpPath; private String[] partitionColumns; private boolean dynamicGrouped; private LinkedHashMap<String, String> staticPartitions; private ObjectIdentifier identifier; private PartitionCommitPolicyFactory partitionCommitPolicyFactory; public HiveStagedTable( String hiveVersion, JobConfWrapper jobConfWrapper, Table table, boolean ignoreIfExists) { this.hiveVersion = hiveVersion; this.jobConfWrapper = jobConfWrapper; this.table = table; this.ignoreIfExists = ignoreIfExists; } @Override public void begin() { // init hive metastore client client = HiveMetastoreClientFactory.create( HiveConfUtils.create(jobConfWrapper.conf()), hiveVersion); } @Override public void commit() { try { // create table first client.createTable(table); try { List<PartitionCommitPolicy> policies = Collections.emptyList(); if (partitionCommitPolicyFactory != null) { policies = partitionCommitPolicyFactory.createPolicyChain( Thread.currentThread().getContextClassLoader(), () -> { try { return fsFactory.create(tmpPath.toUri()); } catch (IOException e) { throw new RuntimeException(e); } }); } FileSystemCommitter committer = new FileSystemCommitter( fsFactory, msFactory, overwrite, tmpPath, partitionColumns.length, false, identifier, staticPartitions, policies); committer.commitPartitions(); } catch (Exception e) { throw new TableException("Exception in two phase commit", e); } finally { try { fsFactory.create(tmpPath.toUri()).delete(tmpPath, true); } catch (IOException ignore) { } } } catch (AlreadyExistsException alreadyExistsException) { if (!ignoreIfExists) { throw new FlinkHiveException(alreadyExistsException); } } catch (Exception e) { throw new FlinkHiveException(e); } finally { client.close(); } } @Override public void abort() { // do nothing } public void setFsFactory(FileSystemFactory fsFactory) { this.fsFactory = fsFactory; } public void setMsFactory(TableMetaStoreFactory msFactory) { this.msFactory = msFactory; } public void setOverwrite(boolean overwrite) { this.overwrite = overwrite; } public void setTmpPath(Path tmpPath) { this.tmpPath = tmpPath; } public void setPartitionColumns(String[] partitionColumns) { this.partitionColumns = partitionColumns; } public void setDynamicGrouped(boolean dynamicGrouped) { this.dynamicGrouped = dynamicGrouped; } public void setStaticPartitions(LinkedHashMap<String, String> staticPartitions) { this.staticPartitions = staticPartitions; } public void setIdentifier(ObjectIdentifier identifier) { this.identifier = identifier; } public void setPartitionCommitPolicyFactory( PartitionCommitPolicyFactory partitionCommitPolicyFactory) { this.partitionCommitPolicyFactory = partitionCommitPolicyFactory; } public Table getTable() { return table; } }
Jdbc Demo
JdbcTableSink implements the applyStaging API:
@Override public StagedTable applyStaging(StagingContext context) { ... ... StagedTable stagedTable = new JdbcStagedTable( new ObjectPath(tablePath.getDatabaseName(), tablePath.getObjectName() + "_" + System.currentTimeMillis()), tablePath, tableSchem, jdbcUrl, jdbcUserName, jdbcPassword); return stagedTable; }
JdbcStagedTable implements the core logic
/** An implementation of {@link StagedTable} for Jdbc to support atomic ctas. */ public class JdbcStagedTable implements StagedTable { private final ObjectPath tmpTablePath; private final ObjectPath finalTablePath; private final Map<String, String> schema; private final String jdbcUrl; private final String userName; private final String password; public JdbcStagedTable( ObjectPath tmpTablePath, ObjectPath finalTablePath, Map<String, String> schema, String jdbcUrl, String userName, String password) { this.tmpTablePath = tmpTablePath; this.finalTablePath = finalTablePath; this.schema = schema; this.jdbcUrl = jdbcUrl; this.userName = userName; this.password = password; } @Override public void begin() { // create tmp table, writing data to the tmp table Connection connection = getConnection(); connection .prepareStatement("create table " + tmpTablePath.getFullName() + "( ... ... )") .execute(); } @Override public void commit() { // Rename the tmp table to the final table name Connection connection = getConnection(); connection .prepareStatement( "rename table " + tmpTablePath.getFullName() + " to " + finalTablePath.getFullName()) .execute(); } @Override public void abort() { // drop tmp table Connection connection = getConnection(); connection.prepareStatement("drop table " + tmpTablePath.getFullName()).execute(); } private Connection getConnection() { // get jdbc connection return JDBCDriver.getConnection(); } }
Compatibility, Deprecation, and Migration Plan
It is a new feature with no implication for backwards compatibility.
Test Plan
changes will be verified by UT