Discussion thread | https://lists.apache.org/thread/nwm5z7bkowwfd1z6gb5chzcmc2o3b2d9 |
---|---|
Vote thread | https://lists.apache.org/thread/ttj7h5nhqn90w8t8wxpln8tz83g91046 |
Issue | |
Release | - |
This PIP is co-worked by Ming Li , Guojun Li and Fang Yong .
Motivation
Paimon is a Streaming Lakehouse, Flink streaming and batch jobs can use Paimon tables as data sources and sinks. Flink can be combined with Paimon to complete the entire ETL processing, and the overall process is as follows.
...
1. Table lineage in Paimon. We want to create lineage between Paimon source and sink tables based on Flink job name, then we can manage the entire ETL topology for Paimon tables and Flink jobs.
...
2. Locating and analyzing issues. When there is a delay in the data of a Paimon table, users can get the upstream tables and jobs, check whether there are any issues in them.
3. Data management for streaming and batch ETL, for example, according to the data lineage, users can easily submit debug jobs, compare the results between the debug job and streaming running job.
4. Data correction for streaming ETL. When users need to correct data, they can rollback all tables and jobs to a unify "version" based on the data lineage, and recover the ETL after data correction.
Architecture
Overall
We would like to introduce System Metadata Database named "paimon" for each catalog in Paimon, it stores the properties of catalog and lineage information. Users can create a Paimon catalog with table and data lineage options, and if the source or sink tables of ETL job are in the catalog, the lineage information will be saved. The overall process is as followed.
...
1. Create a catalog with options table-lineage
and data-lineage
. NOTICE: these two options are immutable after a catalog is created.
Code Block |
---|
-- Create a catalog with lineage options CREATE CATALOG my_catalog WITH ( 'type' = 'paimon', 'warehouse' = 'warehouse path', 'table-lineage' = 'true', 'data-lineage' = 'true' ); USE CATALOG my_catalog; -- Select the options for the current catalog SELECT * FROM paimon.catalog_options; +-----------------------+ | name | value | +-----------------------+ | table-lineage | true | +-----------------------+ | data-lineage | true | +-----------------------+ 2 row in set |
...
draw.io Diagram | ||||||||||||||||||||
---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
|
Proposed Changes
...
Metadata Database
When users add option table-lineage
or data-lineage
to catalog DDL, Paimon will check and create system database Metadata Database 'paimon' automatically, then create options table, table lineage table and data lineage table.
...
Table souce_job_lineage stores relationship between source table and ETL job. The job, database and table fields form a joint primary key.
Column Name | Column Type | Example |
job | STRING | "myJob" |
database | STRING | "myDatabase" |
table | STRING | "myTableName" |
create_time | DATETIME | "2023-06-05 13:29:01" |
3. sink_job_lineage
Table sink_job_lineage stores relationship between sink table and ETL job. The job, database and table fields form a joint primary key.
Column Name | Column Type | Example |
job | STRING | "myJob" |
database | STRING | "myDatabase" |
table | STRING | "myTableName" |
4. source_snapshot_lineage
create_time | DATETIME | "2023-06-05 13:29:01" |
4. source_snapshot_lineage
Table source_snapshot_lineage stores the data lineage between source and ETL job, it will record which snapshot id in source table is consumed by the specific barrier id of ETL job. The job, barrier_id, database and table fields form a joint primary key.
Column Name | Column Type | Example |
job | STRING | "myJob" |
barrier_id | LONG | 1 |
database | STRING | "myDatabase" |
table | STRING | "myTableName" |
snapshot_id | LONG | 1 |
create_time | DATETIME | "2023-06-05 13:29:01" |
5. sink_snapshot_lineage
Table sink_snapshot_lineage stores the data lineage between ETL job and sink, it will record which snapshot id in sink table is produced by the specific barrier id of ETL job. The job, barrier_id, database and table fields form a joint primary key.
Column Name | Column Type | Example |
job | STRING | "myJob" |
barrier_id | LONG | 1 |
database | STRING | "myDatabase" |
table | STRING | "myTableName" |
snapshot_id | LONG | 1 |
create_time | DATETIME | "2023-06-05 13:29:01" |
Lineage Use Case
1. Query Table Lineage
...
Code Block |
---|
-- Query upstream table snapshot for given table and snapshot SELECT S.database, S.table, S.snapshot_id FROM source_snapshot_lineage S JOIN sink_snapshot_lineage T ON S.job=T.job AND S.barrier_id=T.barrier_id where T.`database`='myDatabase' and T.`table`='myTable' and T.snapshot_id=123; -- Query downstream table snapshot for given table and snapshot SELECT T.database, T.table, T.snapshot_id FROM source_job_lineage S JOIN sink_job_lineage T ON S.job=T.job AND S.barrier_id=T.barrier_id where S.`database`='myDatabase' and S.`table`='myTable' and S.snapshot_id=123; |
Table Lineage
Metadata Database Implementation
The tables in Metadata Database can be divided into two type: the catalog options table and lineage information table. For catalog options table, we can store json format data in the catalog directory, while for lineage tables the requirements are much more complex:
1. Multiple jobs will concurrently write data to the same table, for example, the source operators will write data to source_snapshot_lineage
table.
2. One table has multiple writers, but no committer operator
, which is very different from the current data writing in Paimon tables.
3. High data visibility requirements, with basic requirements for real-time visibility for each data write.
According to these requirements, we cannot use regular Paimon tables for Metadata Database. We design the metadata store api for Metadata Database and introduce jdbc as default store. Users can config their jdbc parameters for Metadata Database in create catalog
statement as follows
Code Block |
---|
CREATE CATALOG paimon_catalog1 WITH (
... // other options
'lineage-meta' = 'jdbc',
'jdbc.driver' = 'com.mysql.jdbc.Driver',
'jdbc.url' = 'XXXXX',
'jdbc.database' = 'paimon_cata1', // The default Metadata Database name is `paimon`
'jdbc.username' = 'XXX',
'jdbc.password' = 'XXX'
); |
Paimon catalog will create metadata store based on the options of metadata to store and get lineage information. The metadata store and factory classes are as follows.
Code Block |
---|
/** Factor to create metadata store. */
public interface LineageMetaFactory {
/* The identifier for metadata store factor. */
String identifier();
/* Create metadata store from specific context. */
LineageMeta create(LineageMetaContext context);
/** The context used to create metadata store in the factory. */
public interface LineageMetaContext {
Options options();
}
}
/* Metadata store will manage the options, table lineage and data lineage information for the catalog. */
public interface LineageMeta {
/**
* Store the source table and job lineage.
*
* @param entity the table lineage entity
*/
void storeSourceTableLineage(TableLineageEntity entity);
/**
* Delete the source table lineage for given job.
*
* @param job the job for table lineage
*/
void deleteSourceTableLineage(String job);
/**
* Get source table and job lineages.
*
* @param predicate the predicate for the table lineages
* @return the iterator for source table and job lineages
*/
Iterator<TableLineageEntity> sourceTableLineages(@Nullable Predicate predicate);
/**
* Store the sink table and job lineage.
*
* @param entity the table lineage entity
*/
void storeSinkTableLineage(TableLineageEntity entity);
/**
* Get sink table and job lineages.
*
* @param predicate the predicate for the table lineages
* @return the iterator for source table and job lineages
*/
Iterator<TableLineageEntity> sinkTableLineages(@Nullable Predicate predicate);
/**
* Delete the sink table lineage for given job.
*
* @param job the job for table lineage
*/
void deleteSinkTableLineage(String job);
/**
* Store the source table and job lineage.
*
* @param entity the data lineage entity
*/
void storeSourceDataLineage(DataLineageEntity entity);
/**
* Get source data and job lineages.
*
* @param predicate the predicate for the table lineages
* @return the iterator for source table and job lineages
*/
Iterator<DataLineageEntity> sourceDataLineages(@Nullable Predicate predicate);
/**
* Store the source table and job lineage.
*
* @param entity the data lineage entity
*/
void storeSinkDataLineage(DataLineageEntity entity);
/**
* Get sink data and job lineages.
*
* @param predicate the predicate for the table lineages
* @return the iterator for source table and job lineages
*/
Iterator<DataLineageEntity> sinkDataLineages(@Nullable Predicate predicate);
}
/**
* Table lineage entity with database, table and job for table source and sink lineage.
*/
public interface TableLineageEntity {
public String getDatabase();
public String getTable();
public String getJob();
public String getCreateTime();
}
/**
* Data lineage entity with table lineage, barrier id and snapshot id for table source and sink lineage.
*/
public interface DataLineageEntity extends TableLineageEntity {
public long getBarrierId();
public long getSnapshotId();
} |
We default to implementing the jdbc metadata store, which supports the use of external databases to store this information.
Code Block |
---|
public class JdbcLineageMetaOptions {
/**
* The jdbc driver class for lineage meta.
*/
public static ConfigOption<String> JDBC_DRIVER =
key("jdbc.driver")
.stringType()
.noDefaultValue()
.withDescription("The jdbc driver class for lineage meta.");
/**
* The jdbc url for lineage meta.
*/
public static ConfigOption<String> JDBC_URL =
key("jdbc.url")
.stringType()
.noDefaultValue()
.withDescription("The jdbc url for lineage meta.");
/**
* The jdbc url for lineage meta.
*/
public static ConfigOption<String> JDBC_USERNAME =
key("jdbc.username")
.stringType()
.noDefaultValue()
.withDescription("The jdbc username for lineage meta.");
/**
* The jdbc url for lineage meta.
*/
public static ConfigOption<String> JDBC_PASSWORD =
key("jdbc.password")
.stringType()
.noDefaultValue()
.withDescription("The jdbc password for lineage meta.");
/**
* The jdbc url for lineage meta.
*/
public static ConfigOption<String> JDBC_DATABASE =
key("jdbc.database")
.stringType()
.defaultValue("paimon")
.withDescription("The jdbc database for lineage meta and default database is `paimon`");
}
/* Jdbc metadata store factory. */
public class JdbcLineageMetaFactory implements LineageMetaFactory {
public static final String IDENTIFIER = "jdbc";
@Override
public String identifier() {
return IDENTIFIER;
}
@Override
public LineageMeta create(LineageMetaContext context) {
// Create DataSource from the context
return new LineageMeta(`The DataSource of database`);
}
}
/* Jdbc lineage meta. */
public class JdbcLineageMeta implements LineageMeta {
private DataSource dataSource;
} |
Besides jdbc, users can implements their own metadata store and factory to manage the lineage information in their own external system.
Metadata Database Query And Action
We can implement ReadOnlyTable
for tables in Metadata Database, then users can query data from these table as regular Paimon table in computation engine such as Flink, Spark and Trino.
Code Block |
---|
/** Catalog options table for Paimon. */
public class CatalogOptionsTable implements ReadOnlyTable {
}
/** Source/sink table and job lineage table for Paimon. */
public class TableLineageTable implements ReadOnlyTable {
private final MetadataStore metadataStore;
}
/** Source/sink data and job lineage table for Paimon. */
public class DataLineageTable implements ReadOnlyTable {
private final MetadataStore metadataStore;
} |
Tables in Metadata Database are read-only and users can query data from metadata tables by SQL just like a regular table. But users can not alter the metadata tables and update the data in them. We provide Flink actions for users to delete data from metadata tables as follows
action | argument | note |
---|---|---|
delete-table-lineage | --job <job-name>: specify name of the job. | delete table lineage created by the given job. |
delete-data-lineage | --job <job-name>: specify name of job. | delete data lineage created by the given job. |
Table Lineage
We introduce a new option table-lineage
We introduce a new option table-lineage
for Paimon catalog, users can set this option when they create a new catalog.
...
However, managing Paimon table lineage in this way will have one issue: the lifecycle management of table lineage. When ETL job goes to termination, the table lineage should be deleted manually. Currently we would like to support customized job status listener in FLIP-314 [1], we can solve this issue based on FLIP-314 in the future.
Data Lineage
Job Submission
...
Currently Flink does not support setting a unified version for all sources in the job, we have created an issue [2]. According to the exist ability of Flink, we introduce job_startup table which saves the table snapshot id for job when it starts.
Column Name | Column Type | Example |
job | STRING | "myJob" |
database | STRING | "myDatabase" |
table | STRING | "myTableName" |
snapshot_id | LONG | 1 |
When submitting the job, based on the existing table and snapshot id, we will query the new source table's corresponding snapshot id from the data lineage table. When creating the source for job, we will put the relevant snapshot id in the configuration and write the corresponding data to the job_startup table. During the execution of streaming jobs, the source will directly consume data based on the corresponding snapshot id.
...
Cross Catalog Lineage
There will be a system database Metadata Database in each catalog to manage table and data lineage. For the table and data lineage crossing multiple catalogs, users can join the lineage tables in different catalogs. For example, there are two catalogs, Flink ETL job reads data from catalog1 and writes data to catalog2. Users can query data lineage from source_snapshot_lineage in catalog1 and sink_snapshot_lineage in catalog2.
...
1. Implement lineage management based on FLIP-314
a) Currently we store table lineage when job creates its source and sink based on pipeline name, in FLIP-314 we get the the source/sink for job directly in the customized listener.
b) The table lineage should be deleted when user submit job failed or job goes to termination, we can get the job status and do the specific operations in FLIP-314
2. The number of snapshots that support checkpoint can be configured
...