Discussion thread | |
---|---|
Vote thread | |
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.
We would like to manage Flink Streaming & Batch ETL processing in Paimon based on the current ability of Flink, including:
1. Table lineage in Paimon. We want to create lineage between Paimon source and sink tables based on Flink job name, then we can the entire ETL topology for Paimon tables and Flink jobs.
2. Data lineage for Table. Flink streaming job creates a snapshot for Paimon table when it completes a checkpoint, we can create the snapshot lineage for source and sink tables.
Based on the above lineage, we could support the following capabilities in the integrated Flink ETL processing and Paimon tables
1. Query Paimon tables dependency, which allows users to manage their tables and Flink jobs better.
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 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 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.
Use Case
1. Create a catalog with options table-lineage
and data-lineage
. NOTICE: these two options are immutable.
-- 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
2. Create two tables and submit Flink ETL job.
-- Create a word table CREATE TABLE word_table (word STRING); -- Create a word count table CREATE TABLE word_count_table ( word STRING PRIMARY KEY NOT ENFORCED, cnt BITING ); -- Submit a flink job to ingest data from kafka to word_table INSERT INTO word_table SELECT FROM ...; SET 'pipeline.name'='job1'; INSERT INTO word_count_table SELECT word, count(*) FROM word_table GROUP BY word;
Paimon will create a relationship between word_table
and word_count_table
with job1
, as mentioned above, it includes table lineage and data lineage. For example, Checkpoint 1 in job1 consumes Snapshot 3 of word_table, and products Snapshot 5 in word_count_table, and so on. The information of table and data lineage is shown as follows.
3. Users can debug and validate their data of tables based on the table and data lineage, the steps are
a) Users can get the information that "Snapshot 5 of word_table generates Snapshot 7 in word_count_table" from lineage tables.
b) According to the time travel ability of Paimon, users can submit a Flink or Spark batch job to read Snapshot 5 from word_table, generates data to a new table word_count_table_snapshot_result whose schema is same as word_count_table.
c) Compare the data in word_count_table_snapshot_result and Snapshot 7 of word_count_table to validate the result of streaming ETL.
-- Create table word_count_table_snapshot_result from word_count_table CREATE TABLE word_count_table_snapshot_result LIKE word_count_table; -- Read snapshot 5 from word_table and write data to word_count_table_snapshot_result -- in batch mode INSERT INTO word_count_table_snapshot_result SELECT word, count(*) FROM word_table /*+ OPTIONS('scan.snapshot-id'='5') */ GROUP BY word; -- Compare the data in word_count_table_snapshot_result with -- snapshot 7 in word_count_table and get the diffs SELECT * FROM word_count_table_snapshot_result as L FULL OUTER JOIN word_count_table /*+ OPTIONS('scan.snapshot-id'='7') */ as R ON L.word=R.word where L.word IS NULL OR R.word IS NULL OR L.cnt != R.cnt;
In addition to the above data validation, users can also debug their jobs. For example, users can modify and submit a new job which will read Snapshot 5 of word_table and product results in a new table. Then they can compare the data with Snapshot 7 in word_count_table to check whether the new job meets their expectations.
4. Users can also correct data based on table and data lineages. Suppose that users want to correct data for table1 from Snapshot 4, and recompute data for downstream tables.
a) Full Recover
1> Create new table_s1 based on table1 and insert Snapshot 4 of table1 to table_s1, then users can correct data in table_s1
2> According to table lineage, create empty downstream table as table_s3
3> Submit new ETL job to read full data of Snapshot 4 from table_s1 and Snapshot 6 from table2, then read incremental data
Full recover will read the full data first, then read the incremental data, which may result in a large amount of computation and a lot of time. To avoid this situation, incremental recover can be considered.
b) Increment Recover
Increment recover will start a new Flink ETL job with savepoint, suppose that in the above data lineage, Snapshot 4 in table1 is related to Savepoint 5 in Flink ETL job, then the steps of increment recover:
1> Create new table_s1 based on table1 and insert Snapshot 4 of table1 to table_s1, then users can correct data in table_s1
2> Create downstream table table_s3 from table3 and insert all data of snapshot 2 from table3 to table_s3
3> Submit new Flink ETL job with Savepoint 5 and start to consume increment data from Snapshot 4 in table_s1 and Snapshot 6 in table2.
Proposed Changes
System Database
When users add option table-lineage
or data-lineage
to catalog DDL, Paimon will check and create system database 'paimon' automatically, then create options table, table lineage table and data lineage table.
Table Schema
1. catalog_options
Catalog options are stored in catalog_options table, it is used for Flink ETL to read lineage related options and create source and sink. There are key and value fields in the catalog_options and key is the primary key of the table.
Column Name | Column Type | Example |
key | STRING | "table-lineage" |
value | STRING | "true" |
2. source_job_lineage
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" |
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
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 |
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 |
Lineage Use Case
1. Query Table Lineage
-- Query upstream tables for given table SELECT S.database, S.table FROM source_job_lineage S JOIN sink_job_lineage T ON S.job=T.job where T.`database`='myDatabase' and T.`table`='myTable'; -- Query downstream tables for given table SELECT T.database, T.table FROM source_job_lineage S JOIN sink_job_lineage T ON S.job=T.job where S.`database`='myDatabase' and S.`table`='myTable';
2. Query Data Lineage
-- 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
We introduce a new option table-lineage
for Paimon catalog, users can set this option when they create a new catalog.
FlinkTableFactory
in Paimon will create source and sink for Flink ETL job, it can read pipeline.name
from Context
and save table lineage information for the ETL job.
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 FLIP-314.
Data Lineage
Job Submission
Similar to table lineage, we add a new option data-lineage
in Paimon catalog for data lineage. One Paimon table may be consumed by multiple Flink ETL jobs and produce different sink table.
We would like to align data lineage in the same table for different ETL jobs. For example, when ETL 1 generates Snapshot 9 in Table 5 From Snapshot 13 in Table 2 and Snapshot 12 in Table 3, we want that the ETL 2 will consume the same Snapshot 13 in Table 2 and Snapshot 12 in Table 3 to produce result Snapshot 15 in Table 6. In this way, we ensure that the result snapshots in Table 5 and Table 6 are consistent, users can perform queries on Table 5 and Table 6 to get a consistency results from Snapshot 9 and Snapshot 15. On the other hand, when users submit a new ETL job to consume Table 5 and Table 6, they can manage the data pipeline from table 1/2/3/4 with a unified version.
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.
In the first step we only support snapshot in source and checkpoint in ETL job correspond one-to-one, that means ETL job will trigger checkpoint when it finishes reading one snapshot from source. This helps us simplify runtime management without the need to coordinate checkpoints for multiple jobs. We will support this ability in the future.
Job Runtime
To align snapshot and checkpoint for table and ETL job, we introduce AlignedSnapshotSource
for Flink which will provide AlignedSourceReader
and AlignedEnumerator
.
1. AlignedSourceReader implements ExternallyInducedSourceReader
, the main functions are
a) Receive checkpoint events
b) Discover snapshots in source table
c) Send splits of snapshot to downstream operator ReadOperator
which will read data from table for the split
d) Check whether the received checkpoint should be triggered based on the snapshot and splits
2. AlignedEnumerator will do nothing except sending a CheckpointEvent
to AlignedSourceReader
when it receives checkpoint from CheckpointCoordinator
to notify SourceReader
to trigger checkpoint.
3. CheckpointEvent
is a special SourceEvent
which contains checkpoint id and will be sent to SourceReader
.
public class AlignedSnapshotSource implements Source { @Override public SplitEnumerator<SplitT, EnumChkT> createEnumerator(SplitEnumeratorContext<SplitT> enumContext) throws Exception { } @Override public SourceReader<T, SplitT> createReader(SourceReaderContext readerContext) throws Exception { return new AlignedSourceReader(readerContext); } } public class AlignedSourceReader implements ExternallyInducedSourceReader { private Map<Long, List<Split>> discoveredSplits; private boolean alignedSnapshotSended; private Optional<Long> receivedCheckpoint; @Override public void start() { // start async scan to find new snapshot and add to discoveredSplits } @Override public InputStatus pollNext(ReaderOutput<T> output) throws Exception { if (!alignedSnapshotSended && !discoveredSplits.isEmpty()) { // send current aligned splits alignedSnapshotSended = true; } return InputStatus.NOTHING_AVAILABLE; } @Override public void handleSourceEvents(SourceEvent sourceEvent) { // handle checkpoint Events receivedCheckpoint = Optional.of(((CheckpointEvent)sourceEvent).checkpointID); } @Override public Optional<Long> shouldTriggerCheckpoint() { if (alignedSnapshotSended && receivedCheckpoint.isPresent()) { alignedSnapshotSended = false; Optional<Long> lastReceivedCheckpoint = receivedCheckpoint; receivedCheckpoint = Optional.empty(); return lastReceivedCheckpoint; } return Optional.empty(); } } public class AlignedEnumerator implements SplitEnumerator { private final SplitEnumeratorContext<SourceSplit> context; @Override public Void snapshotState(long checkpointId) throws Exception { CheckpointEvent event = new CheckpointEvent(checkpointId); for (int i = 0; i < context.currentParallelism(); i++) { context.sendEventToSourceReader(i, event); } } } public class CheckpointEvent implements SourceEvent { private final long checkpointId; public CheckpointEvent(long checkpointId) { this.checkpointId = checkpointId; } }
The detailed processes are:
1. AlignedSourceReader
discover the latest snapshot from source table and add it to the list.
2. When pollNext
method in AlignedSourceReader
is called, it will check whether the snapshot belongs to the current checkpoint, if yes, it will send all snapshots of the snapshot to downstream operator.
3. AlignedEnumerator
will create CheckpointEvent
with checkpoint id and send it to AlignedSourceReader
when the method snapshotState
is called.
4. When AlignedSourceReader
has received CheckpointEvent
and the splits of specific snapshot been sent to downstream operator, it will return the checkpoint id in shouldTriggerCheckpoint
to trigger the checkpoint.
5. After checkpoint is completed, it will store checkpoint id and snapshot id to source_job_lineage table.
Besides source data lineage, we need to save sink data lineage too. We introduce CommitListener
for Committer
interface. Flink sink will notify CommitListener
to write checkpoint id and snapshot id to sink data lineage when the specific snapshot is committed in the sink table.
public interface Committer extends AutoCloseable { void commit(List<ManifestCommittable> globalCommittables, CommitListener commitListener) throws IOException, InterruptedException; } public interface CommitListener { /** Notifies the listener that the commit with the given {@code identifier} completed. */ void notifyCommitComplete(long identifier, long snapshotId, Snapshot.CommitKind commitKind); }
Cross Catalog Lineage
There will be a system 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.
SELECT k.snapshot_id FROM catalog2.paimon.sink_job_lineage k JOIN catalog1.paimon.source_job_lineage s ON s.job=k.job AND s.barrier_id=k.barrier_id WHERE s.job='job_name' AND s.snapshot_id=10
Plan For The Future
1. Implement lineage management based on FLIP-314
2. The number of snapshots that support checkpoint can be configured
3. Support branch in Paimon table that data management does not require copying data
[1] FLIP-314: Support Customized Job Lineage Listener
[2] https://issues.apache.org/jira/browse/FLINK-32042