Flink CDC版本：3.2.1
说明：本文从SchemaOperator接收到，表结构变更事件开始，表结构变更事件应由source端产生，本文不讨论。
可以先看流程图，研究源码。
参考文章：
Flink cdc3.0动态变更表结构——源码解析
一、源码解析
以Sink to doris举例：
SchemaOperator
org.apache.flink.cdc.runtime.operators.schema.SchemaOperator
判断是否是SchemaChangeEvent事件，调用processSchemaChangeEvents方法
/** * This method is guaranteed to not be called concurrently with other methods of the operator. */
@Override
public void processElement(StreamRecord<Event> streamRecord) throws InterruptedException, TimeoutException, ExecutionException { Event event = streamRecord.getValue(); if (event instanceof SchemaChangeEvent) { // (0)processSchemaChangeEvents((SchemaChangeEvent) event); } else if (event instanceof DataChangeEvent) { // (13)processDataChangeEvents(streamRecord, (DataChangeEvent) event); } else { throw new RuntimeException("Unknown event type in Stream record: " + event); }
}
调用handleSchemaChangeEvent方法：
private void processSchemaChangeEvents(SchemaChangeEvent event) throws InterruptedException, TimeoutException, ExecutionException { TableId tableId = event.tableId(); LOG.info( "{}> Table {} received SchemaChangeEvent {} and start to be blocked.", subTaskId, tableId, event); handleSchemaChangeEvent(tableId, event); // Update caches originalSchema.put(tableId, getLatestOriginalSchema(tableId)); schemaDivergesMap.put(tableId, checkSchemaDiverges(tableId)); List<TableId> optionalRoutedTable = getRoutedTables(tableId); if (!optionalRoutedTable.isEmpty()) { tableIdMappingCache .get(tableId) .forEach(routed -> evolvedSchema.put(routed, getLatestEvolvedSchema(routed))); } else { evolvedSchema.put(tableId, getLatestEvolvedSchema(tableId)); }
}
handleSchemaChangeEvent调用requestSchemaChange方法，请求修改Schema：
response.isAccepted()就是注册中心接收了此修改需求。进入if后，重点来了：output.collect(new StreamRecord<>(new FlushEvent(tableId))); 。注意这里发送了一个new FlushEvent(tableId)事件，这个事件会在SinkWriter用到，就是通知SinkWriter要执行flush，即把数据刷入到sink端数据库，和jdbc的commit相似。
FlushEvent内容非常简单只有tableId但是其类型是FlushEvent，此类的doc内容是：
- An {@link Event} from {@code SchemaOperator} to notify {@code DataSinkWriterOperator} that it
- start flushing.
也就是FlushEvent作为特殊数据传递事件，接收到此数据的DataSinkWriterOperator会触发其执行flushing操作，也就是将目前收到的所有数据都写入目标数据库。可以理解为：
schema修改后的数据 --> FlushEvent（新插入） --> schema修改前的数据
发送FlushEvent事件后执行requestSchemaChangeResult方法，此方法是while阻塞的方法，简而言之是等所有writer都完成了FlushEvent前数据的（旧表结构的数据）写入前，一直阻塞不发送新表结构的数据至下游。
最后finishedSchemaChangeEvents.forEach(e -> output.collect(new StreamRecord<>(e))); ，简略的说其内部是handler方法中生成的SchemaRegistryRequestHandler#applySchemaChange事件，将原始的SchemaChangeEvent转换成新的数据，还是根据Flink CDC的schema.change.behavior转换，其类型如下：
![[image-20250106113512324.png]]
具体将这些时间发送至下游怎么用暂时没有研究了。
private void handleSchemaChangeEvent(TableId tableId, SchemaChangeEvent schemaChangeEvent) throws InterruptedException, TimeoutException { if (schemaChangeBehavior == SchemaChangeBehavior.EXCEPTION && schemaChangeEvent.getType() != SchemaChangeEventType.CREATE_TABLE) { // CreateTableEvent should be applied even in EXCEPTION mode throw new RuntimeException( String.format( "Refused to apply schema change event %s in EXCEPTION mode.", schemaChangeEvent)); } // The request will block if another schema change event is being handled SchemaChangeResponse response = requestSchemaChange(tableId, schemaChangeEvent); // (1)if (response.isAccepted()) { // (3)LOG.info("{}> Sending the FlushEvent for table {}.", subTaskId, tableId); output.collect(new StreamRecord<>(new FlushEvent(tableId))); // (4)List<SchemaChangeEvent> expectedSchemaChangeEvents = response.getSchemaChangeEvents(); schemaOperatorMetrics.increaseSchemaChangeEvents(expectedSchemaChangeEvents.size()); // The request will block until flushing finished in each sink writer SchemaChangeResultResponse schemaEvolveResponse = requestSchemaChangeResult(); // (5) List<SchemaChangeEvent> finishedSchemaChangeEvents = schemaEvolveResponse.getFinishedSchemaChangeEvents(); // Update evolved schema changes based on apply results finishedSchemaChangeEvents.forEach(e -> output.collect(new StreamRecord<>(e))); } else if (response.isDuplicate()) { LOG.info( "{}> Schema change event {} has been handled in another subTask already.", subTaskId, schemaChangeEvent); } else if (response.isIgnored()) { LOG.info( "{}> Schema change event {} has been ignored. No schema evolution needed.", subTaskId, schemaChangeEvent); } else { throw new IllegalStateException("Unexpected response status " + response); }
}
requestSchemaChange是一个阻塞的方法（while (true)），发送SchemaChangeRequest直到返回的response不是Busy。可以看到发送的的SchemaChangeRequest。
private SchemaChangeResponse requestSchemaChange( TableId tableId, SchemaChangeEvent schemaChangeEvent) throws InterruptedException, TimeoutException { long schemaEvolveTimeOutMillis = System.currentTimeMillis() + rpcTimeOutInMillis; while (true) { SchemaChangeResponse response = sendRequestToCoordinator( new SchemaChangeRequest(tableId, schemaChangeEvent, subTaskId)); if (response.isRegistryBusy()) { // (2)if (System.currentTimeMillis() < schemaEvolveTimeOutMillis) { LOG.info( "{}> Schema Registry is busy now, waiting for next request...", subTaskId); Thread.sleep(1000); } else { throw new TimeoutException("TimeOut when requesting schema change"); } } else { return response; } }
}
sendRequestToCoordinator方法是org.apache.flink.runtime.jobgraph.tasks.TaskOperatorEventGateway类的，也就Flink的内部类。
实习类有：
（1）org.apache.flink.runtime.taskmanager.NoOpTaskOperatorEventGateway
（2）org.apache.flink.runtime.taskexecutor.rpc.RpcTaskOperatorEventGateway
内部具体逻辑暂不深入了解。
其实际发送至 SchemaRegistry#handleEventFromOperator
private <REQUEST extends CoordinationRequest, RESPONSE extends CoordinationResponse> RESPONSE sendRequestToCoordinator(REQUEST request) { try { CompletableFuture<CoordinationResponse> responseFuture = toCoordinator.sendRequestToCoordinator( getOperatorID(), new SerializedValue<>(request)); return CoordinationResponseUtils.unwrap(responseFuture.get()); } catch (Exception e) { throw new IllegalStateException( "Failed to send request to coordinator: " + request.toString(), e); }
}
requestSchemaChangeResult执行的操作非常简单，就是等待返回，如果跳出while方法结束，就代表sink端已经完成所有旧数据的flush，在此之前SchemaOperator类不会向下游发送新数据，因为FlushEvent后的数据都是schema变更的后的新数据了。
private SchemaChangeResultResponse requestSchemaChangeResult() throws InterruptedException, TimeoutException { CoordinationResponse coordinationResponse = sendRequestToCoordinator(new SchemaChangeResultRequest()); long nextRpcTimeOutMillis = System.currentTimeMillis() + rpcTimeOutInMillis; while (coordinationResponse instanceof SchemaChangeProcessingResponse) { // (6) (7)if (System.currentTimeMillis() < nextRpcTimeOutMillis) { Thread.sleep(1000); coordinationResponse = sendRequestToCoordinator(new SchemaChangeResultRequest()); } else { throw new TimeoutException("TimeOut when requesting release upstream"); } } return ((SchemaChangeResultResponse) coordinationResponse);
}
这里的toCoordinator.sendRequestToCoordinator也是使用flink内部的调用过程，暂不做研究。
这个发送过程也是被org.apache.flink.cdc.runtime.operators.schema.coordinator.SchemaRegistry#handleCoordinationRequest接收了，并在if (request instanceof SchemaChangeResultRequest)内处理其逻辑。
private <REQUEST extends CoordinationRequest, RESPONSE extends CoordinationResponse> RESPONSE sendRequestToCoordinator(REQUEST request) { try { CompletableFuture<CoordinationResponse> responseFuture = toCoordinator.sendRequestToCoordinator( getOperatorID(), new SerializedValue<>(request)); return CoordinationResponseUtils.unwrap(responseFuture.get()); } catch (Exception e) { throw new IllegalStateException( "Failed to send request to coordinator: " + request.toString(), e); }
}
SchemaRegistry
org.apache.flink.cdc.runtime.operators.schema.coordinator.SchemaRegistry
toCoordinator.sendRequestToCoordinator方法就由handleCoordinationRequest接收，进入request instanceof SchemaChangeRequest中的handleSchemaChangeRequest方法。
@Override
public CompletableFuture<CoordinationResponse> handleCoordinationRequest( CoordinationRequest request) { CompletableFuture<CoordinationResponse> responseFuture = new CompletableFuture<>(); runInEventLoop( () -> { try { if (request instanceof SchemaChangeRequest) { SchemaChangeRequest schemaChangeRequest = (SchemaChangeRequest) request; requestHandler.handleSchemaChangeRequest( schemaChangeRequest, responseFuture); } else if (request instanceof SchemaChangeResultRequest) { requestHandler.getSchemaChangeResult(responseFuture); } else if (request instanceof GetEvolvedSchemaRequest) { handleGetEvolvedSchemaRequest( ((GetEvolvedSchemaRequest) request), responseFuture); } else if (request instanceof GetOriginalSchemaRequest) { handleGetOriginalSchemaRequest( (GetOriginalSchemaRequest) request, responseFuture); } else { throw new IllegalArgumentException( "Unrecognized CoordinationRequest type: " + request); } } catch (Throwable t) { context.failJob(t); throw t; } }, "handling coordination request %s", request); return responseFuture;
}
SchemaRegistry#handleEventFromOperator方法用于处理DataSinkWriterOperator#handleFlushEvent发送而来的FlushSuccessEvent事件。还是使用handler执行具体逻辑：SchemaRegistryRequestHandler#flushSuccess
@Override
public void handleEventFromOperator(int subtask, int attemptNumber, OperatorEvent event) { runInEventLoop( () -> { try { if (event instanceof FlushSuccessEvent) { FlushSuccessEvent flushSuccessEvent = (FlushSuccessEvent) event; LOG.info( "Sink subtask {} succeed flushing for table {}.", flushSuccessEvent.getSubtask(), flushSuccessEvent.getTableId().toString()); requestHandler.flushSuccess( flushSuccessEvent.getTableId(), flushSuccessEvent.getSubtask(), currentParallelism); } else if (event instanceof SinkWriterRegisterEvent) { requestHandler.registerSinkWriter( ((SinkWriterRegisterEvent) event).getSubtask()); } else { throw new FlinkException("Unrecognized Operator Event: " + event); } } catch (Throwable t) { context.failJob(t); throw t; } }, "handling event %s from subTask %d", event, subtask);
}
SchemaRegistryRequestHandler
SchemaRegistryRequestHandler是SchemaRegistry的执行器，类中schemaChangeStatus是自己的状态记录状态的。
而pendingSubTaskIds是记录待处理的任务id的，即数据流ID，是含有一个任务所有的并行度的子任务ID。
此处：
（1）pendingSubTaskIds空 -> 继续执行
（2）requestSubTaskId和发送过来的一样，则为移除头一个。
（3）其他pendingSubTaskIds不为空情形，则直接返回SchemaChangeResponse.busy()，此处的busy就和SchemaOperator的response.isRegistryBusy()对应上了。
继续执行：
calculateDerivedSchemaChangeEvents方法是对事件作息写转换，根据的是flink的schema evolution的策略进行转换，例如通过返回空集合的方式进行忽略 。
`schema.change.behavior` is of enum type, and could be set to `exception`, `evolve`, `try_evolve`, `lenient` or `ignore`.
而后此handler的状态修改为WAITING_FOR_FLUSH。
并返回ResponseCode.ACCEPTED的状态，此时程序跳转回SchemaOperator#handleSchemaChangeEvent方法。
SchemaRegistryRequestHandler#handleSchemaChangeRequest方法：
/** * Handle the {@link SchemaChangeRequest} and wait for all sink subtasks flushing. * * @param request the received SchemaChangeRequest */public void handleSchemaChangeRequest( SchemaChangeRequest request, CompletableFuture<CoordinationResponse> response) { // We use requester subTask ID as the pending ticket, because there will be at most 1 schema // change requests simultaneously from each subTask int requestSubTaskId = request.getSubTaskId(); synchronized (schemaChangeRequestLock) { // Make sure we handle the first request in the pending list to avoid out-of-order // waiting and blocks checkpointing mechanism. if (schemaChangeStatus == RequestStatus.IDLE) { if (pendingSubTaskIds.isEmpty()) { LOG.info( "Received schema change event request {} from table {} from subTask {}. Pending list is empty, handling this.", request.getSchemaChangeEvent(), request.getTableId().toString(), requestSubTaskId); } else if (pendingSubTaskIds.get(0) == requestSubTaskId) { LOG.info( "Received schema change event request {} from table {} from subTask {}. It is on the first of the pending list, handling this.", request.getSchemaChangeEvent(), request.getTableId().toString(), requestSubTaskId); pendingSubTaskIds.remove(0); } else { LOG.info( "Received schema change event request {} from table {} from subTask {}. It is not the first of the pending list ({}).", request.getSchemaChangeEvent(), request.getTableId().toString(), requestSubTaskId, pendingSubTaskIds); if (!pendingSubTaskIds.contains(requestSubTaskId)) { pendingSubTaskIds.add(requestSubTaskId); } response.complete(wrap(SchemaChangeResponse.busy())); // (2) return; } SchemaChangeEvent event = request.getSchemaChangeEvent(); // If this schema change event has been requested by another subTask, ignore it. if (schemaManager.isOriginalSchemaChangeEventRedundant(event)) { LOG.info("Event {} has been addressed before, ignoring it.", event); clearCurrentSchemaChangeRequest(); LOG.info( "SchemaChangeStatus switched from WAITING_FOR_FLUSH to IDLE for request {} due to duplicated request.", request); response.complete(wrap(SchemaChangeResponse.duplicate())); return; } schemaManager.applyOriginalSchemaChange(event); List<SchemaChangeEvent> derivedSchemaChangeEvents = calculateDerivedSchemaChangeEvents(request.getSchemaChangeEvent()); // (14)// If this schema change event is filtered out by LENIENT mode or merging table // route strategies, ignore it. if (derivedSchemaChangeEvents.isEmpty()) { LOG.info("Event {} is omitted from sending to downstream, ignoring it.", event); clearCurrentSchemaChangeRequest(); LOG.info( "SchemaChangeStatus switched from WAITING_FOR_FLUSH to IDLE for request {} due to ignored request.", request); response.complete(wrap(SchemaChangeResponse.ignored())); return; } LOG.info( "SchemaChangeStatus switched from IDLE to WAITING_FOR_FLUSH, other requests will be blocked."); // This request has been accepted. schemaChangeStatus = RequestStatus.WAITING_FOR_FLUSH; // (3)currentDerivedSchemaChangeEvents = new ArrayList<>(derivedSchemaChangeEvents); response.complete(wrap(SchemaChangeResponse.accepted(derivedSchemaChangeEvents))); // (3) } else { LOG.info( "Schema Registry is busy processing a schema change request, could not handle request {} for now. Added {} to pending list ({}).", request, requestSubTaskId, pendingSubTaskIds); if (!pendingSubTaskIds.contains(requestSubTaskId)) { pendingSubTaskIds.add(requestSubTaskId); } response.complete(wrap(SchemaChangeResponse.busy())); // (2) } }
}
SchemaRegistryRequestHandler#getSchemaChangeResult方法：
内容就是检查类成员变量SchemaRegistryRequestHandler#schemaChangeStatus的状态：
- FINISHED -> 重置自身状态并返回FINISHED状态
- 非FINISHED -> 返回Processing状态，
SchemaOperator#requestSchemaChangeResult接到SchemaChangeProcessingResponse会在while一直循环等待阻塞。
public void getSchemaChangeResult(CompletableFuture<CoordinationResponse> response) { Preconditions.checkState( schemaChangeStatus != RequestStatus.IDLE, "Illegal schemaChangeStatus: should not be IDLE before getting schema change request results."); if (schemaChangeStatus == RequestStatus.FINISHED) { // (12)schemaChangeStatus = RequestStatus.IDLE; LOG.info( "SchemaChangeStatus switched from FINISHED to IDLE for request {}", currentDerivedSchemaChangeEvents); // This request has been finished, return it and prepare for the next request List<SchemaChangeEvent> finishedEvents = clearCurrentSchemaChangeRequest(); SchemaChangeResultResponse resultResponse = new SchemaChangeResultResponse(finishedEvents); response.complete(wrap(resultResponse)); } else { // Still working on schema change request, waiting it response.complete(wrap(new SchemaChangeProcessingResponse())); }
}
方法flushSuccess用于处理DataSinkWriterOperator返回的FlushSuccessEvent事件。这里有点不好理解。
activeSinkWriters是记录所有可用的writer的索引，也就是说writer的并行度可能大于1，activeSinkWriters记录的是writer的索引，接收的FlushSuccessEvent只是其中一个writer发送的。需要等待所有writer都完成flush才能确定所有的schema修改前的数据都写入数据库了。
（1）if (activeSinkWriters.size() < parallelism)内的就是上述过程。
（2）if (flushedSinkWriters.equals(activeSinkWriters))代表所有writer都完成了flush。而后修改handler状态为RequestStatus.APPLYING，即此handler正在apply schema change。接下来执行applySchemaChange方法 。
/** * Record flushed sink subtasks after receiving FlushSuccessEvent. * * @param tableId the subtask in SchemaOperator and table that the FlushEvent is about * @param sinkSubtask the sink subtask succeed flushing */public void flushSuccess(TableId tableId, int sinkSubtask, int parallelism) { flushedSinkWriters.add(sinkSubtask); if (activeSinkWriters.size() < parallelism) { LOG.info( "Not all active sink writers have been registered. Current {}, expected {}.", activeSinkWriters.size(), parallelism); return; } if (flushedSinkWriters.equals(activeSinkWriters)) { Preconditions.checkState( schemaChangeStatus == RequestStatus.WAITING_FOR_FLUSH, "Illegal schemaChangeStatus state: should be WAITING_FOR_FLUSH before collecting enough FlushEvents, not " + schemaChangeStatus); schemaChangeStatus = RequestStatus.APPLYING; // (9)LOG.info( "All sink subtask have flushed for table {}. Start to apply schema change.", tableId.toString()); schemaChangeThreadPool.submit( () -> applySchemaChange(tableId, currentDerivedSchemaChangeEvents)); }
}
SchemaRegistryRequestHandler#applySchemaChange方法：
内部主要是schemaManager.applyEvolvedSchemaChange(changeEvent)即执行表结构变更操作，其接口类org.apache.flink.cdc.common.sink.MetadataApplier的doc内容：
- {@code MetadataApplier} is used to apply metadata changes to external systems.
可以看到schemaManager至对外部数据执行的表结构变更，其实就是sink端的数据库，其内部一般是收到需要变更的内容，拼接SQL并发送到数据库执行。
最后，修改handler状态为RequestStatus.FINISHED。
好像此FlushSuccessEvent没有继续向SchemaOperator继续传递，其实不然，SchemaOperator是不断向SchemaRegistry发送请求的：SchemaOperator#requestSchemaChangeResult。
而SchemaRegistry是根据handler状态判断返回值类型的
SchemaRegistryRequestHandler#getSchemaChangeResult，此时handler状态已经是RequestStatus.FINISHED，SchemaRegistry就会给CompletableFuture填充非SchemaChangeProcessingResponse了，SchemaOperator类就中断阻塞，继续向下游发送数据了。
/** * Apply the schema change to the external system. * * @param tableId the table need to change schema * @param derivedSchemaChangeEvents list of the schema changes */private void applySchemaChange( TableId tableId, List<SchemaChangeEvent> derivedSchemaChangeEvents) { for (SchemaChangeEvent changeEvent : derivedSchemaChangeEvents) { if (changeEvent.getType() != SchemaChangeEventType.CREATE_TABLE) { if (schemaChangeBehavior == SchemaChangeBehavior.IGNORE) { currentIgnoredSchemaChanges.add(changeEvent); continue; } } if (!metadataApplier.acceptsSchemaEvolutionType(changeEvent.getType())) { LOG.info("Ignored schema change {} to table {}.", changeEvent, tableId); currentIgnoredSchemaChanges.add(changeEvent); } else { try { metadataApplier.applySchemaChange(changeEvent); LOG.info("Applied schema change {} to table {}.", changeEvent, tableId); schemaManager.applyEvolvedSchemaChange(changeEvent); currentFinishedSchemaChanges.add(changeEvent); } catch (Throwable t) { LOG.error( "Failed to apply schema change {} to table {}. Caused by: {}", changeEvent, tableId, t); if (!shouldIgnoreException(t)) { currentChangeException = t; break; } else { LOG.warn( "Failed to apply event {}, but keeps running in tolerant mode. Caused by: {}", changeEvent, t); } } } } Preconditions.checkState( schemaChangeStatus == RequestStatus.APPLYING, "Illegal schemaChangeStatus state: should be APPLYING before applySchemaChange finishes, not " + schemaChangeStatus); schemaChangeStatus = RequestStatus.FINISHED; LOG.info( "SchemaChangeStatus switched from APPLYING to FINISHED for request {}.", currentDerivedSchemaChangeEvents);
}
SchemaRegistryRequestHandler.RequestStatus类是就handler类状态的类型。具体状态流程可见文档。
// Schema change event state could transfer in the following way:
//
// -------- B --------
// | |
// v |
// -------- ---------------------
// | IDLE | --- A --> | WAITING_FOR_FLUSH |
// -------- ---------------------
// ^ |
// E C
// \ v
// ------------ ------------
// | FINISHED | <-- D -- | APPLYING |
// ------------ ------------
//
// A: When a request came to an idling request handler.
// B: When current request is duplicate or ignored by LENIENT / routed table merging
// strategies.
// C: When schema registry collected enough flush success events, and actually started to apply
// schema changes.
// D: When schema change application finishes (successfully or with exceptions)
// E: When current schema change request result has been retrieved by SchemaOperator, and ready
// for the next request.
private enum RequestStatus { IDLE, WAITING_FOR_FLUSH, APPLYING, FINISHED
}
接下来看下：Sink端的事件处理：
DataSinkWriterOperator
org.apache.flink.cdc.runtime.operators.sink.DataSinkWriterOperator
org.apache.flink.cdc.runtime.operators.sink.DataSinkWriterOperator#processElement方法：
重点是对FlushEvent的处理
@Override
public void processElement(StreamRecord<Event> element) throws Exception { Event event = element.getValue(); // FlushEvent triggers flush if (event instanceof FlushEvent) { handleFlushEvent(((FlushEvent) event)); return; } // CreateTableEvent marks the table as processed directly if (event instanceof CreateTableEvent) { processedTableIds.add(((CreateTableEvent) event).tableId()); this.<OneInputStreamOperator<Event, CommittableMessage<CommT>>>getFlinkWriterOperator() .processElement(element); return; } // Check if the table is processed before emitting all other events, because we have to make // sure that sink have a view of the full schema before processing any change events, // including schema changes. ChangeEvent changeEvent = (ChangeEvent) event; if (!processedTableIds.contains(changeEvent.tableId())) { emitLatestSchema(changeEvent.tableId()); processedTableIds.add(changeEvent.tableId()); } processedTableIds.add(changeEvent.tableId()); this.<OneInputStreamOperator<Event, CommittableMessage<CommT>>>getFlinkWriterOperator() .processElement(element);
}
handleFlushEvent方法内只有两个操作：
- flush： 将目前已经接受到所有数据写入目标库（相当于jdbc的commit操作）。
- 发送事件：发送FlushSuccess。
notifyFlushSuccess内容见类SchemaEvolutionClient
private void handleFlushEvent(FlushEvent event) throws Exception { copySinkWriter.flush(false); // (8) schemaEvolutionClient.notifyFlushSuccess( getRuntimeContext().getIndexOfThisSubtask(), event.getTableId()); // (9)
}
SchemaEvolutionClient
org.apache.flink.cdc.runtime.operators.sink.SchemaEvolutionClient
org.apache.flink.cdc.runtime.operators.sink.SchemaEvolutionClient#notifyFlushSuccess方法：
发送了FlushSuccessEvent事件至SchemaRegistry类的handleEventFromOperator方法。
public void notifyFlushSuccess(int subtask, TableId tableId) throws IOException { toCoordinator.sendOperatorEventToCoordinator( schemaOperatorID, new SerializedValue<>(new FlushSuccessEvent(subtask, tableId)));
}
TaskOperatorEventGateway
org.apache.flink.runtime.jobgraph.tasks.TaskOperatorEventGateway
SchemaOperator和DataSinkWriterOperator中的toCoordinator都是此类对象。
/*
Gateway to send an OperatorEvent or CoordinationRequest from a Task to the OperatorCoordinator JobManager side.
This is the first step in the chain of sending Operator Events and Requests from Operator to Coordinator. Each layer adds further context, so that the inner layers do not need to know about the complete context, which keeps dependencies small and makes testing easier.OperatorEventGateway takes the event, enriches the event with the OperatorID, and forwards it to:
TaskOperatorEventGateway enriches the event with the ExecutionAttemptID and forwards it to the:
JobMasterOperatorEventGateway which is RPC interface from the TaskManager to the JobManager.
*/
public interface TaskOperatorEventGateway { /** * Sends an event from the operator (identified by the given operator ID) to the operator * coordinator (identified by the same ID). */ void sendOperatorEventToCoordinator(OperatorID operator, SerializedValue<OperatorEvent> event); /** * Sends a request from current operator to a specified operator coordinator which is identified * by the given operator ID and return the response. */ CompletableFuture<CoordinationResponse> sendRequestToCoordinator( OperatorID operator, SerializedValue<CoordinationRequest> request);
}
MetadataApplier
org.apache.flink.cdc.common.sink.MetadataApplier
此类负责将表结构修改的事件，转化成为DDL，发送给目标sink端数据库执行。
/** {@code MetadataApplier} is used to apply metadata changes to external systems. */
@PublicEvolving
public interface MetadataApplier extends Serializable { /** Apply the given {@link SchemaChangeEvent} to external systems. */ void applySchemaChange(SchemaChangeEvent schemaChangeEvent) throws SchemaEvolveException; // (10) /** Sets enabled schema evolution event types of current metadata applier. */ default MetadataApplier setAcceptedSchemaEvolutionTypes( Set<SchemaChangeEventType> schemaEvolutionTypes) { return this; } /** Checks if this metadata applier should this event type. */ default boolean acceptsSchemaEvolutionType(SchemaChangeEventType schemaChangeEventType) { return true; } /** Checks what kind of schema change events downstream can handle. */ default Set<SchemaChangeEventType> getSupportedSchemaEvolutionTypes() { return Arrays.stream(SchemaChangeEventTypeFamily.ALL).collect(Collectors.toSet()); }
}
DorisMetadataApplier
org.apache.flink.cdc.connectors.doris.sink.DorisMetadataApplier 实现 MetadataApplier
org.apache.flink.cdc.connectors.doris.sink.DorisMetadataApplier#applySchemaChange：
以：
// (10)
@Override
public void applySchemaChange(SchemaChangeEvent event) throws SchemaEvolveException { try { // send schema change op to doris if (event instanceof CreateTableEvent) { applyCreateTableEvent((CreateTableEvent) event); } else if (event instanceof AddColumnEvent) { applyAddColumnEvent((AddColumnEvent) event); } else if (event instanceof DropColumnEvent) { applyDropColumnEvent((DropColumnEvent) event); } else if (event instanceof RenameColumnEvent) { applyRenameColumnEvent((RenameColumnEvent) event); } else if (event instanceof AlterColumnTypeEvent) { applyAlterColumnTypeEvent((AlterColumnTypeEvent) event); } else { throw new UnsupportedSchemaChangeEventException(event); } } catch (Exception ex) { throw new SchemaEvolveException(event, ex.getMessage(), null); }
}
applyAddColumnEvent举例说明：
这里仅做一些转换
private void applyAddColumnEvent(AddColumnEvent event) throws IOException, IllegalArgumentException { TableId tableId = event.tableId(); List<AddColumnEvent.ColumnWithPosition> addedColumns = event.getAddedColumns(); for (AddColumnEvent.ColumnWithPosition col : addedColumns) { Column column = col.getAddColumn(); FieldSchema addFieldSchema = new FieldSchema( column.getName(), buildTypeString(column.getType()), column.getDefaultValueExpression(), column.getComment()); schemaChangeManager.addColumn( tableId.getSchemaName(), tableId.getTableName(), addFieldSchema); }
}
SchemaChangeManager
org.apache.doris.flink.sink.schema.SchemaChangeManager
org.apache.doris.flink.sink.schema.SchemaChangeManager#addColumn方法：
SchemaChangeHelper是拼接SQL用的。schemaChange方法向数据库发送需要执行的SQL。
public boolean addColumn(String database, String table, FieldSchema field) throws IOException, IllegalArgumentException { if (checkColumnExists(database, table, field.getName())) { LOG.warn( "The column {} already exists in table {}, no need to add it again", field.getName(), table); return true; } String tableIdentifier = getTableIdentifier(database, table); String addColumnDDL = SchemaChangeHelper.buildAddColumnDDL(tableIdentifier, field); return schemaChange( database, table, buildRequestParam(false, field.getName()), addColumnDDL);
}
SchemaChangeHelper
org.apache.doris.flink.sink.schema.SchemaChangeHelper
org.apache.doris.flink.sink.schema.SchemaChangeHelper#buildAddColumnDDL
用ADD_DDL字符串模板拼接SQL：
// (11)
private static final String ADD_DDL = "ALTER TABLE %s ADD COLUMN %s %s";
public static String buildAddColumnDDL(String tableIdentifier, FieldSchema fieldSchema) { String name = fieldSchema.getName(); String type = fieldSchema.getTypeString(); String defaultValue = fieldSchema.getDefaultValue(); String comment = fieldSchema.getComment(); StringBuilder addDDL = new StringBuilder( String.format( ADD_DDL, DorisSchemaFactory.quoteTableIdentifier(tableIdentifier), DorisSchemaFactory.identifier(name), type)); if (defaultValue != null) { addDDL.append(" DEFAULT ").append(DorisSchemaFactory.quoteDefaultValue(defaultValue)); } commentColumn(addDDL, comment); return addDDL.toString();
}
流程总结：
- SchemaOperator接收到SchemaChangeEvent，发送
SchemaChangeRequest至SchemaRegistry。 - SchemaRegistry内部执行器是SchemaRegistryRequestHandler，简称handler，handler内部持有有状态
schemaChangeStatus其判断是否正在执行之前的Request，如果是则返回busy状态。如果不是则返回accept状态。其状态修改由RequestStatus.IDLE为RequestStatus.WAITING_FOR_FLUSH。 - SchemaOperator如果收到busy状态则sleep后再次发起请求，阻塞直到，收到accept状态，则发送一条
FlushEvent至下游，之后发送SchemaChangeResultRequest至SchemaRegistry,等待返回结果如果是SchemaChangeProcessingResponse则认为SchemaChange还没有结束，sleep后再次发起请求，阻塞直至收到非SchemaChangeProcessingResponse。此时阻塞，不再发送新的表结构的数据至下游。 - SchemaRegistry收到
SchemaChangeResultRequest,handler会检查自身状态schemaChangeStatus，如果不是RequestStatus.FINISHED，则返回SchemaChangeProcessingResponse。 - DataSinkWriterOperator收到
FlushEvent，并执行flush操作，将所有已经收到的老表结构的数据写入数据库。并发送FlushSuccessEvent给SchemaRegistry。 - SchemaRegistry的handler收集
FlushSuccessEvent，当收到所有的subtask的FlushSuccessEvent后，修改自身状态为RequestStatus.APPLYING。后使用MetadataApplier执行sink端（外）数据库的表结构变更。执行后修改自身状态为RequestStatus.FINISHED。 - 当SchemaOperator再次发送
SchemaChangeResultRequest，且SchemaRegistry的handler的状态为RequestStatus.FINISHED，SchemaRegistry返回给其结果为非SchemaChangeProcessingResponse，SchemaOperator将不再阻塞，开始将新的表结构的数据继续发送至下游。
二、流程图
下图中的序号已经在源码中表示，可以在源码中搜索。
