[debezium 源码分析] BinlogReader和SnapshotReader

debezium工作方式主要有2种，分别是全量和增量订阅。 全量对应SnapshotReader，增量订阅对应BinlogReader;前者会读取全表数据，后者会对binlog事件作出对应的处理逻辑。

SnapshotReader

该类内最重要的方法就是execute()，调用该方法也就是意味着一次snapshot操作，全量任务是在子线程里运行的。现在按序分析该方法的执行过程。

step 0

...
            logger.info("Step 0: disabling autocommit and enabling repeatable read transactions");
            mysql.setAutoCommit(false);
            sql.set("SET TRANSACTION ISOLATION LEVEL REPEATABLE READ");
            mysql.execute(sql.get());
...

将mysql事务隔离级别设置为REPEATABLE READ;

step 1

...
try {
    logger.info("Step 1: flush and obtain global read lock to prevent writes to database");
    sql.set("FLUSH TABLES WITH READ LOCK");
    mysql.execute(sql.get());
    lockAcquired = clock.currentTimeInMillis();
    metrics.globalLockAcquired();
    isLocked = true;
} catch (SQLException e) {
    logger.info("Step 1: unable to flush and acquire global read lock, will use table read locks after reading table names");
    // Continue anyway, since RDS (among others) don't allow setting a global lock
    assert !isLocked;
}
...

对数据库加全局只读锁。如果是生产环境，需要注意这一点。

step 2

logger.info("Step 2: start transaction with consistent snapshot");
sql.set("START TRANSACTION WITH CONSISTENT SNAPSHOT");
mysql.execute(sql.get());
isTxnStarted = true;

对存储引擎使用一致性读。根据mysql官网内容，除了REPEATABLE READ事务隔离级别，其它级别下WITH CONSISTENTSNAPSHOT会被忽略。

step 3

step = 3;
if (isLocked) {
    // Obtain the binlog position and update the SourceInfo in the context. This means that all source records
    // generated as part of the snapshot will contain the binlog position of the snapshot.
    readBinlogPosition(step++, source, mysql, sql);
}

// -------------------
// READ DATABASE NAMES
// -------------------
// Get the list of databases ...
if (!isRunning()) return;
logger.info("Step {}: read list of available databases", step++);
final List<String> databaseNames = new ArrayList<>();
sql.set("SHOW DATABASES");
mysql.query(sql.get(), rs -> {
    while (rs.next()) {
        databaseNames.add(rs.getString(1));
    }
});
logger.info("\t list of available databases is: {}", databaseNames);

如果全局读锁加锁成功，会调用readBinlogPosition方法。看readBinlogPosition方法

protected void readBinlogPosition(int step, SourceInfo source, JdbcConnection mysql, AtomicReference<String> sql) throws SQLException {
    logger.info("Step {}: read binlog position of MySQL master", step);
    String showMasterStmt = "SHOW MASTER STATUS";
    sql.set(showMasterStmt);
    mysql.query(sql.get(), rs -> {
        if (rs.next()) {
            String binlogFilename = rs.getString(1);
            long binlogPosition = rs.getLong(2);
            source.setBinlogStartPoint(binlogFilename, binlogPosition);
            if (rs.getMetaData().getColumnCount() > 4) {
                // This column exists only in MySQL 5.6.5 or later ...
                String gtidSet = rs.getString(5);// GTID set, may be null, blank, or contain a GTID set
                source.setCompletedGtidSet(gtidSet);
                logger.info("\t using binlog '{}' at position '{}' and gtid '{}'", binlogFilename, binlogPosition,
                            gtidSet);
            } else {
                logger.info("\t using binlog '{}' at position '{}'", binlogFilename, binlogPosition);
            }
            source.startSnapshot();
        } else {
            throw new IllegalStateException("Cannot read the binlog filename and position via '" + showMasterStmt
                    + "'. Make sure your server is correctly configured");
        }
    });
}

该方法会读取当前binlog文件名以及位点，并保存。现在回到step 3剩余代码，获取所有可用的数据库名。

step 4

logger.info("Step {}: read list of available tables in each database", step++);
List<TableId> tableIds = new ArrayList<>();
final Map<String, List<TableId>> tableIdsByDbName = new HashMap<>();
final Set<String> readableDatabaseNames = new HashSet<>();
for (String dbName : databaseNames) {
    try {
        // MySQL sometimes considers some local files as databases (see DBZ-164),
        // so we will simply try each one and ignore the problematic ones ...
        sql.set("SHOW FULL TABLES IN " + quote(dbName) + " where Table_Type = 'BASE TABLE'");
        mysql.query(sql.get(), rs -> {
            while (rs.next() && isRunning()) {
                TableId id = new TableId(dbName, null, rs.getString(1));
                if (filters.tableFilter().test(id)) {
                    tableIds.add(id);
                    tableIdsByDbName.computeIfAbsent(dbName, k -> new ArrayList<>()).add(id);
                    logger.info("\t including '{}'", id);
                } else {
                    logger.info("\t '{}' is filtered out, discarding", id);
                }
            }
        });
        readableDatabaseNames.add(dbName);
    } catch (SQLException e) {
        // We were unable to execute the query or process the results, so skip this ...
        logger.warn("\t skipping database '{}' due to error reading tables: {}", dbName, e.getMessage());
    }
}
final Set<String> includedDatabaseNames = readableDatabaseNames.stream().filter(filters.databaseFilter()).collect(Collectors.toSet());
logger.info("\tsnapshot continuing with database(s): {}", includedDatabaseNames);

获取数据库所有可用的表，之后对其进行过滤，得到用户想要的表。

step 5

if (!isLocked) {
    // ------------------------------------
    // LOCK TABLES and READ BINLOG POSITION
    // ------------------------------------
    // We were not able to acquire the global read lock, so instead we have to obtain a read lock on each table.
    // This requires different privileges than normal, and also means we can't unlock the tables without
    // implicitly committing our transaction ...
    if (!context.userHasPrivileges("LOCK TABLES")) {
        // We don't have the right privileges
        throw new ConnectException("User does not have the 'LOCK TABLES' privilege required to obtain a "
                + "consistent snapshot by preventing concurrent writes to tables.");
    }
    // We have the required privileges, so try to lock all of the tables we're interested in ...
    logger.info("Step {}: flush and obtain read lock for {} tables (preventing writes)", step++, tableIds.size());
    String tableList = tableIds.stream()
            .map(tid -> quote(tid))
            .reduce((r, element) -> r+ "," + element)
            .orElse(null);
    if (tableList != null) {
        sql.set("FLUSH TABLES " + tableList + " WITH READ LOCK");
        mysql.execute(sql.get());
    }
    lockAcquired = clock.currentTimeInMillis();
    metrics.globalLockAcquired();
    isLocked = true;
    tableLocks = true;

    // Our tables are locked, so read the binlog position ...
    readBinlogPosition(step++, source, mysql, sql);
}

如果数据库全局读锁加锁失败，就对每个表加表级读锁，然后调用readBinlogPosition方法。

step 6

schema.applyDdl(source, null, setSystemVariablesStatement, this::enqueueSchemaChanges);

// Add DROP TABLE statements for all tables that we knew about AND those tables found in the databases ...
Set<TableId> allTableIds = new HashSet<>(schema.tables().tableIds());
allTableIds.addAll(tableIds);
allTableIds.stream()
           .filter(id -> isRunning()) // ignore all subsequent tables if this reader is stopped
           .forEach(tableId -> schema.applyDdl(source, tableId.schema(),
                                               "DROP TABLE IF EXISTS " + quote(tableId),
                                               this::enqueueSchemaChanges));

// Add a DROP DATABASE statement for each database that we no longer know about ...
schema.tables().tableIds().stream().map(TableId::catalog)
      .filter(Predicates.not(readableDatabaseNames::contains))
      .filter(id -> isRunning()) // ignore all subsequent tables if this reader is stopped
      .forEach(missingDbName -> schema.applyDdl(source, missingDbName,
                                                "DROP DATABASE IF EXISTS " + quote(missingDbName),
                                                this::enqueueSchemaChanges));

// Now process all of our tables for each database ...
for (Map.Entry<String, List<TableId>> entry : tableIdsByDbName.entrySet()) {
    if (!isRunning()) break;
    String dbName = entry.getKey();
    // First drop, create, and then use the named database ...
    schema.applyDdl(source, dbName, "DROP DATABASE IF EXISTS " + quote(dbName), this::enqueueSchemaChanges);
    schema.applyDdl(source, dbName, "CREATE DATABASE " + quote(dbName), this::enqueueSchemaChanges);
    schema.applyDdl(source, dbName, "USE " + quote(dbName), this::enqueueSchemaChanges);
    for (TableId tableId : entry.getValue()) {
        if (!isRunning()) break;
        sql.set("SHOW CREATE TABLE " + quote(tableId));
        mysql.query(sql.get(), rs -> {
            if (rs.next()) {
                schema.applyDdl(source, dbName, rs.getString(2), this::enqueueSchemaChanges);
            }
        });
    }
}
context.makeRecord().regenerate();

没看懂

step 7

if (minimalBlocking && isLocked) {
    if (tableLocks) {
        // We could not acquire a global read lock and instead had to obtain individual table-level read locks
        // using 'FLUSH TABLE <tableName> WITH READ LOCK'. However, if we were to do this, the 'UNLOCK TABLES'
        // would implicitly commit our active transaction, and this would break our consistent snapshot logic.
        // Therefore, we cannot unlock the tables here!
        // https://dev.mysql.com/doc/refman/5.7/en/flush.html
        logger.info("Step {}: tables were locked explicitly, but to get a consistent snapshot we cannot "
                + "release the locks until we've read all tables.", step++);
    } else {
        // We are doing minimal blocking via a global read lock, so we should release the global read lock now.
        // All subsequent SELECT should still use the MVCC snapshot obtained when we started our transaction
        // (since we started it "...with consistent snapshot"). So, since we're only doing very simple SELECT
        // without WHERE predicates, we can release the lock now ...
        logger.info("Step {}: releasing global read lock to enable MySQL writes", step);
        sql.set("UNLOCK TABLES");
        mysql.execute(sql.get());
        isLocked = false;
        long lockReleased = clock.currentTimeInMillis();
        metrics.globalLockReleased();
        logger.info("Step {}: blocked writes to MySQL for a total of {}", step++,
                    Strings.duration(lockReleased - lockAcquired));
    }
}

解除数据库级别的全局读锁。

step 8

第8部代码很长慢慢分析；

BufferedBlockingConsumer<SourceRecord> bufferedRecordQueue = BufferedBlockingConsumer.bufferLast(super::enqueueRecord);

对enqueueRecord方法的封装，此方法内部有一个records队列，这是一个BlockingQueue，该队列的数据在kafka connect调用Task的poll方法时被读取。

protected void enqueueRecord(SourceRecord record) throws InterruptedException {
    if (record != null) {
        if (logger.isTraceEnabled()) {
            logger.trace("Enqueuing source record: {}", record);
        }
        this.records.put(record);
    }
}

Iterator<TableId> tableIdIter = tableIds.iterator();
while (tableIdIter.hasNext()) {
    ...
}

遍历tableIds所有数据，这个List里包含了所有需要全量的表。

//1.
RecordsForTable recordMaker = context.makeRecord().forTable(tableId, null, bufferedRecordQueue);
if (recordMaker != null) {

    //2.
    // Switch to the table's database ...
    sql.set("USE " + quote(tableId.catalog()) + ";");
    mysql.execute(sql.get());

    AtomicLong numRows = new AtomicLong(-1);
    AtomicReference<String> rowCountStr = new AtomicReference<>("<unknown>");
    //3.
    StatementFactory statementFactory = this::createStatementWithLargeResultSet;
    if (largeTableCount > 0) {
        try {
            // Choose how we create statements based on the # of rows.
            // This is approximate and less accurate then COUNT(*),
            // but far more efficient for large InnoDB tables.
            sql.set("SHOW TABLE STATUS LIKE '" + tableId.table() + "';");
            mysql.query(sql.get(), rs -> {
                if (rs.next()) numRows.set(rs.getLong(5));
            });
            if (numRows.get() <= largeTableCount) {
                //4.
                statementFactory = this::createStatement;
            }
            rowCountStr.set(numRows.toString());
        } catch (SQLException e) {
            // Log it, but otherwise just use large result set by default ...
            logger.debug("Error while getting number of rows in table {}: {}", tableId, e.getMessage(), e);
        }
    }

    // Scan the rows in the table ...
    long start = clock.currentTimeInMillis();
    logger.info("Step {}: - scanning table '{}' ({} of {} tables)", step, tableId, ++counter, tableIds.size());

    //5.
    Map<TableId, String> selectOverrides = getSnapshotSelectOverridesByTable();
    String selectStatement = selectOverrides.getOrDefault(tableId, "SELECT * FROM " + quote(tableId));
    logger.info("For table '{}' using select statement: '{}'", tableId, selectStatement);
    sql.set(selectStatement);

    //snapshot
    try {
        int stepNum = step;
        //6.
        mysql.query(sql.get(), statementFactory, rs -> {
            long rowNum = 0;
            try {
                // The table is included in the connector's filters, so process all of the table records
                // ...
                final Table table = schema.tableFor(tableId);
                final int numColumns = table.columns().size();
                final Object[] row = new Object[numColumns];
                while (rs.next()) {
                    for (int i = 0, j = 1; i != numColumns; ++i, ++j) {
                        Column actualColumn = table.columns().get(i);
                        row[i] = readField(rs, j, actualColumn);
                    }
                    recorder.recordRow(recordMaker, row, ts); // has no row number!
                    ++rowNum;
                    if (rowNum % 100 == 0 && !isRunning()) {
                        // We've stopped running ...
                        break;
                    }
                    if (rowNum % 10_000 == 0) {
                        long stop = clock.currentTimeInMillis();
                        logger.info("Step {}: - {} of {} rows scanned from table '{}' after {}",
                                    stepNum, rowNum, rowCountStr, tableId, Strings.duration(stop - start));
                    }
                }

                totalRowCount.addAndGet(rowNum);
                if (isRunning()) {
                    long stop = clock.currentTimeInMillis();
                    logger.info("Step {}: - Completed scanning a total of {} rows from table '{}' after {}",
                                stepNum, rowNum, tableId, Strings.duration(stop - start));
                }
            } catch (InterruptedException e) {
                Thread.interrupted();
                // We were not able to finish all rows in all tables ...
                logger.info("Step {}: Stopping the snapshot due to thread interruption", stepNum);
                interrupted.set(true);
            }
        });
    } finally {
        metrics.completeTable();
        if (interrupted.get()) break;
    }
}
++completedCounter;

1. 先获取某个表的消息生产器recordMaker；如果recordMaker不为Null，切换到该表所在的数据库。
2. 初始化statementFactory，默认是一个用于大量数据的statement构造方法。
3. 执行SHOW TABLE STATUS LIKE '" + tableId.table() + "';，获取该表的数据行数。如果得到的数据行数小于定义的大表数据量，就将statementFactory设为普通statement生产方式。
4. 构建全量SQL语句，调用getSnapshotSelectOverridesByTable方法，用户设置的带有过滤条件的语句能够在这里变处理。
5. 如果用户没有设置自己的SQL，那么执行的语句变为"SELECT * FROM " + quote(tableId) 语句。
6. 执行SQL，对于每条数据，调用recorder.recordRow(recordMaker, row, ts);方法，该方法将每条数据放入一个BlockingQueue内，也就是AbstractReader的records变量里。

step 9

if (isTxnStarted) {
    if (interrupted.get() || !isRunning()) {
        // We were interrupted or were stopped while reading the tables,
        // so roll back the transaction and return immediately ...
        logger.info("Step {}: rolling back transaction after abort", step++);
        sql.set("ROLLBACK");
        mysql.execute(sql.get());
        metrics.abortSnapshot();
        return;
    }
    // Otherwise, commit our transaction
    logger.info("Step {}: committing transaction", step++);
    sql.set("COMMIT");
    mysql.execute(sql.get());
    metrics.completeSnapshot();
}

Commit或者Rollback;

step 10

if (isLocked) {
    if (tableLocks) {
        logger.info("Step {}: releasing table read locks to enable MySQL writes", step++);
    } else {
        logger.info("Step {}: releasing global read lock to enable MySQL writes", step++);
    }
    sql.set("UNLOCK TABLES");
    mysql.execute(sql.get());
    isLocked = false;
    long lockReleased = clock.currentTimeInMillis();
    metrics.globalLockReleased();
    if (tableLocks) {
        logger.info("Writes to MySQL prevented for a total of {}", Strings.duration(lockReleased - lockAcquired));
    } else {
        logger.info("Writes to MySQL tables prevented for a total of {}", Strings.duration(lockReleased - lockAcquired));
    }
}

释放表级别的读锁。

SnapshotReader核心代码分析完成。

---

BinlogReader

这里分析BinlogReader几部分重要的代码;

BinlogReader初始化

client = new BinaryLogClient(context.hostname(), context.port(), context.username(), context.password());
...
client.registerEventListener(context.bufferSizeForBinlogReader() == 0 ? this::handleEvent : (new EventBuffer(context.bufferSizeForBinlogReader(), this))::add);
...

初始化一个Binlog处理客户端，添加对事件的处理方式。这里看handleEvent方法:

...
eventHandlers.getOrDefault(eventType, this::ignoreEvent).accept(event);
...

主要是一句，client对于事件的处理逻辑里，会调用eventHandlers的方法。eventHandlers在doStart里会注册对不同事件的处理逻辑。

EventDeserializer eventDeserializer = new EventDeserializer() {
    @Override
    public Event nextEvent(ByteArrayInputStream inputStream) throws IOException {
        try {
            // Delegate to the superclass ...
            Event event = super.nextEvent(inputStream);

            // We have to record the most recent TableMapEventData for each table number for our custom deserializers ...
            if (event.getHeader().getEventType() == EventType.TABLE_MAP) {
                TableMapEventData tableMapEvent = event.getData();
                tableMapEventByTableId.put(tableMapEvent.getTableId(), tableMapEvent);
            }
            return event;
        }
        // DBZ-217 In case an event couldn't be read we create a pseudo-event for the sake of logging
        catch(EventDataDeserializationException edde) {
            EventHeaderV4 header = new EventHeaderV4();
            header.setEventType(EventType.INCIDENT);
            header.setTimestamp(edde.getEventHeader().getTimestamp());
            header.setServerId(edde.getEventHeader().getServerId());

            if(edde.getEventHeader() instanceof EventHeaderV4) {
                header.setEventLength(((EventHeaderV4)edde.getEventHeader()).getEventLength());
                header.setNextPosition(((EventHeaderV4)edde.getEventHeader()).getNextPosition());
                header.setFlags(((EventHeaderV4)edde.getEventHeader()).getFlags());
            }

            EventData data = new EventDataDeserializationExceptionData(edde);
            return new Event(header, data);
        }
    }
};
// Add our custom deserializers ...
eventDeserializer.setEventDataDeserializer(EventType.STOP, new StopEventDataDeserializer());
eventDeserializer.setEventDataDeserializer(EventType.GTID, new GtidEventDataDeserializer());
eventDeserializer.setEventDataDeserializer(EventType.WRITE_ROWS,
                                           new RowDeserializers.WriteRowsDeserializer(tableMapEventByTableId));
eventDeserializer.setEventDataDeserializer(EventType.UPDATE_ROWS,
                                           new RowDeserializers.UpdateRowsDeserializer(tableMapEventByTableId));
eventDeserializer.setEventDataDeserializer(EventType.DELETE_ROWS,
                                           new RowDeserializers.DeleteRowsDeserializer(tableMapEventByTableId));
eventDeserializer.setEventDataDeserializer(EventType.EXT_WRITE_ROWS,
                                           new RowDeserializers.WriteRowsDeserializer(
                                                   tableMapEventByTableId).setMayContainExtraInformation(true));
eventDeserializer.setEventDataDeserializer(EventType.EXT_UPDATE_ROWS,
                                           new RowDeserializers.UpdateRowsDeserializer(
                                                   tableMapEventByTableId).setMayContainExtraInformation(true));
eventDeserializer.setEventDataDeserializer(EventType.EXT_DELETE_ROWS,
                                           new RowDeserializers.DeleteRowsDeserializer(
                                                   tableMapEventByTableId).setMayContainExtraInformation(true));
client.setEventDeserializer(eventDeserializer);

初始化client事件解析器; 对于不同的binlog事件类型，注册不同的binlog数据解析器，比如解析Date,Datetime等类型。

BinlogReader doStart方法

eventHandlers.put(EventType.STOP, this::handleServerStop);
eventHandlers.put(EventType.HEARTBEAT, this::handleServerHeartbeat);
eventHandlers.put(EventType.INCIDENT, this::handleServerIncident);
eventHandlers.put(EventType.ROTATE, this::handleRotateLogsEvent);
eventHandlers.put(EventType.TABLE_MAP, this::handleUpdateTableMetadata);
eventHandlers.put(EventType.QUERY, this::handleQueryEvent);
eventHandlers.put(EventType.WRITE_ROWS, this::handleInsert);
eventHandlers.put(EventType.UPDATE_ROWS, this::handleUpdate);
eventHandlers.put(EventType.DELETE_ROWS, this::handleDelete);
eventHandlers.put(EventType.EXT_WRITE_ROWS, this::handleInsert);
eventHandlers.put(EventType.EXT_UPDATE_ROWS, this::handleUpdate);
eventHandlers.put(EventType.EXT_DELETE_ROWS, this::handleDelete);
eventHandlers.put(EventType.VIEW_CHANGE, this::viewChange);
eventHandlers.put(EventType.XA_PREPARE, this::prepareTransaction);

根据binlog事件，注册相应的处理方法。对于delete,update,insert变更，eventHandlers分别使用handleDelete,handleUpdate,handleInsert方法来处理，这3个方法内会将数据添加到AbstractReader的BlockingQueue(records)。

总结

到这里debezium全量和增量订阅的处理逻辑已经非常清晰了。