回复：Flink 1.10 的 JDBCUpsertOutputFormat flush方法的重试机制无法生效

[lucas.wu]

hi ,我们之前写mysql遇到过这种问题，因为任务的open函数是任务启动的时候调用的，是只调用一次的。而open函数里面会对jdbc connection进行初始化，当jdbc conection因为各种原因断开的时候，例如空闲时间超过max_idel_timeout。这都会导致flush失败，进而导致整个task重启。所以我们后面参照官方的这个JDBCUpsertOutputFormat 自己写了一个ouputFormat，加上了连接的检查和重连，如果对一致性要求不高的话，还可以对flush进行异常捕捉。


原始邮件
发件人:[hidden email]
收件人:[hidden email]
抄送:[hidden email]
发送时间:2020年3月23日(周一) 11:05
主题:Flink 1.10 的 JDBCUpsertOutputFormat flush方法的重试机制无法生效


我们在测试环境测试JDBC写入postgresql的场景，用tcpkill模拟链接被关闭的情况，测试对异常的兼容性，我们发现一直打印类似的日志 2020-03-20 21:16:21.247 [jdbc-upsert-output-format-thread-1] ERROR org.apache.flink.api.java.io.jdbc.JDBCUpsertOutputFormatnbsp; - JDBC executeBatch error, retry times = 1 org.postgresql.util.PSQLException: This connection has been closed. at org.postgresql.jdbc.PgConnection.checkClosed(PgConnection.java:857) at org.postgresql.jdbc.PgConnection.getAutoCommit(PgConnection.java:817) at org.postgresql.jdbc.PgStatement.internalExecuteBatch(PgStatement.java:813) at org.postgresql.jdbc.PgStatement.executeBatch(PgStatement.java:873) at org.postgresql.jdbc.PgPreparedStatement.executeBatch(PgPreparedStatement.java:1569) at org.apache.flink.api.java.io.jdbc.writer.AppendOnlyWriter.executeBatch(AppendOnlyWriter.java:62) at org.apache.flink.api.java.io.jdbc.JDBCUpsertOutputFormat.flush(JDBCUpsertOutputFormat.java:159) at org.apache.flink.api.java.io.jdbc.JDBCUpsertOutputFormat.lambda$open$0(JDBCUpsertOutputFormat.java:124) at java.util.concurrent.Executors$RunnableAdapter.call(Executors.java:511) at java.util.concurrent.FutureTask.runAndReset(FutureTask.java:308) at java.util.concurrent.ScheduledThreadPoolExecutor$ScheduledFutureTask.access$301(ScheduledThreadPoolExecutor.java:180) at java.util.concurrent.ScheduledThreadPoolExecutor$ScheduledFutureTask.run(ScheduledThreadPoolExecutor.java:294) at java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1149) at java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:624) at java.lang.Thread.run(Thread.java:748) 从日志上看来，链接被关闭，只会导致重试一次，这里是Flink的重试代码 //JDBCUpsertOutputFormat.javapublic synchronized void flush() throws Exception { checkFlushException(); for (int i = 1; i = maxRetryTimes; i++) { try { jdbcWriter.executeBatch(); batchCount = 0; break; } catch (SQLException e) { LOG.error("JDBC executeBatch error, retry times = {}", i, e); if (i gt;= maxRetryTimes) { throw e; } Thread.sleep(1000 * i); } } } 通过远程debug分析，在第一次执行 JDBCUpsertOutputFormat.flush nbsp; -gt; AppendOnlyWriter.executeBatch nbsp; nbsp; nbsp;... nbsp; nbsp; nbsp;-gt; PgConnection.getAutoCommit 抛出PSQLException: This connection has been closed时，batchStatements在这之前已经被清空 // PgStatement.java private BatchResultHandler internalExecuteBatch() throws SQLException { // Construct query/parameter arrays. transformQueriesAndParameters(); // Empty arrays should be passed to toArray // see http://shipilev.net/blog/2016/arrays-wisdom-ancients/ Query[] queries = batchStatements.toArray(new Query[0]); ParameterList[] parameterLists = batchParameters.toArray(new ParameterList[0]); batchStatements.clear(); // 这里已经被清空 batchParameters.clear(); ... if (connection.getAutoCommit()) { // 抛出异常 flags |= QueryExecutor.QUERY_SUPPRESS_BEGIN; } ... } 所以在Flink第二次重试执行这段代码的时候jdbcWriter.executeBatch的时候，batchStatements已经为Empty并直接返回了，导致Flink认为是成功的，这样会导致其实语句没有被执行？ // PgStatement.java public int[] executeBatch() throws SQLException { checkClosed(); closeForNextExecution(); if (batchStatements == null || batchStatements.isEmpty()) { //这里就直接返回了 return new int[0]; } return internalExecuteBatch().getUpdateCount(); } 目前的想法是，当我出现异常捕获的时候，在重试之前，判断这个链接是否已经被关闭，如果已经被关闭，则重新open，不知道大家有啥看法，这是我提的issue https://issues.apache.org/jira/browse/FLINK-16708