1 背景
1.使用 flink-1.16 的 table sql 消费 kafka数据,并使用 sql 计算指标,然后写入 doris;
2.指标计算时,需要统计当日数据条数,考虑到作业异常退出被重新拉起时,需要从零点开始消费,所以指定 'scan.startup.mode' = 'timestamp','scan.startup.timestamp-millis' = '当日零点时间戳' 方式创建 kafka table:
s"""
|CREATE TABLE qysfxjKafkaTable (
|xlid STRING,
|available_status STRING,
|sendtime STRING,
|`ts` TIMESTAMP(3) METADATA FROM 'timestamp'
|) WITH (
|'connector' = 'kafka',
|'topic' = '${param.getProperty("qysfxjTopic")}',
|'scan.startup.mode' = 'timestamp','scan.startup.timestamp-millis' = '当日零点时间戳'
|'properties.group.id' = '${param.getProperty("qysfxjTopicGroupId")}',
|'properties.bootstrap.servers' = '${param.getProperty("brokers")}',
|'properties.auto.offset.reset' = 'earliest',
|'json.ignore-parse-errors' = 'true',
|'json.fail-on-missing-field' = 'false',
|'format' = 'json')
|""".stripMargin
3.启动时报 kakfa 的错误,Invalid negative offset,即 flink 使用了一个不正确的 offset 到 kafka 消费数据,经过排查 topic 中最新一条数据的时间,在今日零点之前,也就是说,kafka table sql 中指定今日零点的时间戳,落后于 kafka 最新数据的时间;
2 解决方案
1.两种解决方案,① 从检查点启动作业;② 根据 kafka 数据时间,调整消费的时间;考虑到第一次启动可能 topic 也没有数据,且如果检查点失败会导致作业无法从检查点恢复的情况,决定采用 ② 方案解决;
2.方案步骤
1.使用 kafka java api,获取 topic 中最后一条数据,根据数据的时间戳初始化创建 kafka table sql 的启动时间;
2.获取到 kafka 最后一条数据的场景有两种:① kafka 中最新一条数据时间早于零点(报错的场景);② kafka 中最新一条数据时间晚于零点;
3.根据以上步骤,实现代码,代码会返回一个时间戳,0或者最后一条数据时间戳:
def getTopicLatestRecordTimeStamp(brokers: String,topic: String): Long ={
var retMillis = 0L
val props = new Properties();
props.put("bootstrap.servers", brokers);
props.put("group.id", "test");
props.put("key.deserializer", "org.apache.kafka.common.serialization.StringDeserializer");
props.put("value.deserializer", "org.apache.kafka.common.serialization.StringDeserializer");
val consumer = new KafkaConsumer[String,String](props);
try {
// 订阅 topic
consumer.subscribe(Collections.singletonList(topic))
// 获取并订阅全部分区
var assigment = consumer.assignment()
while (assigment.size() == 0) {
consumer.poll(Duration.ofMillis(1000L))
assigment = consumer.assignment()
}
println(assigment)
val earliestOffset = consumer.beginningOffsets(assigment)
val lastOffset = consumer.endOffsets(assigment)
println("assigment: " + assigment + ",topic earliestOffset:" + earliestOffset + ",topic endOffsets:" + lastOffset)
// 遍历所有分区,获取最新的 offset
val lastOffsetIter = lastOffset.entrySet().iterator()
while (lastOffsetIter.hasNext){
val ele = lastOffsetIter.next()
if(ele.getValue != 0L){
// 分区有数据
consumer.seek(ele.getKey, ele.getValue - 1)
val records = consumer.poll(1000).iterator()
while (records.hasNext){
val next = records.next()
if(next.timestamp() > retMillis){
retMillis = next.timestamp()
}
System.out.println("Timestamp of last record: " + next.timestamp())
}
}
}
retMillis
} finally {
consumer.close();
}
}
4.根据返回的时间戳,于当日零点判断,如果返回的时间戳早于零点,使用 latest-offset,返回的时间戳晚于当日零点,使用零点启动即可,以下代码返回使用的是时间戳启动,还是 latest-offset 启动:
if(parameterTool.get("qysfxjTopicStartFrom","latest").equals("latest")){
val toAssignmentTime = TimeHandler.getMidNightMillions()
val latestTime = KfkUtil.getTopicLatestRecordTimeStamp(pro.get("brokers").toString,pro.get("qysfxjTopic").toString)
// 如果最后一条数据在 toAssignmentTime 之前,则使用 latest-offset 启动消费
if(toAssignmentTime > latestTime){
pro.put("qysfxjTopicStartFrom","latest-offset")
}else {
pro.put("qysfxjTopicStartFrom",(toAssignmentTime).toString)
}
}else {
pro.put("qysfxjTopicStartFrom",parameterTool.get("qysfxjTopicStartFrom"))
}
5.根据时间戳,还是 latest-offset,生成 sql 中的 scan 片段:
/**
* 根据 startFrom,判断是从什么位置消费。
*
* @param startFrom:earliest-offset,latest-offset,group-offsets,timestamp
* @return
*/
def getKafkaSQLScanStr(startFrom: String): String = {
var scanStartup = ""
if(Character.isDigit(startFrom.trim()(0))){
scanStartup =
"'scan.startup.mode' = 'timestamp'," +
s"'scan.startup.timestamp-millis' = '${startFrom.trim()}',"
}else {
scanStartup =
s"'scan.startup.mode' = '${startFrom}',"
}
scanStartup
}
6.完整table sql 拼接:
val qysfxjKafkaSource =
s"""
|CREATE TABLE qysfxjKafkaTable (
|xlid STRING,
|available_status STRING,
|sendtime STRING,
|`ts` TIMESTAMP(3) METADATA FROM 'timestamp'
|) WITH (
|'connector' = 'kafka',
|'topic' = '${param.getProperty("qysfxjTopic")}',
|${TXGJUtils.getKafkaSQLScanStr(qysfxjTopicStartFrom)}
|'properties.group.id' = '${param.getProperty("qysfxjTopicGroupId")}',
|'properties.bootstrap.servers' = '${param.getProperty("brokers")}',
|'properties.auto.offset.reset' = 'earliest',
|'json.ignore-parse-errors' = 'true',
|'json.fail-on-missing-field' = 'false',
|'format' = 'json')
|""".stripMargin