@Override
public void run(SourceContext<OUT> ctx) throws Exception {
if (iteratorToRead == null) {
throw new IllegalStateException("Kafka iterator not initialized properly.");
}
final Object checkpointLock = ctx.getCheckpointLock();
while (running && iteratorToRead.hasNext()) {
MessageAndMetadata<byte[], byte[]> message = iteratorToRead.next();
if (lastOffsets.getState()[message.partition()] >= message.offset()) {
LOG.info(
"Skipping message with offset {} from partition {}",
message.offset(),
message.partition());
continue;
}
OUT next = deserializationSchema.deserialize(message.message());
if (deserializationSchema.isEndOfStream(next)) {
LOG.info("DeserializationSchema signaled end of stream for this source");
break;
}
// make the state update and the element emission atomic
synchronized (checkpointLock) {
lastOffsets.getState()[message.partition()] = message.offset();
ctx.collect(next);
}
if (LOG.isTraceEnabled()) {
LOG.trace(
"Processed record with offset {} from partition {}",
message.offset(),
message.partition());
}
}
}
@Override
public void open(Configuration parameters) throws Exception {
super.open(parameters);
ConsumerConnector consumer = Consumer.createJavaConsumerConnector(this.consumerConfig);
// we request only one stream per consumer instance. Kafka will make sure that each consumer
// group
// will see each message only once.
Map<String, Integer> topicCountMap = Collections.singletonMap(topicName, 1);
Map<String, List<KafkaStream<byte[], byte[]>>> streams =
consumer.createMessageStreams(topicCountMap);
if (streams.size() != 1) {
throw new RuntimeException("Expected only one message stream but got " + streams.size());
}
List<KafkaStream<byte[], byte[]>> kafkaStreams = streams.get(topicName);
if (kafkaStreams == null) {
throw new RuntimeException(
"Requested stream not available. Available streams: " + streams.toString());
}
if (kafkaStreams.size() != 1) {
throw new RuntimeException(
"Requested 1 stream from Kafka, bot got " + kafkaStreams.size() + " streams");
}
LOG.info(
"Opening Consumer instance for topic '{}' on group '{}'",
topicName,
consumerConfig.groupId());
this.iteratorToRead = kafkaStreams.get(0).iterator();
this.consumer = consumer;
zkClient =
new ZkClient(
consumerConfig.zkConnect(),
consumerConfig.zkSessionTimeoutMs(),
consumerConfig.zkConnectionTimeoutMs(),
new KafkaZKStringSerializer());
// most likely the number of offsets we're going to store here will be lower than the number of
// partitions.
int numPartitions = getNumberOfPartitions();
LOG.debug("The topic {} has {} partitions", topicName, numPartitions);
this.lastOffsets =
getRuntimeContext().getOperatorState("offset", new long[numPartitions], false);
this.commitedOffsets = new long[numPartitions];
// check if there are offsets to restore
if (!Arrays.equals(lastOffsets.getState(), new long[numPartitions])) {
if (lastOffsets.getState().length != numPartitions) {
throw new IllegalStateException(
"There are "
+ lastOffsets.getState().length
+ " offsets to restore for topic "
+ topicName
+ " but "
+ "there are only "
+ numPartitions
+ " in the topic");
}
LOG.info("Setting restored offsets {} in ZooKeeper", Arrays.toString(lastOffsets.getState()));
setOffsetsInZooKeeper(lastOffsets.getState());
} else {
// initialize empty offsets
Arrays.fill(this.lastOffsets.getState(), -1);
}
Arrays.fill(this.commitedOffsets, 0); // just to make it clear
running = true;
}