...
How it works for end-of-stream:
When an input stream is consumed to the end, Samza sends an Eos message to the control channel which includes the input topic and partition.
Once the EOS messages are received from all the partitions of this input, we know the input is end-of-stream. Then the ControlStreamConsumer will inspect the stream graph and find out the intermediate stream that all its input streams to it have been all end-of-stream. If so, we mark the intermediate stream pending end-of-stream. After that, whenever a marked intermediate stream partition reaches its highest offset (high watermark in Kafka), we can emit end-of-stream message for this partition. It’s guaranteed that the partition reaches end of stream.
Approach 2: In-band control messages
In this approach we don’t use a separate stream to keep the control messages. We use the intermediate streams themselves as both data and control.
How it works for end-of-stream:
When an input stream is consumed to the end, Samza finds out the following intermediate streams that all the inputs have been end-of-stream (through the topology of the operator graph).
The task will send an Eos message to all the partitions of the intermediate streams in 1.
Each consumer of the intermediate streams will count the watermark messages received for each partition and declare end of stream once all the EOS messages have been received.
Comparisons of the two approaches:
Pros | Cons | |
Approach 1 | - Intermediate streams are clean with only user data. This is convenient if user wants to consume it elsewhere. - Simple recovery from failure, just read the control stream from the beginning. - Less number of messages. The control messages needed is the same as the input stream partition count (n partitions). So the total will be n messages. | - Need to correlate the out-of-band control message with the source stream, which is complex to track and requires synchronization between input streams and control stream. - Need to maintain a separate stream for control messages |
Approach 2 | - No synchronization coordination needed between control message and input messages. For watermark messages, input messages can be concluded once this watermark is receivedWhen a control message is received, it is a marker that the messages sent before the control message have been consumed completely. This is critical to support general event-time watermarks. | - Complicated failure scenario of the second job. It . The consumer of control messages needs to checkpoint all the control messages received, so when it recovered from failure, it can still countresume. - More control messages required. For each intermediate stream (m partitions), we need to write to each partition task of the downstream processorproducer (n tasks) into it. So the total will be n*m messages. |
Based on the pros and cons above, we propose to use the in-band approach to support control messages.
Detail details
Intermediate Stream Message Format:
The format of the intermediate stream message:
Code Block |
---|
IntermediateMessage => [MessageType MessageData] MessageType => byte MessageData => byte[] MessageType => [0(UserMessage), 1(Watermark), 2(EndOfStream)] MessageData => [UserMessage/ControlMessage] ControlMessage => Type => int Version => int Other Message Data (based on different types of control message) |
For user message, we will use the user provided serde (default is the system serde). For control message, we will use JSON serde since it is built in Samza and easy to parse.
ControlMessage
We will support two types of ControlMessage: EndOfStreamMessage and WatermarkMessage
Code Block |
---|
public abstract class ControlMessage {
private final int version = 1;
public int getVersion() { return version; }
}
public class EndOfStreamMessage extends ControlMessage{
private final String taskName;
private final int taskCount;
private EndOfStreamMessage(String taskName, int taskCount) {
super(ControlMessageType.EndOfStream.ordinal());
this.timestamp = timestamp;
this.taskName = taskName;
this.taskCount = taskCount;
}
public long getTimestamp() { return timestamp; }
public String getTaskName() { return taskName; }
public int getTaskCount() { return taskCount; }
}
public class WatermarkMessage extends ControlMessage{
private final long timestamp;
private final String taskName;
private final int taskCount;
private WatermarkMessage(long timestamp, String taskName, int taskCount) {
super(ControlMessageType.Watermark.ordinal());
this.timestamp = timestamp;
this.taskName = taskName;
this.taskCount = taskCount;
}
public long getTimestamp() { return timestamp; }
public String getTaskName() { return taskName; }
public int getTaskCount() { return taskCount; }
} |
Reconciliation
When SystemConsumers gets EOS/watermark messages, Samza needs to reconcile based on the task counts. The reconciliation works as follows:
For each intermediate stream partition, Samza keeps track of the end-of-stream/watermark messages received from the producing tasks, and counts the number of tasks that it has been received in the messages.
When the count matches the total task count, Samza will emit a end-of-stream/watermark message to the task that’s assigned for this stream partition.
When Samza received further watermark messages, it will emit a watermark with the earliest event time across all the stream partitions. No emission if the earliest event time doesn’t change.
Checkpoint control messages
For failure scenario, the latest control message received from each intermediate stream partition will be lost without checkpointing. The checkpoint of control messages need to preserve both intermediate stream partition and the producing task information. A checkpoint will be:
Code Block |
---|
Key => IntermediateStreamPartition.ControlMessageType Value => ControlMessageCheckpoint public class ControlMessageCheckpoint { int taskCount; Map<String, Long> tasksToEventTime; } |
...