Step 1:
The example below reads from a Kafka topic with two partitions that each contains “A”, “B”, “C”, ”D”, “E” as messages. We set the offset to zero for both partitions.
Step 2:
In the second step, the Kafka consumer starts reading messages from partition 0. Message “A” is processed “in-flight” and the offset of the first consumer is changed to 1.
Step 3:
In the third step, message “A” arrives at the Flink Map Task. Both consumers read their next records (message “B” for partition 0 and message “A” for partition 1). The offsets are updated to 2 and 1 respectively for both partitions. At the same time, Flink’s Job Master decides to trigger a checkpoint at the source.
Step 4:
In the following step, the Kafka consumer tasks have already created a snapshot of their states (“offset = 2, 1”) which is now stored in Apache Flink’s Job Master. The sources emit a checkpoint barrier after messages “B” and “A” from partitions 0 and 1 respectively. The checkpoint barriers are used to align the checkpoints of all operator tasks and guarantee the consistency of the overall checkpoint. Message “A” arrives at the Flink Map Task while the top consumer continues reading its next record (message “C”).
Step 5:
This step shows that the Flink Map Task receives the checkpoint barriers from both sources and checkpoints its state to the Job Master. Meanwhile, the consumers continue reading more events from the Kafka partitions.
Step 6:
This step shows that the Flink Map Task communicates to Flink Job Master once it checkpointed its state. When all tasks of a job acknowledge that their state is checkpointed, the Job Master completes the checkpoint. From now on, the checkpoint can be used to recover from a failure. It’s worth mentioning here that Apache Flink does not rely on the Kafka offsets for restoring from potential system failures.
Recovery in case of a failure
In case of a failure (for instance, a worker failure) all operator tasks are restarted and their state is reset to the last completed checkpoint. As it is depicted in the following illustration.
The Kafka sources start from offset 2 and 1 respectively as this was the offset of the completed checkpoint. When the job is restarted we can expect a normal system operation as if no failure occurred before.
网友评论