Checkpointing under backpressure

Normally aligned checkpointing time is dominated by the synchronous and asynchronous parts of the checkpointing process. However, when a Flink job is running under heavy backpressure, the dominant factor in the end-to-end time of a checkpoint can be the time to propagate checkpoint barriers to all operators/subtasks. This is explained in the overview of the checkpointing process). and can be observed by high alignment time and start delay metrics. When this happens and becomes an issue, there are three ways to address the problem:

  1. Remove the backpressure source by optimizing the Flink job, by adjusting Flink or JVM configurations, or by scaling up.
  2. Reduce the amount of buffered in-flight data in the Flink job.
  3. Enable unaligned checkpoints.

These options are not mutually exclusive and can be combined together. This document focuses on the latter two options.

Buffer debloating

Flink 1.14 introduced a new tool to automatically control the amount of buffered in-flight data between Flink operators/subtasks. The buffer debloating mechanism can be enabled by setting the property taskmanager.network.memory.buffer-debloat.enabled to true.

This feature works with both aligned and unaligned checkpoints and can improve checkpointing times in both cases, but the effect of the debloating is most visible with aligned checkpoints. When using buffer debloating with unaligned checkpoints, the added benefit will be smaller checkpoint sizes and quicker recovery times (there will be less in-flight data to persist and recover).

For more information on how the buffer debloating feature works and how to configure it, please refer to the network memory tuning guide. Keep in mind that you can also manually reduce the amount of buffered in-flight data which is also described in the aforementioned tuning guide.

Unaligned checkpoints

Starting with Flink 1.11, checkpoints can be unaligned. Unaligned checkpoints contain in-flight data (i.e., data stored in buffers) as part of the checkpoint state, allowing checkpoint barriers to overtake these buffers. Thus, the checkpoint duration becomes independent of the current throughput as checkpoint barriers are effectively not embedded into the stream of data anymore.

You should use unaligned checkpoints if your checkpointing durations are very high due to backpressure. Then, checkpointing time becomes mostly independent of the end-to-end latency. Be aware unaligned checkpointing adds to I/O to the state storage, so you shouldn’t use it when the I/O to the state storage is actually the bottleneck during checkpointing.

In order to enable unaligned checkpoints you can:

Java

  1. StreamExecutionEnvironment env = StreamExecutionEnvironment.getExecutionEnvironment();
  2. // enables the unaligned checkpoints
  3. env.getCheckpointConfig().enableUnalignedCheckpoints();

Scala

  1. val env = StreamExecutionEnvironment.getExecutionEnvironment()
  2. // enables the unaligned checkpoints
  3. env.getCheckpointConfig.enableUnalignedCheckpoints()

Python

  1. env = StreamExecutionEnvironment.get_execution_environment()
  2. # enables the unaligned checkpoints
  3. env.get_checkpoint_config().enable_unaligned_checkpoints()

or in the flink-conf.yml configuration file:

  1. execution.checkpointing.unaligned: true

Aligned checkpoint timeout

After enabling unaligned checkpoints, you can also specify the aligned checkpoint timeout programmatically:

  1. StreamExecutionEnvironment env = StreamExecutionEnvironment.getExecutionEnvironment();
  2. env.getCheckpointConfig().setAlignedCheckpointTimeout(Duration.ofSeconds(30));

or in the flink-conf.yml configuration file:

  1. execution.checkpointing.aligned-checkpoint-timeout: 30 s

When activated, each checkpoint will still begin as an aligned checkpoint, but when the global checkpoint duration exceeds the aligned-checkpoint-timeout, if the aligned checkpoint has not completed, then the checkpoint will proceed as an unaligned checkpoint.

Limitations

Concurrent checkpoints

Flink currently does not support concurrent unaligned checkpoints. However, due to the more predictable and shorter checkpointing times, concurrent checkpoints might not be needed at all. However, savepoints can also not happen concurrently to unaligned checkpoints, so they will take slightly longer.

Interplay with watermarks

Unaligned checkpoints break with an implicit guarantee in respect to watermarks during recovery. Currently, Flink generates the watermark as the first step of recovery instead of storing the latest watermark in the operators to ease rescaling. In unaligned checkpoints, that means on recovery, Flink generates watermarks after it restores in-flight data. If your pipeline uses an operator that applies the latest watermark on each record will produce different results than for aligned checkpoints. If your operator depends on the latest watermark being always available, the workaround is to store the watermark in the operator state. In that case, watermarks should be stored per key group in a union state to support rescaling.

Interplay with long-running record processing

Despite that unaligned checkpoints barriers are able to overtake all other records in the queue. The handling of this barrier still can be delayed if the current record takes a lot of time to be processed. This situation can occur when firing many timers all at once, for example in windowed operations. Second problematic scenario might occur when system is being blocked waiting for more than one network buffer availability when processing a single input record. Flink can not interrupt processing of a single input record, and unaligned checkpoints have to wait for the currently processed record to be fully processed. This can cause problems in two scenarios. Either as a result of serialisation of a large record that doesn’t fit into single network buffer or in a flatMap operation, that produces many output records for one input record. In such scenarios back pressure can block unaligned checkpoints until all the network buffers required to process the single input record are available. It also can happen in any other situation when the processing of the single record takes a while. As result, the time of the checkpoint can be higher than expected or it can vary.

Certain data distribution patterns are not checkpointed

There are types of connections with properties that are impossible to keep with channel data stored in checkpoints. To preserve these characteristics and ensure no state corruption or unexpected behaviour, unaligned checkpoints are disabled for such connections. All other exchanges still perform unaligned checkpoints.

Pointwise connections

We currently do not have any hard guarantees on pointwise connections regarding data orderliness. However, since data was structured implicitly in the same way as any preceding source or keyby, some users relied on this behaviour to divide compute-intensive tasks into smaller chunks while depending on orderliness guarantees.

As long as the parallelism does not change, unaligned checkpoints (UC) retain these properties. With the addition of rescaling of UC that has changed.

Consider a job

Pointwise connection

If we want to rescale from parallelism p = 2 to p = 3, suddenly the records inside the keyby channels need to be divided into three channels according to the key groups. That is easily possible by using the key group ranges of the operators and a way to determine the key(group) of the record ( independent of the actual approach). For the forward channels, we lack the key context entirely. No record in the forward channel has any key group assigned; it’s also impossible to calculate it as there is no guarantee that the key is still present.

Broadcast connections

Broadcast connections bring another problem to the table. There are no guarantees that records are consumed at the same rate in all channels. This can result in some tasks applying state changes corresponding to a specific broadcasted event while others don’t, as depicted in the figure.

Broadcast connection

Broadcast partitioning is often used to implement a broadcast state which should be equal across all operators. Flink implements the broadcast state by checkpointing only a single copy of the state from subtask 0 of the stateful operator. Upon restore, we send that copy to all of the operators. Therefore it might happen that an operator will get the state with changes applied for a record that it will soon consume from its checkpointed channels.

Troubleshooting

Corrupted in-flight data

Actions described below are a last resort as they will lead to data loss.

In case of the in-flight data corrupted or by another reason when the job should be restored without the in-flight data, it is possible to use recover-without-channel-state.checkpoint-id property. This property requires to specify a checkpoint id for which in-flight data will be ignored. Do not set this property, unless a corruption inside the persisted in-flight data has lead to an otherwise unrecoverable situation. The property can be applied only after the job will be redeployed which means this operation makes sense only if externalized checkpoint is enabled.