...

...

...

...

...

JIRA: here (<- link to https://issues.apache.org/jira/browse/FLINK-XXXX)

...

Please keep the discussion on the mailing list rather than commenting on the wiki (wiki discussions get unwieldy fast).

...

Figure 1: Checkpoint barriers overtake in-flight records. Colored buffers need to be persisted together with the operator state.

Preliminary work

We first evaluated with POCs two options when to persist:

• Persist in an adhoc fashion on checkpoint. While keeping the data volume to a minimum this approach could potentially overload external systems during checkpointing.
• Persist continuously to avoid such overload on the cost of storing much more data and potentially decreasing throughput.

In the following experiments, we had a simple topology (source → map → map → sleepy map → measure map), where each channel was a random shuffle. The sleepy map slept on average 0, 0.01, and 0.1 ms per record to induce backpressure. We compared a POC for adhoc spilling, a POC for continuous spilling, and the base commit in the master (1.11-SNAPSHOT). For each approach and sleep time, we measured 3 times and took the median while persisting on HDFS in an EMR cluster with 1 master and 5 slaves, each having 32 cores and 256 GB RAM, with a total parallelism of 160.

Image AddedImage Added

We can see that for the baseline, the checkpoint duration strongly increases with sleep time and thus backpressure. For both POCs, the checkpoint duration, however, remained stable and even decreases for higher backpressure because of lower overall data volume. Surprisingly, the checkpointing times for continuous spilling are higher than adhoc for lower sleep times. We suspect that we have additional backpressure for continuously write large amounts of data. With increasing backpressure and thus decreasing data volume, continuous spilling reaches sub-seconds checkpointing times.

Nevertheless, continuous spilling seems to have rather big impact of overall throughput. While adhoc POC showed 10% performance decrease, continuous POC is clearly bottlenecked for higher volume. Primarily for that reason, we decided to go with the adhoc approach in this FLIP. While sub-seconds checkpointing times of continuous spilling surely would be a nice asset, the primary goal of decoupling checkpointing times from backpressure is also reached with adhoc spilling.

Proposed Changes

In this FLIP, we

• Change the handling of checkpoint barriers, to allow them to overtake other records,
• Enhance the checkpoint formatAdd persist the inflight data inside of checkpoints, 
• Recover from the new checkpoint while also allowing new checkpoints to be taken during recovering to guarantee progress, and
• Propose a solution of how the operators can be rescaled in the future

...

Note, that during downscaling the size limit of “max inflight data” might be temporarily exceeded during recovery (see recovery section).

Key decisions

1. snapshot channel state when a checkpoint arrives (as opposed to continuously writing all output buffers): POC showed it's more efficient and has lower latency
2. use the existing checkpointing mechanism (see Persistence)
3. restore upstream/downstream parts of a channel in the corresponding tasks (as opposed to restoring the whole channel in the downstream task): less coupling and easier to implement if use the existing checkpointing mechanism

Recovery

...

Persistence

For persistence, existing state management primitives will be reused to colocate operator state and inflight data, which offers the following advantages.

• Don’t duplicate code and avoid inconsistencies between channel and operator state distribution (especially keyed)
• Simplicity on recovery and rescaling: having state not split into channel and operator parts
• Avoid inconsistencies between operator and channel state (e.g. different retention times)
• Being able to use existing snapshotting mechanism with the possibility to reuse incremental checkpoints

However, some drawbacks may require to use alternatives ways in the future.

• Less flexibility
• Risk of break snapshotting
• Increased checkpoints size

The checkpoint format is only implicitly extended by adding more (keyed) state with conventional naming.

Components

In general, inflight data is stored in state handles per operator sub task that are ultimately managed by CheckpointCoordinator. We need to add or modify the following components.

1. Checkpoint Metadata
1. Channel StateHandle extends StreamStateHandle, contains subtask index and information to place it to the correct subpartition/inputchannel
2. TaskStateSnapshot, PendingCheckpoint, and CompletedCheckpoint - contains a collection of ChannelStateHandle
2. Writer. Writes network buffers using a provided CheckpointStreamFactory and returns state handles
3. Reader. Reads data from state handle and returns buffers (loads and deserializes)
4. Buffer (de)serializer
5. StateAssignmentOperation (existing class) - reads snapshot metadata and distributes state handles across subtasks; need to add channel states distribution

TaskStateManagerImpl (existing class) - holds state assigned to task

Recovery

From a high level perspective, inflight data is restored where it was previously spilled:

• Outgoing buffers are restored on upstream side into output.
• Incoming buffers are restored on downstream side into the input.

This approach facilitates an easy integration into existing checkpointing mechanism. Restored buffers take precedence over all newly produced/received buffers. The following steps explain the assignment of state and mapping it to in-memory data structures (without rescaling) in more detail:

• Channel state consists of upstream and downstream parts; they are recovered by upstream/downstream subtasks separately
• Each part (InputChannel/SubPartition) is represented as StateHandle and included in the snapshot
• Each part in metadata has subtask id and target or source subtask id
• On recovery, checkpoint coordinator distributes state according to subtask index
1. Without rescaling, we don’t need to do anything to colocate operator and channel state (state placement isn't changed)
2. Raw state is distributed in the same manner
• Subtasks place state into proper InputChannel/SubPartitions according to target/source id

Rescaling

Since we hook into existing checkpointing mechanism, most of the rescaling challenges are solved in the same way.

Non-keyed

To place a state into a task CheckpointCoordinator uses modulo operation. For example, when scaling out:

...

• i.e., if we have same some channel state on upstream, and channel state on downstream corresponds to it (as ensured by step 4); then it’s enough to sort them by channel id ([src_id : dst_id]); this will ensure the order of records
• there is a (technical?) problem with the proposed solution to match multi-buffer records: we need to alternate load channel and process network operations which can be tricky
• another solution would be to have temporary "virtual" channels to process data from "imported" channels

Keyed

For each key group, we need to place channel and operator state on the same subtask. For output buffers, we know (at least conceptually) the keys, so we can store state as KeyedStateHandle and then reuse operator state redistribution code.For input buffers, we only know a set of groups - the groups assigned to this task.

But, given that placement of key groups among the tasks is deterministic, we can: 1) compute old and new placements and 2) give each new task input buffers that could contain it’s key groups. The task then must filter out irrelevant records by keys (after deserialization). For regular (single-buffer) records it’s trivial.

For multi-buffer record, if it doesn’t belong to this subtask, it will never receive remaining buffers and therefore can’t deserialize and filter it out. Possible solutions:

1. Send restored records from upstream to all downstreams that could need it (not according to key); we can use the same logic as in distribution for that
2. Use sequence numbers: if we expect next part of a multi-buffer record on this channel, but receive a buffer with a different SN; then discard this stored buffer
3. Write key in the beginning of each buffer (need to ensure the key itself isn’t splitted)

Order of processing restored buffers

...

Guaranteed progress

To guarantee progress even during recovery, we need to implement three features:

...

Thus, in an environment with frequent crashes (under backpressure), progress can be made as soon as the operators themselves are recovered fast enough.

The first, state-based storage will not satisfy all requirements. Nevertheless, assuming that the in-flight data is rather small, we should still quickly come to the point where new output is produced.

Persistence

For persistence, existing mechanisms will be reused. The benefits are:

Including channel state into snapshot:

1. Don’t duplicate re-distribution code and avoid inconsistencies between channel and operator state distribution (especially keyed)
2. Simplicity: having state not split into channel and operator parts
3. Being able to use existing snapshotting mechanism (below)

Reading/writing using the existing mechanisms (versus custom storage):

1. Don’t duplicate code for provider-agnostic storage, configuration, local recovery
2. Ease deployment by having single storage for operator and channel state
3. Avoid inconsistencies between operator and channel state (e.g. different retention times)
4. Possibility to reuse incremental checkpoints; other alternatives avoid doubly-storing buffers differently, but probably at the cost of more complex cleanup

Disadvantages

1. Less flexibility
2. Risk of break snapshotting
3. Increased checkpoints size

Components

1. Checkpoint Metadata
1. Channel StateHandle extends StreamStateHandle, contains subtask index and information to place it to the correct subpartition/inputchannel
2. TaskStateSnapshot, PendingCheckpoint, and CompletedCheckpoint - contains a collection of ChannelStateHandle
2. Writer. Writes network buffers using a provided CheckpointStreamFactory and returns state handles
3. Reader. Reads data from state handle and returns buffers (loads and deserializes)
4. (de)Serializer
5. StateAssignmentOperation (existing class) - reads snapshot metadata and distributes state handles across subtasks; need to add channel states distribution

TaskStateManagerImpl (existing class) - holds state assigned to task

Efficiency

The following optimizations won't necessarily be implemented in MVP:

1. not writing the same buffer multiple times (when the job is backpressured); can be addressed by implementing incremental checkpointing for FS backend
2. incremental loading and processing of state
3. no additional memory to load channel state: ideally, existing network buffers should be reused
4. reduced number of files: single file could be reused for several checkpoints 

Compatibility, Deprecation, and Migration Plan

...

Compatibility, Deprecation, and Migration Plan

Unaligned checkpoints will initially be an optional feature. After collecting experience and implementing all necessary extensions, unaligned checkpoint will probably be enabled by default for exactly once.

For compatibility, in

...

documentation, clearly state that users of the checkpoint API can only use unaligned checkpoints if they do not require a consistent state across all operators.

Known Limitations

• State size increase
• Up to a couple of GB per task (especially painful on IO bound clusters)
• Depends on the actual policy used (probably UNALIGNED_WITH_MAX_INFLIGHT_DATA  is the more plausible default)
• Longer and heavier recovery depending on the increased state size
• Can potentially trigger death spiral after a first failure
• More up-to-date checkpoints will most likely still be an improvement about the current checkpoint behavior during backpressure
For the MVP
• no re-scaling
max in-flight checkpoints = 1

Test Plan

• Correctness tests with induced failures
• Compare checkpoints times under backpressure with current state
• Compare throughput under backpressure with current state
• Compare progress under backpressure with frequently induced failures
• Compare throughput with no checkpointing

Roadmap

There are two primary goals. First, we want to provide a minimal viable product (MVP) that breaks up the vicious cycle of overload and instabilities of a cluster with slow checkpoints and more accumulated load resulting from recoveries from outdated checkpoints. Second, we aim for the full FLIP implementation that will among other things allow rescaling from unaligned checkpoints to directly counter overload.

In particular, the following improvements will be achieved through full release over MVP:

• Re-scaling on unaligned checkpoints (need to use savepoint in MVP)
  
• Incremental checkpointing to not write the same buffer multiple times (when the job is backpressured)
• Advanced triggers for unaligned checkpoints such as timeouts on alignment or meeting the maximum threshold of checkpoint sizes
• Support for concurrent checkpoints
• Incremental loading and processing of state
• No additional memory to load channel state: ideally, existing network buffers should be reused
• Reduced number of files: single file could be reused for several checkpoints

We aim to make unaligned checkpoints the default behavior after the full implementation.