This document describes Flink’s fault tolerance mechanism for streaming data flows.
Apache Flink offers a fault tolerance mechanism to consistently recover the state of data streaming applications. The mechanism ensures that even in the presence of failures, the program’s state will eventually reflect every record from the data stream exactly once. Note that there is a switch to downgrade the guarantees to at least once (described below).
The fault tolerance mechanism continuously draws snapshots of the distributed streaming data flow. For streaming applications with small state, these snapshots are very light-weight and can be drawn frequently without impacting the performance much. The state of the streaming applications is stored at a configurable place (such as the master node, or HDFS).
In case of a program failure (due to machine-, network-, or software failure), Flink stops the distributed streaming dataflow. The system then restarts the operators and resets them to the latest successful checkpoint. The input streams are reset to the point of the state snapshot. Any records that are processed as part of the restarted parallel dataflow are guaranteed to not have been part of the checkpointed state before.
Note: For this mechanism to realize its full guarantees, the data stream source (such as message queue or broker) needs to be able to rewind the stream to a defined recent point. Apache Kafka has this ability and Flink’s connector to Kafka exploits this ability.
Note: Because Flink’s checkpoints are realized through distributed snapshots, we use the words snapshot and checkpoint interchangeably.
The central part of Flink’s fault tolerance mechanism is drawing consistent snapshots of the distributed data stream and operator state. These snapshots act as consistent checkpoints to which the system can fall back in case of a failure. Flink’s mechanism for drawing these snapshots is described in “Lightweight Asynchronous Snapshots for Distributed Dataflows”. It is inspired by the standard Chandy-Lamport algorithm for distributed snapshots and is specifically tailored to Flink’s execution model.
A core element in Flink’s distributed snapshotting are the stream barriers. These barriers are injected into the data stream and flow with the records as part of the data stream. Barriers never overtake records, the flow strictly in line. A barrier separates the records in the data stream into the set of records that goes into the current snapshot, and the records that go into the next snapshot. Each barrier carries the ID of the snapshot whose records it pushed in front of it. Barriers do not interrupt the flow of the stream and are hence very lightweight. Multiple barriers from different snapshots can be in the stream at the same time, which means that various snapshots may happen concurrently.
Stream barriers are injected into the parallel data flow at the stream sources. The point where the barriers for snapshot n are injected (let’s call it Sn) is the position in the source stream up to which the snapshot covers the data. For example, in Apache Kafka, this position would be the last record’s offset in the partition. This position Sn is reported to the checkpoint coordinator (Flink’s JobManager).
The barriers then flow downstream. When an intermediate operator has received a barrier for snapshot n from all of its input streams, it emits itself a barrier for snapshot n into all of its outgoing streams. Once a sink operator (the end of a streaming DAG) has received the barrier n from all of its input streams, it acknowledges that snapshot n to the checkpoint coordinator. After all sinks acknowledged a snapshot, it is considered completed.
When snapshot n is completed, it is certain that no records from before Sn will be needed any more from the source, because these records (and their descendant records) have passed through the entire data flow topology.
Operators that receive more than one input stream need to align the input streams on the snapshot barriers. The figure above illustrates this:
When operators contain any form of state, this state must be part of the snapshots as well. Operator state comes in different forms:
filter()). User-defined state can either be a simple variable in the function’s java object, or the associated key/value state of a function (see State in Streaming Applications for details).
Operators snapshot their state at the point in time when they received all snapshot barriers from their input streams, before emitting the barriers to their output streams. At that point, all updates to the state from records before the barriers will have been made, and no updates that depend on records from after the barriers have been applied. Because the state of a snapshot may be potentially large, it is stored in a configurable state backend. By default, this is the JobManager’s memory, but for serious setups, a distributed reliable storage should be configured (such as HDFS). After the state has been stored, the operator acknowledges the checkpoint, emits the snapshot barrier into the output streams, and proceeds.
The resulting snapshot now contains:
The alignment step may add latency to the streaming program. Usually, this extra latency is in the order of a few milliseconds, but we have seen cases where the latency of some outliers increased noticeably. For applications that require consistently super low latencies (few milliseconds) for all records, Flink has a switch to skip the stream alignment during a checkpoint. Checkpoint snapshots are still drawn as soon as an operator has seen the checkpoint barrier from each input.
When the alignment is skipped, an operator keeps processing all inputs, even after some checkpoint barriers for checkpoint n arrived. That way, the operator also processes elements that belong to checkpoint n+1 before the state snapshot for checkpoint n was taken. On a restore, these records will occur as duplicates, because they are both included in the state snapshot of checkpoint n, and will be replayed as part of the data after checkpoint n.
NOTE: Alignment happens only for operators with multiple predecessors (joins) as well as operators with multiple senders (after a stream repartitioning/shuffle).
Because of that, dataflows with only embarrassingly parallel streaming operations (
filter(), …) actually give exactly once guarantees even
in at least once mode.
Recovery under this mechanism is straightforward: Upon a failure, Flink selects the latest completed checkpoint k. The system then re-deploys the entire distributed dataflow, and gives each operator the state that was snapshotted as part of checkpoint k. The sources are set to start reading the stream from position Sk. For example in Apache Kafka, that means telling the consumer to start fetching from offset Sk.
If state was snapshotted incrementally, the operators start with the state of the latest full snapshot and then apply a series of incremental snapshot updates to that state.