Flink DataStream programs are typically designed to run for long periods of time such as weeks, months, or even years. As with all long-running services, Flink streaming applications need to be maintained, which includes fixing bugs, implementing improvements, or migrating an application to a Flink cluster of a later version.
This document describes how to update a Flink streaming application and how to migrate a running streaming application to a different Flink cluster.
The line of action for upgrading a streaming application or migrating an application to a different cluster is based on Flink’s Savepoint feature. A savepoint is a consistent snapshot of the state of an application at a specific point in time.
There are two ways of taking a savepoint from a running streaming application.
It is recommended to periodically take savepoints in order to be able to restart an application from a previous point in time.
This means that the application is canceled immediately after the savepoint completed, i.e., no other checkpoints are taken after the savepoint.
Given a savepoint taken from an application, the same or a compatible application (see Application State Compatibility section below) can be started from that savepoint. Starting an application from a savepoint means that the state of its operators is initialized with the operator state persisted in the savepoint. This is done by starting an application using a savepoint.
The operators of the started application are initialized with the operator state of the original application (i.e., the application the savepoint was taken from) at the time when the savepoint was taken. The started application continues processing from exactly this point on.
Note: Even though Flink consistently restores the state of an application, it cannot revert writes to external systems. This can be an issue if you resume from a savepoint that was taken without stopping the application. In this case, the application has probably emitted data after the savepoint was taken. The restarted application might (depending on whether you changed the application logic or not) emit the same data again. The exact effect of this behavior can be very different depending on the
SinkFunction and storage system. Data that is emitted twice might be OK in case of idempotent writes to a key-value store like Cassandra but problematic in case of appends to a durable log such as Kafka. In any case, you should carefully check and test the behavior of a restarted application.
When upgrading an application in order to fix a bug or to improve the application, usually the goal is to replace the application logic of the running application while preserving its state. We do this by starting the upgraded application from a savepoint which was taken from the original application. However, this does only work if both applications are state compatible, meaning that the operators of upgraded application are able to initialize their state with the state of the operators of original application.
In this section, we discuss how applications can be modified to remain state compatible.
When an application is restarted from a savepoint, Flink matches the operator state stored in the savepoint to stateful operators of the started application. The matching is done based on operator IDs, which are also stored in the savepoint. Each operator has a default ID that is derived from the operator’s position in the application’s operator topology. Hence, an unmodified application can always be restarted from one of its own savepoints. However, the default IDs of operators are likely to change if an application is modified. Therefore, modified applications can only be started from a savepoint if the operator IDs have been explicitly specified. Assigning IDs to operators is very simple and done using the
uid(String) method as follows:
Note: Since the operator IDs stored in a savepoint and IDs of operators in the application to start must be equal, it is highly recommended to assign unique IDs to all operators of an application that might be upgraded in the future. This advice applies to all operators, i.e., operators with and without explicitly declared operator state, because some operators have internal state that is not visible to the user. Upgrading an application without assigned operator IDs is significantly more difficult and may only be possible via a low-level workaround using the
Important: As of 1.3.x this also applies to operators that are part of a chain.
By default all state stored in a savepoint must be matched to the operators of a starting application. However, users can explicitly agree to skip (and thereby discard) state that cannot be matched to an operator when starting a application from a savepoint. Stateful operators for which no state is found in the savepoint are initialized with their default state. Users may enforce best practices by calling
ExecutionConfig#disableAutoGeneratedUIDs which will fail the job submission if any operator does not contain a custom unique ID.
When upgrading an application, user functions and operators can be freely modified with one restriction. It is not possible to change the data type of the state of an operator. This is important because, state from a savepoint can (currently) not be converted into a different data type before it is loaded into an operator. Hence, changing the data type of operator state when upgrading an application breaks application state consistency and prevents the upgraded application from being restarted from the savepoint.
Operator state can be either user-defined or internal.
User-defined operator state: In functions with user-defined operator state the type of the state is explicitly defined by the user. Although it is not possible to change the data type of operator state, a workaround to overcome this limitation can be to define a second state with a different data type and to implement logic to migrate the state from the original state into the new state. This approach requires a good migration strategy and a solid understanding of the behavior of key-partitioned state.
Internal operator state: Operators such as window or join operators hold internal operator state which is not exposed to the user. For these operators the data type of the internal state depends on the input or output type of the operator. Consequently, changing the respective input or output type breaks application state consistency and prevents an upgrade. The following table lists operators with internal state and shows how the state data type relates to their input and output types. For operators which are applied on a keyed stream, the key type (KEY) is always part of the state data type as well.
|Operator||Data Type of Internal Operator State|
|ReduceFunction[IOT]||IOT (Input and output type) [, KEY]|
|FoldFunction[IT, OT]||OT (Output type) [, KEY]|
|WindowFunction[IT, OT, KEY, WINDOW]||IT (Input type), KEY|
|AllWindowFunction[IT, OT, WINDOW]||IT (Input type)|
|JoinFunction[IT1, IT2, OT]||IT1, IT2 (Type of 1. and 2. input), KEY|
|CoGroupFunction[IT1, IT2, OT]||IT1, IT2 (Type of 1. and 2. input), KEY|
|Built-in Aggregations (sum, min, max, minBy, maxBy)||Input Type [, KEY]|
Besides changing the logic of one or more existing operators, applications can be upgraded by changing the topology of the application, i.e., by adding or removing operators, changing the parallelism of an operator, or modifying the operator chaining behavior.
When upgrading an application by changing its topology, a few things need to be considered in order to preserve application state consistency.
This section describes the general way of upgrading Flink across versions and migrating your jobs between the versions.
In a nutshell, this procedure consists of 2 fundamental steps:
Besides those two fundamental steps, some additional steps can be required that depend on the way you want to change the Flink version. In this guide we differentiate two approaches to upgrade across Flink versions: in-place upgrade and shadow copy upgrade.
For in-place update, after taking savepoints, you need to:
For shadow copy, you need to:
In the following, we will first present the preconditions for successful job migration and then go into more detail about the steps that we outlined before.
Before starting the migration, please check that the jobs you are trying to migrate are following the best practices for savepoints. Also, check out the API Migration Guides to see if there is any API changes related to migrating savepoints to newer versions.
In particular, we advise you to check that explicit
uids were set for operators in your job.
This is a soft precondition, and restore should still work in case you forgot about assigning
If you run into a case where this is not working, you can manually add the generated legacy vertex ids from previous
Flink versions to your job using the
setUidHash(String hash) call. For each operator (in operator chains: only the
head operator) you must assign the 32 character hex string representing the hash that you can see in the web ui or logs
for the operator.
Besides operator uids, there are currently two hard preconditions for job migration that will make migration fail:
We do not support migration for state in RocksDB that was checkpointed using
semi-asynchronous mode. In case your old job was using this mode, you can still change your job to use
fully-asynchronous mode before taking the savepoint that is used as the basis for the migration.
Another important precondition is that for savepoints taken before Flink 1.3.x, all the savepoint data must be accessible from the new installation and reside under the same absolute path. Before Flink 1.3.x, the savepoint data is typically not self-contained in just the created savepoint file. Additional files can be referenced from inside the savepoint file (e.g. the output from state backend snapshots). Since Flink 1.3.x, this is no longer a limitation; savepoints can be relocated using typical filesystem operations..
First major step in job migration is taking a savepoint of your job running in the older Flink version. You can do this with the command:
For more details, please read the savepoint documentation.
In this step, we update the framework version of the cluster. What this basically means is replacing the content of the Flink installation with the new version. This step can depend on how you are running Flink in your cluster (e.g. standalone, on Mesos, …).
If you are unfamiliar with installing Flink in your cluster, please read the deployment and cluster setup documentation.
As the last step of job migration, you resume from the savepoint taken above on the updated cluster. You can do this with the command:
Again, for more details, please take a look at the savepoint documentation.
Savepoints are compatible across Flink versions as indicated by the table below:
|Created with \ Resumed with||1.1.x||1.2.x||1.3.x||1.4.x||1.5.x||1.6.x||1.7.x||1.8.x||1.9.x||Limitations|
|1.1.x||O||O||O||The maximum parallelism of a job that was migrated from Flink 1.1.x to 1.2.x+ is currently fixed as the parallelism of the job. This means that the parallelism can not be increased after migration. This limitation might be removed in a future bugfix release.|
When migrating from Flink 1.2.x to Flink 1.3.x+, changing parallelism at the same
time is not supported. Users have to first take a savepoint after migrating to Flink 1.3.x+, and then change
Savepoints created for CEP applications cannot be restored in 1.4.x+.
Savepoints from Flink 1.2 that contain a Scala TraversableSerializer are not compatible with Flink 1.8 anymore because of an update in this serializer. You can get around this restriction by first upgrading to a version between Flink 1.3 and Flink 1.7 and then updating to Flink 1.8.
|1.3.x||O||O||O||O||O||O||O||Migrating from Flink 1.3.0 to Flink 1.4.[0,1] will fail if the savepoint contains Scala case classes. Users have to directly migrate to 1.4.2+ instead.|
|1.5.x||O||O||O||O||O||There is a known issue with resuming broadcast state created with 1.5.x in versions 1.6.x up to 1.6.2, and 1.7.0: FLINK-11087. Users upgrading to 1.6.x or 1.7.x series need to directly migrate to minor versions higher than 1.6.2 and 1.7.0, respectively.|