[
https://issues.apache.org/jira/browse/FLINK-9114?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16428587#comment-16428587
]
Jacob Park commented on FLINK-9114:
-----------------------------------
I was thinking of creating a ConfigurableCheckpointRecoveryFactory interface
with a configure(Configuration config, Executor executor) method instead of a
constructor for ease of class-loading and reflection. It would also exist in a
separate package to prevent issues with class-loading conflicts when building a
JAR (like flink-metrics).
The ConfigurableCheckpointRecoveryFactory will be instantiated with the help of
a ConfigurableCheckpointRecoveryFactoryLoader invoked in a new abstract class
that subclasses HighAvailabilityServices and overrides
getCheckpointRecoveryFactory(). This new abstract class would be the parent
class for YarnHighAvailabilityServices, StandaloneHaServices, and
EmbeddedHaServices.
I hope this approach won't be too invasive for the existing
StandaloneCheckpointRecoveryFactory as the configure() method will be a no-op
for it, and it would not impact how a JobManager utilizes the
CheckpointRecoveryFactory to create CompletedCheckpointStore in the
ExecutionGraph.
> Enable user-provided, custom CheckpointRecoveryFactory for non-HA deployments
> -----------------------------------------------------------------------------
>
> Key: FLINK-9114
> URL: https://issues.apache.org/jira/browse/FLINK-9114
> Project: Flink
> Issue Type: Improvement
> Components: Configuration, State Backends, Checkpointing
> Reporter: Jacob Park
> Assignee: Jacob Park
> Priority: Major
>
> When you operate a Flink application that uses externalized checkpoints to
> S3, it becomes difficult to determine which checkpoint is the latest to
> recover from. Because S3 provides read-after-write consistency only for PUTS,
> listing a S3 path is not guaranteed to be consistent, so we do not know what
> checkpoint to recover from.
> The goal of this improvement is to allow users to provide a custom
> CheckpointRecoveryFactory for non-HA deployments such that we can use this
> feature to fail checkpoints if we cannot guarantee we will know where a
> checkpoint will be in S3, and co-publish checkpoint metadata to a strongly
> consistent data store.
> I propose the following changes:
> # Modify AbstractNonHaServices and StandaloneHaServices to accept an
> Executor for the custom CheckpointRecoveryFactory.
> # Create a CheckpointRecoveryFactoryLoader to provide the custom
> CheckpointRecoveryFactory from configurations.
> # Add new configurations for this feature.
> We considered the pluggable StateBackend and potential pluggable
> HighAvailabilityServices. These were too convoluted to solve our problem, so
> we would like custom CheckpointRecoveryFactory.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
