[
https://issues.apache.org/jira/browse/FLINK-4022?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15629801#comment-15629801
]
Tzu-Li (Gordon) Tai commented on FLINK-4022:
--------------------------------------------
Hi [~StephanEwen],
Right now, for subtasks that do not have partitions assigned to it, we will
emit a max value watermark from that subtask because it will not read any data.
However, with partition discovery enabled, subtasks which initially do not have
partitions might be assigned one later on and start reading data, and this will
mess up the watermarks downstream.
We also cannot let idle subtasks not emit any watermarks and just wait for
partitions be assigned to it, because then downstream time window states will
unboundly accumulate (FLINK-4341).
So, the watermark service came around as a mean to let source subtasks that
currently do not read any data (thus cannot emit any watermarks for event time)
emit the global low watermark across subtasks instead.
For this I think we still need the watermark service.
> Partition discovery / regex topic subscription for the Kafka consumer
> ---------------------------------------------------------------------
>
> Key: FLINK-4022
> URL: https://issues.apache.org/jira/browse/FLINK-4022
> Project: Flink
> Issue Type: New Feature
> Components: Kafka Connector, Streaming Connectors
> Affects Versions: 1.0.0
> Reporter: Tzu-Li (Gordon) Tai
> Assignee: Tzu-Li (Gordon) Tai
> Fix For: 1.2.0
>
>
> Example: allow users to subscribe to "topic-n*", so that the consumer
> automatically reads from "topic-n1", "topic-n2", ... and so on as they are
> added to Kafka.
> I propose to implement this feature by the following description:
> Since the overall list of partitions to read will change after job
> submission, the main big change required for this feature will be dynamic
> partition assignment to subtasks while the Kafka consumer is running. This
> will mainly be accomplished using Kafka 0.9.x API
> `KafkaConsumer#subscribe(java.util.regex.Pattern,
> ConsumerRebalanceListener)`. Each KafkaConsumers in each subtask will be
> added to the same consumer group when instantiated, and rely on Kafka to
> dynamically reassign partitions to them whenever a rebalance happens. The
> registered `ConsumerRebalanceListener` is a callback that is called right
> before and after rebalancing happens. We'll use this callback to let each
> subtask commit its last offsets of partitions its currently responsible of to
> an external store (or Kafka) before a rebalance; after rebalance and the
> substasks gets the new partitions it'll be reading from, they'll read from
> the external store to get the last offsets for their new partitions
> (partitions which don't have offset entries in the store are new partitions
> causing the rebalancing).
> The tricky part will be restoring Flink checkpoints when the partition
> assignment is dynamic. Snapshotting will remain the same - subtasks snapshot
> the offsets of partitions they are currently holding. Restoring will be a
> bit different in that subtasks might not be assigned matching partitions to
> the snapshot the subtask is restored with (since we're letting Kafka
> dynamically assign partitions). There will need to be a coordination process
> where, if a restore state exists, all subtasks first commit the offsets they
> receive (as a result of the restore state) to the external store, and then
> all subtasks attempt to find a last offset for the partitions it is holding.
> However, if the globally merged restore state feature mentioned by
> [~StephanEwen] in https://issues.apache.org/jira/browse/FLINK-3231 is
> available, then the restore will be simple again, as each subtask has full
> access to previous global state therefore coordination is not required.
> I think changing to dynamic partition assignment is also good in the long run
> for handling topic repartitioning.
> Overall,
> User-facing API changes:
> - New constructor - FlinkKafkaConsumer09(java.util.regex.Pattern,
> DeserializationSchema, Properties)
> - New constructor - FlinkKafkaConsumer09(java.util.regex.Pattern,
> KeyedDeserializationSchema, Properties)
> Implementation changes:
> 1. Dynamic partition assigning depending on KafkaConsumer#subscribe
> - Remove partition list querying from constructor
> - Remove static partition assigning to substasks in run()
> - Instead of using KafkaConsumer#assign() in fetchers to manually assign
> static partitions, use subscribe() registered with the callback
> implementation explained above.
> 2. Restoring from checkpointed states
> - Snapshotting should remain unchanged
> - Restoring requires subtasks to coordinate the restored offsets they hold
> before continuing (unless we are able to have merged restore states).
> 3. For previous consumer functionality (consume from fixed list of topics),
> the KafkaConsumer#subscribe() has a corresponding overload method for fixed
> list of topics. We can simply decide which subscribe() overload to use
> depending on whether a regex Pattern or list of topics is supplied.
> 4. If subtasks don't initially have any assigned partitions, we shouldn't
> emit MAX_VALUE watermark, since it may hold partitions after a rebalance.
> Instead, un-assigned subtasks should be running a fetcher instance too and
> take part as a process pool for the consumer group of the subscribed topics.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
