[
https://issues.apache.org/jira/browse/FLINK-4280?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15690827#comment-15690827
]
ASF GitHub Bot commented on FLINK-4280:
---------------------------------------
Github user tzulitai commented on a diff in the pull request:
https://github.com/apache/flink/pull/2509#discussion_r89365136
--- Diff:
flink-streaming-connectors/flink-connector-kafka-0.8/src/main/java/org/apache/flink/streaming/connectors/kafka/internals/Kafka08Fetcher.java
---
@@ -139,24 +143,65 @@ public void runFetchLoop() throws Exception {
PeriodicOffsetCommitter periodicCommitter = null;
try {
- // read offsets from ZooKeeper for partitions that did
not restore offsets
- {
- List<KafkaTopicPartition>
partitionsWithNoOffset = new ArrayList<>();
- for
(KafkaTopicPartitionState<TopicAndPartition> partition :
subscribedPartitions()) {
- if (!partition.isOffsetDefined()) {
-
partitionsWithNoOffset.add(partition.getKafkaTopicPartition());
- }
+ List<KafkaTopicPartition> partitionsWithNoOffset = new
ArrayList<>();
+ for (KafkaTopicPartitionState<TopicAndPartition>
partition : subscribedPartitions()) {
+ if (!partition.isOffsetDefined()) {
+
partitionsWithNoOffset.add(partition.getKafkaTopicPartition());
}
+ }
+
+ if (partitionsWithNoOffset.size() ==
subscribedPartitions().length) {
+ // if all partitions have no initial offsets,
that means we're starting fresh without any restored state
+ switch (startupMode) {
+ case EARLIEST:
+ LOG.info("Setting starting
point as earliest offset for partitions {}", partitionsWithNoOffset);
+
+ for
(KafkaTopicPartitionState<TopicAndPartition> partition :
subscribedPartitions()) {
+
partition.setOffset(OffsetRequest.EarliestTime());
+ }
+ break;
+ case LATEST:
+ LOG.info("Setting starting
point as latest offset for partitions {}", partitionsWithNoOffset);
+
+ for
(KafkaTopicPartitionState<TopicAndPartition> partition :
subscribedPartitions()) {
+
partition.setOffset(OffsetRequest.LatestTime());
+ }
+ break;
+ default:
+ case GROUP_OFFSETS:
+ LOG.info("Using group offsets
in Zookeeper of group.id {} as starting point for partitions {}",
+
kafkaConfig.getProperty("group.id"), partitionsWithNoOffset);
+
+ Map<KafkaTopicPartition, Long>
zkOffsets = zookeeperOffsetHandler.getCommittedOffsets(partitionsWithNoOffset);
+ for
(KafkaTopicPartitionState<TopicAndPartition> partition :
subscribedPartitions()) {
+ Long offset =
zkOffsets.get(partition.getKafkaTopicPartition());
+ if (offset != null) {
+ // the
committed offset in ZK represents the next record to process,
+ // so we
subtract it by 1 to correctly represent internal state
+
partition.setOffset(offset - 1);
+ } else {
+ // if we can't
find an offset for a partition in ZK when using GROUP_OFFSETS,
+ // we default
to "auto.offset.reset" like the Kafka high-level consumer
+ LOG.warn("No
group offset can be found for partition {} in Zookeeper;" +
+ "
resetting starting offset to 'auto.offset.reset'", partition);
+
+
partition.setOffset(invalidOffsetBehavior);
+ }
+ }
+ }
+ } else if (partitionsWithNoOffset.size() > 0 &&
partitionsWithNoOffset.size() < subscribedPartitions().length) {
--- End diff --
I was adding this as a preparation for the kafka partition discovery task.
But it'd probably make sense to remove it for this PR to avoid confusion.
> New Flink-specific option to set starting position of Kafka consumer without
> respecting external offsets in ZK / Broker
> -----------------------------------------------------------------------------------------------------------------------
>
> Key: FLINK-4280
> URL: https://issues.apache.org/jira/browse/FLINK-4280
> Project: Flink
> Issue Type: New Feature
> Components: Kafka Connector
> Reporter: Tzu-Li (Gordon) Tai
> Assignee: Tzu-Li (Gordon) Tai
> Fix For: 1.2.0
>
>
> Currently, to start reading from the "earliest" and "latest" position in
> topics for the Flink Kafka consumer, users set the Kafka config
> {{auto.offset.reset}} in the provided properties configuration.
> However, the way this config actually works might be a bit misleading if
> users were trying to find a way to "read topics from a starting position".
> The way the {{auto.offset.reset}} config works in the Flink Kafka consumer
> resembles Kafka's original intent for the setting: first, existing external
> offsets committed to the ZK / brokers will be checked; if none exists, then
> will {{auto.offset.reset}} be respected.
> I propose to add Flink-specific ways to define the starting position, without
> taking into account the external offsets. The original behaviour (reference
> external offsets first) can be changed to be a user option, so that the
> behaviour can be retained for frequent Kafka users that may need some
> collaboration with existing non-Flink Kafka consumer applications.
> How users will interact with the Flink Kafka consumer after this is added,
> with a newly introduced {{flink.starting-position}} config:
> {code}
> Properties props = new Properties();
> props.setProperty("flink.starting-position", "earliest/latest");
> props.setProperty("auto.offset.reset", "..."); // this will be ignored (log a
> warning)
> props.setProperty("group.id", "...") // this won't have effect on the
> starting position anymore (may still be used in external offset committing)
> ...
> {code}
> Or, reference external offsets in ZK / broker:
> {code}
> Properties props = new Properties();
> props.setProperty("flink.starting-position", "external-offsets");
> props.setProperty("auto.offset.reset", "earliest/latest"); // default will be
> latest
> props.setProperty("group.id", "..."); // will be used to lookup external
> offsets in ZK / broker on startup
> ...
> {code}
> A thing we would need to decide on is what would the default value be for
> {{flink.starting-position}}.
> Two merits I see in adding this:
> 1. This compensates the way users generally interpret "read from a starting
> position". As the Flink Kafka connector is somewhat essentially a
> "high-level" Kafka consumer for Flink users, I think it is reasonable to add
> Flink-specific functionality that users will find useful, although it wasn't
> supported in Kafka's original consumer designs.
> 2. By adding this, the idea that "the Kafka offset store (ZK / brokers) is
> used only to expose progress to the outside world, and not used to manipulate
> how Kafka topics are read in Flink (unless users opt to do so)" is even more
> definite and solid. There was some discussion in this PR
> (https://github.com/apache/flink/pull/1690, FLINK-3398) on this aspect. I
> think adding this "decouples" more Flink's internal offset checkpointing from
> the external Kafka's offset store.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
