[
https://issues.apache.org/jira/browse/KAFKA-15297?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17752108#comment-17752108
]
Guozhang Wang commented on KAFKA-15297:
---------------------------------------
{{We are not sure if we should not instead decouple caching from forwarding.}}
I'd assume double negations here mean "We think we should just try to decouple
caching from forwarding as the right solution" :) And yes, I'd love to see that
happening as I've advocated for it for many years, and I was thinking about
just "suppressing" the records in the last sink processor of the sub-topology
to achieve the same effect of less send over the network. It may be just
similar to what you meant by "caching on the last state store" or may be having
some corner differences. In either way like you said it will lose some benefit
of processing less records at the later stage of a sub-topology, but I think in
most cases given a sub-topology's size this seems a good trade for simplicity.
It also have many other benefits, just to name a few: 1) we have much simpler
timestamp tracking (today it's as finer-grained as per-processor) with a task
as every record will always go through the whole sub-topology, 2) we have
simpler version tracking within sub-topologies for IQ since now all state
stores have the same version.
> Cache flush order might not be topological order
> -------------------------------------------------
>
> Key: KAFKA-15297
> URL: https://issues.apache.org/jira/browse/KAFKA-15297
> Project: Kafka
> Issue Type: Bug
> Components: streams
> Affects Versions: 3.4.0
> Reporter: Bruno Cadonna
> Priority: Major
> Attachments: minimal_example.png
>
>
> The flush order of the state store caches in Kafka Streams might not
> correspond to the topological order of the state stores in the topology. The
> order depends on how the processors and state stores are added to the
> topology.
> In some cases downstream state stores might be flushed before upstream state
> stores. That means, that during a commit records in upstream caches might end
> up in downstream caches that have already been flushed during the same
> commit. If a crash happens at that point, those records in the downstream
> caches are lost. Those records are lost for two reasons:
> 1. Records in caches are only changelogged after they are flushed from the
> cache. However, the downstream caches have already been flushed and they will
> not be flushed again during the same commit.
> 2. The offsets of the input records that caused the records that now are
> blocked in the downstream caches are committed during the same commit and so
> they will not be re-processed after the crash.
> An example for a topology where the flush order of the caches is wrong is the
> following:
> {code:java}
> final String inputTopic1 = "inputTopic1";
> final String inputTopic2 = "inputTopic2";
> final String outputTopic1 = "outputTopic1";
> final String processorName = "processor1";
> final String stateStoreA = "stateStoreA";
> final String stateStoreB = "stateStoreB";
> final String stateStoreC = "stateStoreC";
> streamsBuilder.stream(inputTopic2, Consumed.with(Serdes.String(),
> Serdes.String()))
> .process(
> () -> new Processor<String, String, String, String>() {
> private ProcessorContext<String, String> context;
> @Override
> public void init(ProcessorContext<String, String> context) {
> this.context = context;
> }
> @Override
> public void process(Record<String, String> record) {
> context.forward(record);
> }
> @Override
> public void close() {}
> },
> Named.as("processor1")
> )
> .to(outputTopic1, Produced.with(Serdes.String(), Serdes.String()));
> streamsBuilder.stream(inputTopic1, Consumed.with(Serdes.String(),
> Serdes.String()))
> .toTable(Materialized.<String, String, KeyValueStore<Bytes,
> byte[]>>as(stateStoreA).withKeySerde(Serdes.String()).withValueSerde(Serdes.String()))
> .mapValues(value -> value, Materialized.<String, String,
> KeyValueStore<Bytes,
> byte[]>>as(stateStoreB).withKeySerde(Serdes.String()).withValueSerde(Serdes.String()))
> .mapValues(value -> value, Materialized.<String, String,
> KeyValueStore<Bytes,
> byte[]>>as(stateStoreC).withKeySerde(Serdes.String()).withValueSerde(Serdes.String()))
> .toStream()
> .to(outputTopic1, Produced.with(Serdes.String(), Serdes.String()));
> final Topology topology = streamsBuilder.build(streamsConfiguration);
> topology.connectProcessorAndStateStores(processorName, stateStoreC);
> {code}
> This code results in the attached topology.
> In the topology {{processor1}} is connected to {{stateStoreC}}. If
> {{processor1}} is added to the topology before the other processors, i.e., if
> the right branch of the topology is added before the left branch as in the
> code above, the cache of {{stateStoreC}} is flushed before the caches of
> {{stateStoreA}} and {{stateStoreB}}.
> You can observe the flush order by feeding some records into the input topics
> of the topology, waiting for a commit, and looking for the following log
> message:
> https://github.com/apache/kafka/blob/2e1947d240607d53f071f61c875cfffc3fec47fe/streams/src/main/java/org/apache/kafka/streams/processor/internals/ProcessorStateManager.java#L513
>
> I changed the log message from trace to debug to avoid too much noise.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
