[
https://issues.apache.org/jira/browse/KAFKA-14768?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
fujian updated KAFKA-14768:
---------------------------
Description:
Hi, Team:
Nice to meet you!
In our business, we found two types of issue which need to improve:
*(1) Take much time to send the first message*
Sometimes, we found the users' functional interaction take a lot of time. At
last, we figure out the root cause is that after we complete deploy or restart
the servers. The first message's delivery on each application server by kafka
client will take much time.
So, we try to find one solution to improve it.
After analyzing the source code about the first time's sending logic. The time
cost is caused by the getting metadata before the sending. The latter's sending
won't take the much time due to the cached metadata. The logic is right and
necessary. Thus, we still want to improve the experience for the first
message's send/user first interaction.
*(2) can't reduce the send message's block time to wanted value.*
Sometimes our application's thread will block for max.block.ms to send message.
When we try to reduce the max.block.ms to reduce the blocking time. It can't
meet the getting metadata's time requirement sometimes. The root cause is the
configured max.block.ms is shared with "get metadata" operation and "send
message" operation. We can refer to follow tables:
|*where to block*
|*when it is blocked*
|*how long it will be blocked?*
|
|org.apache.kafka.clients.producer.KafkaProducer#waitOnMetadata|the first
request which need to load the metadata from kafka|<max.block.ms|
|org.apache.kafka.clients.producer.internals.RecordAccumulator#append|at peak
time for business, if the network can’t send message in short
time.|<max.block.ms|
What's the solution for the above two issues:
I think about current logic and figure out followed possible solution:
(1) send one "warmup" message, thus we can't send any fake message.
(2) provide one extra configure time configure which dedicated for getting
metadata. thus it will break the define for the max.block.ms
(3) change the private to public for the method or provide dedicated method for
this support.
_private ClusterAndWaitTime waitOnMetadata(String topic, Integer partition,
long nowMs, long maxWaitMs)_
so that we can call it before the service is marked as ready. After the ready.
it won't block to get metadata due to cache. And then we can reduce the
max.block.ms to a lower value to reduce thread's blocking time for thread.
After adopt the solution 3. we solve the above issues. For example, we reduce
the first message's send about 4s seconds. The log can refer to followed:
_warmup test_topic at phase phase 2: get metadata from mq start_
_warmup test_topic at phase phase 2: get metadata from mq end consume *4669ms*_
And after the change, we reduce the max.block.ms from 10s to 2s without worry
can't get metadata.
{*}So what's your thought for these two issues and the solution I proposed{*}.
If there is no problem for it. I can create one PR to merge. I hope to get your
feedback and thought for the issues.
was:
{*}Background{*}:
In [KIP-405: Kafka Tiered Storage - Apache Kafka - Apache Software
Foundation|https://cwiki.apache.org/confluence/display/KAFKA/KIP-405%3A+Kafka+Tiered+Storage],
kafka introduced the feature of hierarchical storage.
Also, [KAFKA-9555] Topic-based implementation for the RemoteLogMetadataManager
- ASF JIRA (apache.org) implements the default RLMM - 'TopicBased-RLMM'.
{*}Problem{*}:
TopicBased-RLMM will only subscribe to the Partitions where the current Broker
is Leader or Follower. If the current Broker is not the Leader or Follower,
then RLMM will directly skip the metadata records related to these Partitions.
When reassign user-partitions occurs, rlmm will subscribe to new
user-partitions, assuming that the metadata-partition to which the new
user-partition belongs is 'metadata-partition0', and RLMM has consumed
'metadata-partition0' *to offset = 100* before the reassign partition occurs,
then {*}after reassign{*}, RMLM will *not* consume 'metadata-partition0'
\{*}from the beginning{*}, and finally cause the metadata records related to
the new user-partition to *be lost with offset < 100.*
*Solution*
Let RLMM subscribe to all user-patitions, instead of only subscribing to
partitions where the current broker is leader or follower.
In this way, when reassign partition occurs, RLMM will have new partition's
metadata records.
> proposal to reduce the first message's send time cost and max block time for
> safety
> ------------------------------------------------------------------------------------
>
> Key: KAFKA-14768
> URL: https://issues.apache.org/jira/browse/KAFKA-14768
> Project: Kafka
> Issue Type: Improvement
> Components: core
> Affects Versions: 3.3.1
> Reporter: fujian
> Assignee: hzh0425
> Priority: Blocker
>
> Hi, Team:
>
> Nice to meet you!
>
> In our business, we found two types of issue which need to improve:
>
> *(1) Take much time to send the first message*
> Sometimes, we found the users' functional interaction take a lot of time. At
> last, we figure out the root cause is that after we complete deploy or
> restart the servers. The first message's delivery on each application server
> by kafka client will take much time.
> So, we try to find one solution to improve it.
>
> After analyzing the source code about the first time's sending logic. The
> time cost is caused by the getting metadata before the sending. The latter's
> sending won't take the much time due to the cached metadata. The logic is
> right and necessary. Thus, we still want to improve the experience for the
> first message's send/user first interaction.
>
> *(2) can't reduce the send message's block time to wanted value.*
> Sometimes our application's thread will block for max.block.ms to send
> message. When we try to reduce the max.block.ms to reduce the blocking time.
> It can't meet the getting metadata's time requirement sometimes. The root
> cause is the configured max.block.ms is shared with "get metadata" operation
> and "send message" operation. We can refer to follow tables:
> |*where to block*
> |*when it is blocked*
> |*how long it will be blocked?*
> |
> |org.apache.kafka.clients.producer.KafkaProducer#waitOnMetadata|the first
> request which need to load the metadata from kafka|<max.block.ms|
> |org.apache.kafka.clients.producer.internals.RecordAccumulator#append|at peak
> time for business, if the network can’t send message in short
> time.|<max.block.ms|
>
> What's the solution for the above two issues:
> I think about current logic and figure out followed possible solution:
> (1) send one "warmup" message, thus we can't send any fake message.
> (2) provide one extra configure time configure which dedicated for getting
> metadata. thus it will break the define for the max.block.ms
> (3) change the private to public for the method or provide dedicated method
> for this support.
> _private ClusterAndWaitTime waitOnMetadata(String topic, Integer partition,
> long nowMs, long maxWaitMs)_
>
> so that we can call it before the service is marked as ready. After the
> ready. it won't block to get metadata due to cache. And then we can reduce
> the max.block.ms to a lower value to reduce thread's blocking time for thread.
>
> After adopt the solution 3. we solve the above issues. For example, we reduce
> the first message's send about 4s seconds. The log can refer to followed:
> _warmup test_topic at phase phase 2: get metadata from mq start_
> _warmup test_topic at phase phase 2: get metadata from mq end consume
> *4669ms*_
> And after the change, we reduce the max.block.ms from 10s to 2s without worry
> can't get metadata.
>
> {*}So what's your thought for these two issues and the solution I
> proposed{*}. If there is no problem for it. I can create one PR to merge. I
> hope to get your feedback and thought for the issues.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
