kafka-users mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Matthias J. Sax" <matth...@confluent.io>
Subject Re: High end-to-end latency with processing.guarantee=exactly_once
Date Thu, 20 Dec 2018 12:22:22 GMT
The problem is repartitions topics:

Kafka Streams considers those topics as transient and purges consumed
data aggressively (cf https://issues.apache.org/jira/browse/KAFKA-6150)
resulting in lost producer state for those topics :(


On 12/20/18 3:18 AM, Dmitry Minkovsky wrote:
> Also, I have read through that issue and KIP-360 to the extent my knowledge
> allows and I don't understand why I get this error constantly when exactly
> once is enabled. The KIP says
>> Idempotent/transactional semantics depend on the broker retaining state
> for each active producer id (e.g. epoch and sequence number). When the
> broker loses that state–due to segment deletion or a call to
> DeleteRecords–then additional produce requests will result in the
> How much throughput do you need before this goes away? It seems like this
> happens for me on every call... with the calls seconds apart.
> On Wed, Dec 19, 2018 at 9:12 PM Dmitry Minkovsky <dminkovsky@gmail.com>
> wrote:
>> Hello 王美功,
>> I am using 2.1.0. And, I think you nailed it on the head, because my
>> application is low throughput and I am seeing UNKNOWN_PRODUCER_ID all the
>> time with exactly once enabled. I've googled this before but couldn't
>> identify the cause. Thank you!
>> Setting retry.backoff.ms to 5 brought the latency down from 1.3s to
>> 750ms. That's tolerable for me, but I am wondering: when the KAFKA-7190 is
>> fixed, will the latency drop further?
>> Dmitry
>> On Wed, Dec 19, 2018 at 8:38 PM meigong.wang <meigong.wang@okcoin.com>
>> wrote:
>>> Which version are you using? This bug(
>>> https://issues.apache.org/jira/browse/KAFKA-7190) may increase the
>>> latency of your application, try to reduce the retry.backoff.ms,the
>>> default value is 100 ms.
>>> 王美功
>>> 原始邮件
>>> 发件人:Dmitry Minkovskydminkovsky@gmail.com
>>> 收件人:usersusers@kafka.apache.org
>>> 发送时间:2018年12月20日(周四) 09:25
>>> 主题:High end-to-end latency with processing.guarantee=exactly_once
>>> I have a process that spans several Kafka Streams applications. With the
>>> streams commit interval and producer linger both set to 5ms, when exactly
>>> once delivery is disabled, this process takes ~250ms. With exactly once
>>> enabled, the same process takes anywhere from 800-1200ms. In Enabling
>>> Exactly-Once in Kafka Streams ,">
>>> https://www.confluent.io/blog/enabling-exactly-kafka-streams/, Guozhang
>>> writes In Kafka Streams, because a new transaction is created whenever
>>> commit is called, the average transaction size is determined by the commit
>>> interval: with the same incoming traffic, a shorter commit interval will
>>> result in smaller transactions. In practice users should therefore tune the
>>> commit.interval.ms setting when exactly-once is enabled to make a good
>>> trade-off between throughput versus end-to-end processing latency. But I am
>>> not seeing much of a difference when I tune commit.interval.ms with
>>> exactly once enabled. `though()` and `.to()/.stream()` take 100-250ms even
>>> with commit.interval.ms set to 5ms. Do these latency differences sound
>>> right? Is something off? Thank you, Dmitry

View raw message