Hi Enric, Seems that versions before 2.0-preview just handle the state I/O with the > user thread?
That's true. It's a single-thread model. Best, Zakelly On Tue, Nov 26, 2024 at 3:42 PM Enric Ott <243816...@qq.com> wrote: > Thank you,Zakelly. > Seems that versions before 2.0-preview just handle the state I/O with the > user thread?My gosh. > > > ------------------ 原始邮件 ------------------ > *发件人:* "Zakelly Lan" <zakelly....@gmail.com>; > *发送时间:* 2024年11月26日(星期二) 中午11:57 > *收件人:* "user"<user@flink.apache.org>; > *主题:* Re: How the Async Exuecution Model improved the throughtput > > Hi Enric, > > The asynchronous state processing prevents the task thread from blocking > at the state I/O and instead allows it to perform CPU operations for > another input record in the meantime. Additionally, state I/Os can run in > parallel, reducing the total I/O time. Therefore, it is suitable for the > following scenarios: > > 1. Heavy I/O, which is often the case when the state is large. > 2. No hotspot key, allowing for greater parallelism in state access. > > This approach increases throughput by parallelizing the CPU operations and > multiple state accesses, with better results than simply increasing task > concurrency and lower cost. However, it is important to note that it cannot > solve everything, particularly when the state is small or there is > significant data skew. > > > Best, > Zakelly > > On Tue, Nov 26, 2024 at 11:12 AM Enric Ott <243816...@qq.com> wrote: > >> Hello,Community: >> I'm conducting experiments on flink-release-2.0-preview1,and I >> got puzzled that How the *Async Exuecution Model* achieved significant >> improvement on end to end throughput >> in the scenarios of streaming state processing.The state access is >> relatively light weighted(in my personal opinion),would the async accessing >> just cracked the nut and solved everything ? >> Thanks. >> >
