flink-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Tzu-Li (Gordon) Tai" <tzuli...@apache.org>
Subject Re: High back-pressure after recovering from a save point
Date Fri, 14 Jul 2017 16:38:06 GMT
Can you try starting from the savepoint, but telling Kafka to start from the latest offset?

(@gordon: Is that possible in Flink 1.3.1 or only in 1.4-SNAPSHOT ?)
This is already possible in Flink 1.3.x. `FlinkKafkaConsumer#setStartFromLatest()` would be
it.


On 15 July 2017 at 12:33:53 AM, Stephan Ewen (sewen@apache.org) wrote:

Can you try starting from the savepoint, but telling Kafka to start from the latest offset?

(@gordon: Is that possible in Flink 1.3.1 or only in 1.4-SNAPSHOT ?)

On Fri, Jul 14, 2017 at 11:18 AM, Kien Truong <duckientruong@gmail.com> wrote:
Hi,

Sorry for the version typo, I'm running 1.3.1. I did not test with 1.2.x.

The jobs runs fine with almost 0 back-pressure if it's started from scratch or if I reuse
the kafka consumers group id without specifying the safe point.

Best regards,
Kien
On Jul 14, 2017, at 15:59, Stephan Ewen <sewen@apache.org> wrote:
Hi!

Flink 1.3.2 does not yet exist. Do you mean 1.3.1 or latest master?

Can you tell us whether this occurs only in 1.3.x and worked well in 1.2.x?
If you just keep the job running without savepoint/restore, you do not get into backpressure
situations?

Thanks,
Stephan


On Fri, Jul 14, 2017 at 1:15 AM, Kien Truong <duckientruong@gmail.com> wrote:
Hi Fabian,
This happens to me even when the restore is immediate, so there's not much data in Kafka to
catch up (5 minutes max)

Regards
Kien
On Jul 13, 2017, at 23:40, Fabian Hueske < fhueske@gmail.com> wrote:
I would guess that this is quite usual because the job has to "catch-up" work.
For example, if you took a save point two days ago and restore the job today, the input data
of the last two days has been written to Kafka (assuming Kafka as source) and needs to be
processed.
The job will now read as fast as possible from Kafka to catch-up to the presence. This means
the data is much fast ingested (as fast as Kafka can read and ship it) than during regular
processing (as fast as your sources produce).
The processing speed is bound by your Flink job which means there will be backpressure.

Once the job caught-up, the backpressure should disappear.

Best, Fabian

2017-07-13 15:48 GMT+02:00 Kien Truong <duckientruong@gmail.com>:
Hi all,

I have one job where back-pressure  is significantly higher after resuming from a save point.

Because that job makes heavy use of stateful functions with RocksDBStateBackend ,

I'm suspecting that this is the cause of performance degradation.

Does anyone encounter simillar issues or have any tips for debugging ?


I'm using Flink 1.3.2 with YARN in detached mode.


Regards,

Kien





Mime
View raw message