Re: Trying to figure out why a slot takes a long time to checkpoint

Posted by Stefan Richter on
URL: http://deprecated-apache-flink-user-mailing-list-archive.369.s1.nabble.com/Trying-to-figure-out-why-a-slot-takes-a-long-time-to-checkpoint-tp23108p23258.html

A debug log for state backend and checkpoint coordinator could also help. 

Am 20.09.2018 um 14:19 schrieb Stefan Richter <[hidden email]>:

Hi,

if some tasks take like 50 minutes, could you wait until such a checkpoint is in progress and (let’s say after 10 minutes) log into the node and create a (or multiple over time) thread-dump(s) for the JVM that runs the slow checkpointing task. This could help to figure out where it is stuck or waiting.

Best,

Stefan

Am 19.09.2018 um 22:30 schrieb Julio Biason <[hidden email]>:

Hey guys,

So, switching to Ceph/S3 didn't shine any new lights on the issue. Although the times are a bit higher, just a few slots are taking a magnitude longer to save. So I changed the logs for DEBUG.

The problem is: I'm not seeing anything that seems relevant; only pings from ZooKeeper, heartbeats and the S3 disconnecting from being idle.

Is there anything else that I should change to DEBUG? Akka? Kafka? Haoop? ZooKeeper? (Those are, by the default config, bumped to INFO)

All of those?

On Tue, Sep 18, 2018 at 12:34 PM, Julio Biason <[hidden email]> wrote:

Hey TIll (and others),

We don't have debug logs yet, but we decided to remove a related component: HDFS.

We are moving the storage to our Ceph install (using S3), which is running for longer than our HDFS install and we know, for sure, it runs without any problems (specially 'cause we have more people that understand Ceph than people that know HDFS at this point).

If, for some reason, the problem persists, we know it's not the underlying storage and may be something with our pipeline itself. I'll enable debug logs, then.

On Tue, Sep 18, 2018 at 4:20 AM, Till Rohrmann <[hidden email]> wrote:

This behavior seems very odd Julio. Could you indeed share the debug logs of all Flink processes in order to see why things are taking so long?

The checkpoint size of task #8 is twice as big as the second biggest checkpoint. But this should not cause an increase in checkpoint time of a factor of 8.

Cheers,

Till

On Mon, Sep 17, 2018 at 5:25 AM Renjie Liu <[hidden email]> wrote:

Hi, Julio:

This happens frequently? What state backend do you use? The async checkpoint duration and sync checkpoint duration seems normal compared to others, it seems that most of the time are spent acking the checkpoint.

On Sun, Sep 16, 2018 at 9:24 AM vino yang <[hidden email]> wrote:

Hi Julio,

Yes, it seems that fifty-five minutes is really long. 

However, it is linear with the time and size of the previous task adjacent to it in the diagram. 

I think your real application is concerned about why Flink accesses HDFS so slowly. 

You can call the DEBUG log to see if you can find any clues, or post the log to the mailing list to help others analyze the problem for you.

Thanks, vino.

Julio Biason <[hidden email]> 于2018年9月15日周六 上午7:03写道：

(Just an addendum: Although it's not a huge problem -- we can always increase the checkpoint timeout time -- this anomalous situation makes me think there is something wrong in our pipeline or in our cluster, and that is what is making the checkpoint creation go crazy.)

On Fri, Sep 14, 2018 at 8:00 PM, Julio Biason <[hidden email]> wrote:

Hey guys,

On our pipeline, we have a single slot that it's taking longer to create the checkpoint compared to other slots and we are wondering what could be causing it.

The operator in question is the window metric -- the only element in the pipeline that actually uses the state. While the other slots take 7 mins to create the checkpoint, this one -- and only this one -- takes 55mins.

Is there something I should look at to understand what's going on?

(We are storing all checkpoints in HDFS, in case that helps.)

--

--

--

--

--