Hi, I have seen a few cases where for certain jobs a small imbalance in the state partition assignment did cascade into a larger imbalance of the job. If your max parallelism mod parallelism is not 0, it means that some tasks have one partition more than others. Again, depending on how much partitions you have assigned to each task, in the extremest case when every task has 1 key group, except for one that has 2, imbalance can be 100%. Maybe you could check for that, especially if you were running at a different parallelism in production and stress testing. This would also explain why the any checkpoint duration is longer for a task, because it would have much more state - assuming that the load is kind of balanced between partitions.
Best, Stefan > On 15. Jan 2019, at 11:42, Bruno Aranda <bara...@apache.org> wrote: > > Hi, > > Just an update from our side. We couldn't find anything specific in the logs > and the problem is not easy reproducible. This week, the system is running > fine, which makes me suspicious as well of some resourcing issue. But so far, > we haven't been able to find the reason though we have discarded a few > things. We consume from Kafka, and the load was properly balanced. We > couldn't find a relationship between rate and the task manager checkpoint > being slower. The problem could happen even at the times of day where we get > less messages. After a flink session restart (using AWS EMR), another TM in a > different machine could have been the one with the longer checkpoints. > > We are now trying to reproduce the problem in a different cluster by trying > to send the data that was crossing the system while we saw the problems and > see if we can identify something specific to it. But our data is pretty > uniform, so not sure, and so far we have only seen this problem in our Prod > environment and not when running stress tests which much higher load. > > Will come back if we figure anything out. > > Thanks, > > Bruno > > On Tue, 15 Jan 2019 at 10:33, Till Rohrmann <trohrm...@apache.org > <mailto:trohrm...@apache.org>> wrote: > Same here Pasquale, the logs on DEBUG log level could be helpful. My guess > would be that the respective tasks are overloaded or there is some resource > congestion (network, disk, etc). > > You should see in the web UI the number of incoming and outgoing events. It > would be good to check that the events are similarly sized and can be > computed in roughly the same time. > > Cheers, > Till > > On Mon, Jan 14, 2019 at 4:07 PM Pasquale Vazzana <p.vazz...@mwam.com > <mailto:p.vazz...@mwam.com>> wrote: > I have the same problem, even more impactful. Some subtasks stall forever > quite consistently. > I am using Flink 1.7.1, but I've tried downgrading to 1.6.3 and it didn't > help. > The Backend doesn't seem to make any difference, I've tried Memory, FS and > RocksDB back ends but nothing changes. I've also tried to change the medium, > local spinning disk, SAN or mounted fs but nothing helps. > Parallelism is the only thing which mitigates the stalling, when I set 1 > everything works but if I increase the number of parallelism then everything > degrades, 10 makes it very slow 30 freezes it. > It's always one of two subtasks, most of them does the checkpoint in few > milliseconds but there is always at least one which stalls for minutes until > it times out. The Alignment seems to be a problem. > I've been wondering whether some Kafka partitions where empty but there is > not much data skew and the keyBy uses the same key strategy as the Kafka > partitions, I've tried to use murmur2 for hashing but it didn't help either. > The subtask that seems causing problems seems to be a CoProcessFunction. > I am going to debug Flink but since I'm relatively new to it, it might take a > while so any help will be appreciated. > > Pasquale > > > From: Till Rohrmann <trohrm...@apache.org <mailto:trohrm...@apache.org>> > Sent: 08 January 2019 17:35 > To: Bruno Aranda <bara...@apache.org <mailto:bara...@apache.org>> > Cc: user <user@flink.apache.org <mailto:user@flink.apache.org>> > Subject: Re: Subtask much slower than the others when creating checkpoints > > Hi Bruno, > > there are multiple reasons wh= one of the subtasks can take longer for > checkpointing. It looks as if the=e is not much data skew since the state > sizes are relatively equal. It als= looks as if the individual tasks all > start at the same time with the chec=pointing which indicates that there > mustn't be a lot of back-pressure =n the DAG (or all tasks were equally > back-pressured). This narrows the pro=lem cause down to the asynchronous > write operation. One potential problem =ould be if the external system to > which you write your checkpoint data has=some kind of I/O limit/quota. Maybe > the sum of write accesses deplete the =aximum quota you have. You could try > whether running the job with a lower =arallelism solves the problems. > > For further debug=ing it could be helpful to get access to the logs of the > JobManager and th= TaskManagers on DEBUG log level. It could also be helpful > to learn which =tate backend you are using. > > Cheers, > Til= > > On Tue, Jan 8,=2019 at 12:52 PM Bruno Aranda <mailto:bara...@apache.org > <mailto:bara...@apache.org>> wrote: > Hi, > > We are using Flink =.6.1 at the moment and we have a streaming job configured > to create a chec=point every 10 seconds. Looking at the checkpointing times > in the UI, we c=n see that one subtask is much slower creating the endpoint, > at least in i=s "End to End Duration", and seems caused by a longer > "Chec=point Duration (Async)". > > For instance, in th= attach screenshot, while most of the subtasks take half > a second, one (an= it is always one) takes 2 seconds. > > But we have w=rse problems. We have seen cases where the checkpoint times out > for one ta=ks, while most take one second, the outlier takes more than 5 > minutes (whi=h is the max time we allow for a checkpoint). This can happen if > there is =ack pressure. We only allow one checkpoint at a time as well. > Why could one subtask take more time? This jobs read from kafk= partitions > and hash by key, and we don't see any major data skew betw=en the partitions. > Does one partition do more work? > > We do have a cluster of 20 machines, in EMR, with TMs that have > multiple=slots (in legacy mode). > > Is this something that co=ld have been fixed in a more recent version? > > Than=s for any insight! > > Bruno > > > This e-mail and any attachments are confidential to the addressee(s) and may > contain information that is legally privileged and/or confidential. Please > refer to http://www.mwam.com/email-disclaimer-uk > <http://www.mwam.com/email-disclaimer-uk> for important disclosures regarding > this email. If we collect and use your personal data we will use it in > accordance with our privacy policy, which can be reviewed at > https://www.mwam.com/privacy-policy <https://www.mwam.com/privacy-policy> . > > Marshall Wace LLP is authorised and regulated by the Financial Conduct > Authority. Marshall Wace LLP is a limited liability partnership registered in > England and Wales with registered number OC302228 and registered office at > George House, 131 Sloane Street, London, SW1X 9AT. If you are receiving this > e-mail as a client, or an investor in an investment vehicle, managed or > advised by Marshall Wace North America L.P., the sender of this e-mail is > communicating with you in the sender's capacity as an associated person of > and on behalf of Marshall Wace North America L.P., which is registered with > the US Securities and Exchange Commission as an investment adviser.
