Hi Le, As I wrote, you can try running Flink in job mode, which spawns separate clusters per each job.
Till, is this issue covered by FLINK-11815 <https://issues.apache.org/jira/browse/FLINK-11815> ? Is this the same as: > Known issues: > 1. (…) > 2. if task slots are registered before slot request, the code have a tendency > to group requests together on the same machine because we are using a > LinkedHashMap ? Piotrek > On 4 Mar 2019, at 21:08, Le Xu <sharonx...@gmail.com> wrote: > > Thanks Piotr. > > I didn't realize that the email attachment isn't working so the example I was > referring to was this figure from Flink website: > https://ci.apache.org/projects/flink/flink-docs-stable/fig/slot_sharing.svg > <https://ci.apache.org/projects/flink/flink-docs-stable/fig/slot_sharing.svg> > > So I try to run multiple jobs concurrently in a cluster -- the jobs are > identical and the DAG looks very similar to the one in the figure. Each > machine holds one map task from each job. I end up with X number of sinks on > machine 1 (X being the number of jobs). I assume this is caused by the > operator chaining (so that all sinks are chained to mapper 1 all end up on > machine 1). But I also tried disabling chaining but I still get the same > result. Some how even when the sink and the map belongs to different threads > they are still placed in the same slot. > > My goal was to see whether it is possible to have sinks evenly distributed > across the cluster (instead of all on machine 1). One way to do this is to > see if it is ok to chained the sink to one of the other mapper -- the other > way is to see if we can change the placement of the mapper altogether (like > placing map 1 of job 2 on machine 2, map 1 of job 3 on machine 3 so we end up > with sinks sit evenly throughout the cluster). > > Thanks. > > Le > > On Mon, Mar 4, 2019 at 6:49 AM Piotr Nowojski <pi...@ververica.com > <mailto:pi...@ververica.com>> wrote: > Hi, > > Are you asking the question if that’s the behaviour or you have actually > observed this issue? I’m not entirely sure, but I would guess that the Sink > tasks would be distributed randomly across the cluster, but maybe I’m mixing > this issue with resource allocations for Task Managers. Maybe Till will know > something more about this? > > One thing that might have solve/workaround the issue is to run those jobs in > the job mode (one cluster per job), not in cluster mode, since containers for > Task Managers are created/requested randomly. > > Piotrek > >> On 2 Mar 2019, at 23:53, Le Xu <sharonx...@gmail.com >> <mailto:sharonx...@gmail.com>> wrote: >> >> Hello! >> >> I'm trying to find out if there a way to force task slot sharing within a >> job. The example on the website looks like the following (as in the >> screenshot) >> >> <image.png> >> In this example, the single sink is slot-sharing with source/map (1) and >> window operator (1). If I deploy multiple identical jobs shown above, all >> sink operators would be placed on the first machine (which creates an >> unbalanced scenario). Is there a way to avoid this situation (i.e., to have >> sink operators of different jobs spread evenly across the task slots for the >> entire cluster). Specifically, I was wondering if either of the following >> options are possible: >> 1. To force Sink[1] to be slot sharing with mapper from a different >> partition on other slots such as (source[2] and window[2]). >> 2. If option 1 is not possible, is there a "hacky" way for Flink to deploy >> jobs starting from a different machine: e.g. For job 2, it can allocate >> source/map[1], window[1], sink[1] to machine 2 instead of again on machine >> 1. In this way the slot-sharing groups are still the same, but we end up >> having sinks from the two jobs on different machines. >> >> >> Thanks! >> >> >> >> >
