RE: Flink logs only written to one host

Posted by Emmanuel on
URL: http://deprecated-apache-flink-user-mailing-list-archive.369.s1.nabble.com/Flink-logs-only-written-to-one-host-tp831p834.html

Hi,

I did change my config to have parallelism of 9 and 3 slots on each machine and now it does distribute properly.

The other day I was told i could have many more slots than CPUs available and the system would distribute the load properly between the hosts with available CPU time, but it doesn't seem to be the case here:

with 9 slots on each hosts, the distribute partitioning still send all 9 tasks to the 9 slots on the same machine even though it only has 3CPUs available and the other 6 CPUs on the other machines were not running anything.

The idea of having more slots than CPUs to me is that you can spin out more jobs and they'll be balanced as data comes... but with few jobs it seems odd that the task are all attributed to the same host just because it has slots available.

Anyways, that fixed that issue for me for now.. .looking forward to learning more about Flink

---

Date: Thu, 12 Mar 2015 20:00:32 +0100
Subject: RE: Flink logs only written to one host
From: [hidden email]
To: [hidden email]

Can you check the JM log file how many slots are available?

Slots are configured per TM. If you configure 9 slots and 3 TMs you end up with 27 slots, 9 on each TM.

On Mar 12, 2015 7:55 PM, "Emmanuel" <[hidden email]> wrote:

It appears actually that the slots used are all on the same host.

My guess is because I am using the default partitioning method (forward, which defaults to the same host) 

However I now tried .shuffle() and .distribute() without any luck:

I have a 

DataStream<String> text = env.socketTextStream(inHostName, inPort);

this is the one socket input stream.

Adding text.distribute().map(...)

does not seem to distribute the .map() process on the other hosts.

Is this the correct way to use .distribute() on a stream input? 

Thanks

Emmanuel

---

From: [hidden email]
To: [hidden email]
Subject: Flink logs only written to one host
Date: Thu, 12 Mar 2015 17:30:28 +0000

Hello,

I'm using a 3 nodes (3VMs) cluster, 3CPUs each, parallelism of 9, 

I usually only see taskmanager.out logs generated only on one of the 3 nodes when I use the System.out.println() method, to print debug info in my main processing function.

Is this expected? Or am I just doing something wrong? 

I stream from a socket with socketTextStream; I understand that this job runs on a single process, and I see that in the UI (using one slot only), but the computation task runs on 9 slots. That task includes the System.out.println() statement, yet it only shows on one host's .out log folder. 

The host is not always the same, so I have to tail all logs on all hosts, but I'm surprised of this behavior.

Am I just missing something? 

Are 'print' statement to stdout aggregated on one host somehow? If so how is this controlled? Why would that host change?

I would love to understand what is going on, and if maybe somehow the 9 slots may be running on a single host which would defeat the purpose.

Thanks for the insight

Emmanuel