question Re: Impalad COORDINATOR_ONLY with HDFS Short-Circuit Read in Support Questions

Impalad COORDINATOR_ONLY with HDFS Short-Circuit Read

anandnalya — Fri, 16 Sep 2022 14:31:41 GMT

Hi,

I'm trying to deploy a few COORDINATOR_ONLY impalad on edge nodes in my cluster (CDH 5.16.1). HDFS short-circuit read are enabled on the cluster and works fine.When starting the COORDINATOR_ONLY nodes, I'm getting following error:

Invalid short-circuit reads configuration:
  - Impala cannot read or execute the parent directory of dfs.domain.socket.path
Aborting Impala Server startup due to improper configuration. Impalad exiting.

Is there a way to run COORDINATOR_ONLY impalad with hdfs short circuit reads enabled.

Thanks in advance!

Re: Impalad COORDINATOR_ONLY with HDFS Short-Circuit Read

Gomathinayagam — Wed, 24 Jul 2019 05:01:28 GMT

Can you check if the short-circuit read is configured as per the Documentation HERE ?

Re: Impalad COORDINATOR_ONLY with HDFS Short-Circuit Read

anandnalya — Wed, 24 Jul 2019 07:11:57 GMT

dfs.client.read.shortcircuit was not enabled for HDFS Gateway. I enabled it and then deployed client configuration and restarted all the services in the cluster. Unfortunatley, I'm still getting the same error when starting COORDINATOR_ONLY impalad.

The Short circuit reads config looks like following:

Re: Impalad COORDINATOR_ONLY with HDFS Short-Circuit Read

EricL — Wed, 24 Jul 2019 11:29:33 GMT

Hi,

COORDINATOR_ONLY impala daemon will not perform data reads, rather it only coordinates query execution by distributing jobs to executors.

Why do you need to enable this for COORDINATOR_ONLY impala daemons? Or is it that all your COORDINATOR_ONLY impala daemons fail with this error?

Also, have you checked whether socket /var/run/hdfs-sockets/dn exist on that impala daemon host?

Cheers
Eric

Re: Impalad COORDINATOR_ONLY with HDFS Short-Circuit Read

anandnalya — Wed, 24 Jul 2019 11:45:19 GMT

Hi Eric,

> COORDINATOR_ONLY impala daemon will not perform data reads, rather it only coordinates query execution by distributing jobs to executors.

That's my understanding as well and that's why this failure was a surprise to me.

> Why do you need to enable this for COORDINATOR_ONLY impala daemons? Or is it that all your COORDINATOR_ONLY impala daemons fail with this error?

Yes, only the COORDINATOR_ONLY impala daemons fails. Other work fines.

> Also, have you checked whether socket /var/run/hdfs-sockets/dn exist on that impala daemon host?

> /var/run/hdfs-sockets/dn does not exists on COORDINATOR_ONLY impala daemon hosts as there is no DataNode on them.

Regards,

Anand

Re: Impalad COORDINATOR_ONLY with HDFS Short-Circuit Read

Lars Volker — Wed, 24 Jul 2019 18:26:50 GMT

This is a known issue and has been fixed in CM 6.2. Here is the relevant item in the release notes.

Cheers, Lars

Re: Impalad COORDINATOR_ONLY with HDFS Short-Circuit Read

EricL — Wed, 24 Jul 2019 23:52:03 GMT

Thanks Lars for pointing it out.

So solution is to disable HDFS shortcircuit read for coordinator only impala daemons:

a) create a new role group and add all coordinator only impala daemon hosts to this group

b) Go to "CM -> Cluster -> Impala services -> Configuration";

c) Add the following property into "Impala Daemon HDFS Advanced Configuration Snippet (Safety Valve)" for the new role group that you just created:

<property>
  <name>dfs.client.read.shortcircuit</name>
  <value>false</value>
</property>

d) Save the changes and restart the Impala Daemon instance that are affected.

Hope above can help.

Cheers
Eric

Re: Impalad COORDINATOR_ONLY with HDFS Short-Circuit Read

anandnalya — Thu, 25 Jul 2019 10:28:52 GMT

Got it working on my cluster. Thanks Lars and Eric.

Cheers,

Anand