[
https://issues.apache.org/jira/browse/SPARK-12482?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Kyle Sutton updated SPARK-12482:
--------------------------------
Description:
The issue of the file server using the default IP instead of the IP address
configured through {{spark.driver.host}} still exists in _Spark 1.5.2_
The problem is that, while the file server is listening on all ports on the
file server host, the _Spark_ service attempts to call back to the default port
of the host, to which it may or may not have connectivity.
For instance, the following setup causes a {{java.net.SocketTimeoutException}}
when the _Spark_ service tries to contact the _Spark_ driver host for a JAR:
* Launch host has a default IP of {{192.168.1.2}} and a secondary LAN
connection IP of {{172.30.0.2}}
* A _Spark_ connection is made from the launch host to a service running on
that LAN, {{172.30.0.0/16}}
** {{spark.driver.host}} is set to the IP of that _Spark_ service host
{{172.30.0.3}}
** {{spark.driver.port}} is set to {{50003}}
** {{spark.fileserver.port}} is set to {{50005}}
* Locally (on the launch host), the following listeners are active:
** {{0.0.0.0:50005}}
** {{172.30.0.2:50003}}
* The _Spark_ service calls back to the file server host for a JAR file using
the file server host's default IP: {{http://192.168.1.2:50005/jars/code.jar}}
* The _Spark_ service, being on a different network than the launch host,
cannot see the {{192.168.1.0/24}} address space, and fails to connect to the
file server
** A {{netstat}} on the _Spark_ service host will show the connection to the
file server host as being in {{SYN_SENT}} state until the process gives up
trying to connect
Given that the configured {{spark.driver.host}} must necessarily be accessible
by the _Spark_ service, it makes sense to reuse it for fileserver connections
and any other _Spark_ service -> driver host connections.
\\
----
\\
Original report follows:
I initially inquired about this here:
http://mail-archives.apache.org/mod_mbox/spark-user/201503.mbox/%3ccalq9kxcn2mwfnd4r4k0q+qh1ypwn3p8rgud1v6yrx9_05lv...@mail.gmail.com%3E
If the Spark driver host has multiple IPs and spark.driver.host is set to one
of them, I would expect the fileserver to start on the same IP. I checked
HttpServer and the jetty Server is started the default IP of the machine:
https://github.com/apache/spark/blob/master/core/src/main/scala/org/apache/spark/HttpServer.scala#L75
Something like this might work instead:
{code:title=HttpServer.scala#L75}
val server = new Server(new InetSocketAddress(conf.get("spark.driver.host"), 0))
{code}
was:
The issue of the file server using the default IP instead of the IP address
configured through {{spark.driver.host}} still exists in _Spark 1.5.2_
The problem is that, while the file server is listening on all ports on the
file server host, the _Spark_ service attempts to call back to the default port
of the host, to which it may or may not have connectivity.
For instance, the following setup causes a {{java.net.SocketTimeoutException}}
when the _Spark_ service tries to contact the _Spark_ driver host for a JAR:
* Launch host has a default IP of {{192.168.1.2}} and a secondary LAN
connection IP of {{172.30.0.2}}
* A _Spark_ connection is made from the launch host to a service running on
that LAN, {{172.30.0.0/16}}
** {{spark.driver.host}} is set to the IP of that _Spark_ service host
{{172.30.0.3}}
** {{spark.driver.port}} is set to {{50003}}
** {{spark.fileserver.port}} is set to {{50005}}
* Locally (on the launch host), the following listeners are active:
** {{0.0.0.0:50005}}
** {{172.30.0.2:50003}}
* The _Spark_ service calls back to the file server host for a JAR file using
the file server host's default IP: {{http://192.168.1.2:50005/jars/code.jar}}
* The _Spark_ service, being on a different network than the launch host,
cannot see the {{192.168.1.0/24}} address space, and fails to connect to the
file server
** A {{netstat}} on the _Spark_ service host will show the connection to the
file server host as being in {{SYN_SENT}} state until the process gives up
trying to connect
Given that the configured {{spark.driver.host}} must necessarily be accessible
by the _Spark_ service, it makes sense to reuse it for fileserver connections
and any other _Spark_ service -> driver host connections.
----
Original report follows:
I initially inquired about this here:
http://mail-archives.apache.org/mod_mbox/spark-user/201503.mbox/%3ccalq9kxcn2mwfnd4r4k0q+qh1ypwn3p8rgud1v6yrx9_05lv...@mail.gmail.com%3E
If the Spark driver host has multiple IPs and spark.driver.host is set to one
of them, I would expect the fileserver to start on the same IP. I checked
HttpServer and the jetty Server is started the default IP of the machine:
https://github.com/apache/spark/blob/master/core/src/main/scala/org/apache/spark/HttpServer.scala#L75
Something like this might work instead:
{code:title=HttpServer.scala#L75}
val server = new Server(new InetSocketAddress(conf.get("spark.driver.host"), 0))
{code}
> CLONE - Spark fileserver not started on same IP as using spark.driver.host
> --------------------------------------------------------------------------
>
> Key: SPARK-12482
> URL: https://issues.apache.org/jira/browse/SPARK-12482
> Project: Spark
> Issue Type: Bug
> Components: Spark Core
> Affects Versions: 1.2.1, 1.5.2
> Reporter: Kyle Sutton
>
> The issue of the file server using the default IP instead of the IP address
> configured through {{spark.driver.host}} still exists in _Spark 1.5.2_
> The problem is that, while the file server is listening on all ports on the
> file server host, the _Spark_ service attempts to call back to the default
> port of the host, to which it may or may not have connectivity.
> For instance, the following setup causes a
> {{java.net.SocketTimeoutException}} when the _Spark_ service tries to contact
> the _Spark_ driver host for a JAR:
> * Launch host has a default IP of {{192.168.1.2}} and a secondary LAN
> connection IP of {{172.30.0.2}}
> * A _Spark_ connection is made from the launch host to a service running on
> that LAN, {{172.30.0.0/16}}
> ** {{spark.driver.host}} is set to the IP of that _Spark_ service host
> {{172.30.0.3}}
> ** {{spark.driver.port}} is set to {{50003}}
> ** {{spark.fileserver.port}} is set to {{50005}}
> * Locally (on the launch host), the following listeners are active:
> ** {{0.0.0.0:50005}}
> ** {{172.30.0.2:50003}}
> * The _Spark_ service calls back to the file server host for a JAR file using
> the file server host's default IP: {{http://192.168.1.2:50005/jars/code.jar}}
> * The _Spark_ service, being on a different network than the launch host,
> cannot see the {{192.168.1.0/24}} address space, and fails to connect to the
> file server
> ** A {{netstat}} on the _Spark_ service host will show the connection to the
> file server host as being in {{SYN_SENT}} state until the process gives up
> trying to connect
> Given that the configured {{spark.driver.host}} must necessarily be
> accessible by the _Spark_ service, it makes sense to reuse it for fileserver
> connections and any other _Spark_ service -> driver host connections.
> \\
> ----
> \\
> Original report follows:
> I initially inquired about this here:
> http://mail-archives.apache.org/mod_mbox/spark-user/201503.mbox/%3ccalq9kxcn2mwfnd4r4k0q+qh1ypwn3p8rgud1v6yrx9_05lv...@mail.gmail.com%3E
> If the Spark driver host has multiple IPs and spark.driver.host is set to one
> of them, I would expect the fileserver to start on the same IP. I checked
> HttpServer and the jetty Server is started the default IP of the machine:
> https://github.com/apache/spark/blob/master/core/src/main/scala/org/apache/spark/HttpServer.scala#L75
> Something like this might work instead:
> {code:title=HttpServer.scala#L75}
> val server = new Server(new InetSocketAddress(conf.get("spark.driver.host"),
> 0))
> {code}
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
