[
https://issues.apache.org/jira/browse/SPARK-5113?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Patrick Wendell updated SPARK-5113:
-----------------------------------
Description:
Spark has multiple network components that start servers and advertise their
network addresses to other processes.
We should go through each of these components and make sure they have
consistent and/or documented behavior wrt (a) what interface(s) they bind to
and (b) what hostname they use to advertise themselves to other processes. We
should document this clearly and explain to people what to do in different
cases (e.g. EC2, dockerized containers, etc).
When Spark initializes, it will search for a network interface until it finds
one that is not a loopback address. Then it will do a reverse DNS lookup for a
hostname associated with that interface. Then the network components will use
that hostname to advertise the component to other processes. That hostname is
also the one used for the akka system identifier (akka supports only supplying
a single name which it uses both as the bind interface and as the actor
identifier). In some cases, that hostname is used as the bind hostname also
(e.g. I think this happens in the connection manager and possibly akka) - which
will likely internally result in a re-resolution of this to an IP address. In
other cases (the web UI and netty shuffle) we seem to bind to all interfaces.
The best outcome would be to have three configs that can be set on each machine:
{code}
SPARK_LOCAL_IP # Ip address we bind to for all services
SPARK_INTERNAL_HOSTNAME # Hostname we advertise to remote processes within the
cluster
SPARK_EXTERNAL_HOSTNAME # Hostname we advertise to processes outside the
cluster (e.g. the UI)
{code}
It's not clear how easily we can support that scheme while providing backwards
compatibility. The last one (SPARK_EXTERNAL_HOSTNAME) is easy - it's just an
alias for what is now SPARK_PUBLIC_DNS.
was:
Spark has multiple network components that start servers and advertise their
network addresses to other processes.
We should go through each of these components and make sure they have
consistent and/or documented behavior wrt (a) what interface(s) they bind to
and (b) what hostname they use to advertise themselves to other processes. We
should document this clearly and explain to people what to do in different
cases (e.g. EC2, dockerized containers, etc).
When Spark initializes, it will search for a network interface until it finds
one that is not a loopback address. Then it will do a reverse DNS lookup for a
hostname associated with that interface. Then the network components will use
that hostname to advertise the component to other processes. That hostname is
also the one used for the akka system identifier (akka supports only supplying
a single name which it uses both as the bind interface and as the actor
identifier). In some cases, that hostname is used as the bind hostname also
(e.g. I think this happens in the connection manager and possibly akka) - which
will likely internally result in a re-resolution of this to an IP address. In
other cases (the web UI and netty shuffle) we seem to bind to all interfaces.
> Audit and document use of hostnames and IP addresses in Spark
> -------------------------------------------------------------
>
> Key: SPARK-5113
> URL: https://issues.apache.org/jira/browse/SPARK-5113
> Project: Spark
> Issue Type: Bug
> Reporter: Patrick Wendell
> Priority: Critical
>
> Spark has multiple network components that start servers and advertise their
> network addresses to other processes.
> We should go through each of these components and make sure they have
> consistent and/or documented behavior wrt (a) what interface(s) they bind to
> and (b) what hostname they use to advertise themselves to other processes. We
> should document this clearly and explain to people what to do in different
> cases (e.g. EC2, dockerized containers, etc).
> When Spark initializes, it will search for a network interface until it finds
> one that is not a loopback address. Then it will do a reverse DNS lookup for
> a hostname associated with that interface. Then the network components will
> use that hostname to advertise the component to other processes. That
> hostname is also the one used for the akka system identifier (akka supports
> only supplying a single name which it uses both as the bind interface and as
> the actor identifier). In some cases, that hostname is used as the bind
> hostname also (e.g. I think this happens in the connection manager and
> possibly akka) - which will likely internally result in a re-resolution of
> this to an IP address. In other cases (the web UI and netty shuffle) we seem
> to bind to all interfaces.
> The best outcome would be to have three configs that can be set on each
> machine:
> {code}
> SPARK_LOCAL_IP # Ip address we bind to for all services
> SPARK_INTERNAL_HOSTNAME # Hostname we advertise to remote processes within
> the cluster
> SPARK_EXTERNAL_HOSTNAME # Hostname we advertise to processes outside the
> cluster (e.g. the UI)
> {code}
> It's not clear how easily we can support that scheme while providing
> backwards compatibility. The last one (SPARK_EXTERNAL_HOSTNAME) is easy -
> it's just an alias for what is now SPARK_PUBLIC_DNS.
--
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
