[
https://issues.apache.org/jira/browse/SPARK-58748?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Chao Sun updated SPARK-58748:
-----------------------------
Description:
Spark Kubernetes client-mode drivers commonly bind RPC sockets to 0.0.0.0 or an
IPv6 wildcard while advertising a routable address through spark.driver.host.
When spark.kubernetes.executor.useDriverPodIP=true, SparkContext replaces
spark.driver.host with spark.driver.bindAddress, and BasicExecutorFeatureStep
independently uses that bind address to construct SPARK_DRIVER_URL. This
produces executor URLs such as spark://[email protected]:7078, so
executors cannot register with the driver.
The option was introduced in Spark 4.1 and is enabled by default in Spark 4.3
and current master, making client-mode and Spark Connect drivers vulnerable
without explicitly enabling the setting.
Reproduction: launch a Kubernetes client-mode driver with
spark.driver.bindAddress=0.0.0.0, spark.driver.host set to a routable driver
address, and spark.kubernetes.executor.useDriverPodIP=true. Observe executor
connection failures against 0.0.0.0.
Expected: preserve the advertised spark.driver.host whenever
spark.driver.bindAddress is an IPv4 or IPv6 wildcard; continue using and
normalizing concrete driver Pod IPs. Both SparkContext and executor driver URL
construction need the same wildcard guard.
Affected versions: 4.1.x and 4.2.x when explicitly enabled; 4.3.0 and 5.0.0 by
default. Prioritize a 4.3 backport.
was:SPARK-53944 added {{spark.kubernetes.executor.useDriverPodIP}} to let
Kubernetes executors connect directly to the driver Pod IP. SPARK-57350 enabled
this behavior by default in Spark 4.3.\n\nSome valid Kubernetes deployments,
including Spark Connect servers and other client-mode drivers, bind their RPC
server to all local interfaces while explicitly advertising a separate routable
driver
address:\n\n{code}\nspark.master=k8s://https://kubernetes.example:6443\nspark.driver.bindAddress=0.0.0.0\nspark.driver.host=10.129.36.37\nspark.kubernetes.executor.useDriverPodIP=true\n{code}\n\nBoth
{{SparkContext}} and {{BasicExecutorFeatureStep}} currently substitute
{{spark.driver.bindAddress}} for the advertised {{spark.driver.host}} when
{{useDriverPodIP}} is enabled. This overwrites the routable advertised address
with {{0.0.0.0}} and constructs executor RPC URLs such
as:\n\n{code}\nspark://[email protected]:7078\n{code}\n\nExecutors
cannot register with that unspecified address, so they repeatedly fail and the
application cannot make progress. The same problem applies to IPv6 wildcard
forms such as {{::}}, {{[::]}}, and {{0:0:0:0:0:0:0:0}}.\n\nSpark already
documents {{spark.driver.bindAddress}} and {{spark.driver.host}} as separate
bind and advertised addresses. The driver Pod IP optimization should continue
to use concrete IPv4 and IPv6 bind addresses, including the IPv6 normalization
fixed by SPARK-58719, but should preserve the configured advertised driver host
whenever the bind address is an unspecified/wildcard address. Both driver
address selection sites must agree.\n\nSpark 4.3 is affected by default; Spark
4.1 and 4.2 are affected when users explicitly enable the option.
> Preserve the advertised driver host when the Kubernetes driver bind address
> is a wildcard
> -----------------------------------------------------------------------------------------
>
> Key: SPARK-58748
> URL: https://issues.apache.org/jira/browse/SPARK-58748
> Project: Spark
> Issue Type: Bug
> Components: Connect, Kubernetes, Spark Core
> Affects Versions: 4.1.0, 4.2.0, 4.3.0, 5.0.0
> Reporter: Chao Sun
> Assignee: Chao Sun
> Priority: Major
> Fix For: 4.3.0, 5.0.0
>
>
> Spark Kubernetes client-mode drivers commonly bind RPC sockets to 0.0.0.0 or
> an IPv6 wildcard while advertising a routable address through
> spark.driver.host. When spark.kubernetes.executor.useDriverPodIP=true,
> SparkContext replaces spark.driver.host with spark.driver.bindAddress, and
> BasicExecutorFeatureStep independently uses that bind address to construct
> SPARK_DRIVER_URL. This produces executor URLs such as
> spark://[email protected]:7078, so executors cannot register
> with the driver.
> The option was introduced in Spark 4.1 and is enabled by default in Spark 4.3
> and current master, making client-mode and Spark Connect drivers vulnerable
> without explicitly enabling the setting.
> Reproduction: launch a Kubernetes client-mode driver with
> spark.driver.bindAddress=0.0.0.0, spark.driver.host set to a routable driver
> address, and spark.kubernetes.executor.useDriverPodIP=true. Observe executor
> connection failures against 0.0.0.0.
> Expected: preserve the advertised spark.driver.host whenever
> spark.driver.bindAddress is an IPv4 or IPv6 wildcard; continue using and
> normalizing concrete driver Pod IPs. Both SparkContext and executor driver
> URL construction need the same wildcard guard.
> Affected versions: 4.1.x and 4.2.x when explicitly enabled; 4.3.0 and 5.0.0
> by default. Prioritize a 4.3 backport.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
