[
https://issues.apache.org/jira/browse/HDFS-17790?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17952770#comment-17952770
]
SUYEON LEE commented on HDFS-17790:
-----------------------------------
YARN-11821
> Support pattern matching for YARN proxy addresses.
> --------------------------------------------------
>
> Key: HDFS-17790
> URL: https://issues.apache.org/jira/browse/HDFS-17790
> Project: Hadoop HDFS
> Issue Type: Bug
> Components: router
> Affects Versions: 3.4.1
> Reporter: SUYEON LEE
> Priority: Major
> Attachments: 스크린샷 2025-02-20 오후 6.14.27.png
>
>
> *Issue:*
> In a YARN Federation environment utilizing multiple YARN Routers, if a Global
> Server Load Balancer (GSLB) is used for distributing requests to these
> Routers _without_ sticky session support, users encounter redirection errors.
> !스크린샷 2025-02-20 오후 6.14.27.png!
> These errors typically manifest when accessing application UIs or logs via
> the NodeManager's proxy.
> *Problem Details:*
> The root cause of this issue lies in how the NodeManager (NM) resolves the
> YARN Router address for proxying web requests. The NodeManager uses the
> {{yarn.web-proxy.address}} configuration property to determine the Router's
> address. However, this property is designed to accommodate only a single
> hostname or IP address.
> When a GSLB without sticky sessions distributes requests, an initial request
> might be routed to Router A, while a subsequent request (e.g., for
> application logs, proxied by an NM) might be directed by the GSLB to Router
> B, or the NM might be configured with a specific Router/GSLB VIP that
> resolves to Router B. If the ApplicationMaster (AM) for the application in
> question is registered with Router A, Router B may not be able to correctly
> proxy or redirect the request, leading to errors. The NM's attempt to proxy
> through an "incorrect" or context-unaware Router (relative to where the AM is
> known) due to the single-host limitation in {{yarn.web-proxy.address}}
> results in these failures.
> *Analysis / Root Cause Summary:*
> # *Environment:* YARN Federation with multiple YARN Routers.
> # *Load Balancing:* GSLB used without sticky session support, leading to
> requests for the same session potentially hitting different Routers.
> # *NM Proxy Mechanism:* NodeManagers use {{yarn.web-proxy.address}} to
> identify the YARN Router for proxying UI/log requests.
> # *Configuration Limitation:* {{yarn.web-proxy.address}} supports only a
> single host, making it incompatible with a multi-Router setup where requests
> are not sticky.
> # *Symptom:* Redirection errors occur as the NM may attempt to proxy through
> a Router that does not have the necessary context for the specific
> application/AM, or is not the Router the user was initially interacting with
> via the GSLB.
> *Proposed Solution:*
> To address this limitation, a patch has been developed to introduce
> *pattern-based address matching for the YARN web proxy functionality.*
> This enhancement will allow {{yarn.web-proxy.address.pattern}} to be
> configured with a pattern or a list of Routers. The NodeManager's proxy
> mechanism can then intelligently select or match the appropriate Router
> address, ensuring that requests are proxied correctly even in a federated
> environment with multiple Routers and non-sticky load balancing.
> *Goal of the Patch:*
> * Enable robust YARN web proxying in federated environments with multiple
> Routers load-balanced by a non-sticky GSLB.
> * Improve the usability and reliability of accessing application UIs and
> logs in such configurations.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: hdfs-issues-unsubscr...@hadoop.apache.org
For additional commands, e-mail: hdfs-issues-h...@hadoop.apache.org
