[SPARK-4313][WebUI][Yarn] Fix link issue of the executor thread dump page in yarn-cluster mode

In yarn-cluster mode, the Web UI is running behind a yarn proxy server. Some features(or bugs?) of yarn proxy server will break the links for thread dump.

1. Yarn proxy server will do http redirect internally, so if opening `http://example.com:8088/cluster/app/application_1415344371838_0012/executors`, it will fetch `http://example.com:8088/cluster/app/application_1415344371838_0012/executors/` and return the content but won't change the link in the browser. Then when a user clicks `Thread Dump`, it will jump to `http://example.com:8088/proxy/application_1415344371838_0012/threadDump/?executorId=2`. This is a wrong link. The correct link should be `http://example.com:8088/proxy/application_1415344371838_0012/executors/threadDump/?executorId=2`.

Adding "/" to the tab links will fix it.

2. Yarn proxy server has a bug about the URL encode/decode. When a user accesses `http://example.com:8088/proxy/application_1415344371838_0006/executors/threadDump/?executorId=%3Cdriver%3E`, the yarn proxy server will require `http://example.com:36429/executors/threadDump/?executorId=%25253Cdriver%25253E`. But Spark web server expects `http://example.com:36429/executors/threadDump/?executorId=%3Cdriver%3E`. Related to [YARN-2844](https://issues.apache.org/jira/browse/YARN-2844).

For now, it's a tricky approach to bypass the yarn bug.

![threaddump](https://cloud.githubusercontent.com/assets/1000778/4972567/d1ccba64-68ad-11e4-983e-257530cef35a.png)

Author: zsxwing <zsxwing@gmail.com>

Closes #3183 from zsxwing/SPARK-4313 and squashes the following commits:

3379ca8 [zsxwing] Encode the executor id in the thread dump link and update the comment
abfa063 [zsxwing] Fix link issue of the executor thread dump page in yarn-cluster mode
This commit is contained in:
zsxwing 2014-11-14 13:36:13 -08:00 committed by Andrew Or
parent 5c265ccde0
commit 156cf3333d
3 changed files with 18 additions and 3 deletions

View file

@ -175,7 +175,7 @@ private[spark] object UIUtils extends Logging {
val shortAppName = if (appName.length < 36) appName else appName.take(32) + "..."
val header = activeTab.headerTabs.map { tab =>
<li class={if (tab == activeTab) "active" else ""}>
<a href={prependBaseUri(activeTab.basePath, "/" + tab.prefix)}>{tab.name}</a>
<a href={prependBaseUri(activeTab.basePath, "/" + tab.prefix + "/")}>{tab.name}</a>
</li>
}

View file

@ -17,6 +17,7 @@
package org.apache.spark.ui.exec
import java.net.URLDecoder
import javax.servlet.http.HttpServletRequest
import scala.util.Try
@ -29,7 +30,19 @@ private[ui] class ExecutorThreadDumpPage(parent: ExecutorsTab) extends WebUIPage
private val sc = parent.sc
def render(request: HttpServletRequest): Seq[Node] = {
val executorId = Option(request.getParameter("executorId")).getOrElse {
val executorId = Option(request.getParameter("executorId")).map {
executorId =>
// Due to YARN-2844, "<driver>" in the url will be encoded to "%25253Cdriver%25253E" when
// running in yarn-cluster mode. `request.getParameter("executorId")` will return
// "%253Cdriver%253E". Therefore we need to decode it until we get the real id.
var id = executorId
var decodedId = URLDecoder.decode(id, "UTF-8")
while (id != decodedId) {
id = decodedId
decodedId = URLDecoder.decode(id, "UTF-8")
}
id
}.getOrElse {
return Text(s"Missing executorId parameter")
}
val time = System.currentTimeMillis()

View file

@ -17,6 +17,7 @@
package org.apache.spark.ui.exec
import java.net.URLEncoder
import javax.servlet.http.HttpServletRequest
import scala.xml.Node
@ -139,8 +140,9 @@ private[ui] class ExecutorsPage(
</td>
{
if (threadDumpEnabled) {
val encodedId = URLEncoder.encode(info.id, "UTF-8")
<td>
<a href={s"threadDump/?executorId=${info.id}"}>Thread Dump</a>
<a href={s"threadDump/?executorId=${encodedId}"}>Thread Dump</a>
</td>
} else {
Seq.empty