[SPARK-46977][CORE] A failed request to obtain a token from one NameNode should not skip subsequent token requests#45
Open
giggsoff wants to merge 1 commit into
Open
Conversation
…ode should not skip subsequent token requests
### What changes were proposed in this pull request?
This PR enhances the `HadoopFSDelegationTokenProvider` to tolerate failures when fetching tokens from multiple NameNodes.
### Why are the changes needed?
Let's say we are going to access 3 HDFS, `nn-1`, `nn-2`, `nn-3` in YARN cluster mode with TGT cache, while the `nn-1` is the `defaultFs` which is used by YARN to store aggregated logs, and there are issues in `nn-2` which can not issue the token.
```
spark-submit \
--master yarn \
--deployMode cluster \
--conf spark.kerberos.access.hadoopFileSystems=hdfs://nn-1,hdfs://nn-2,hdfs://nn-3 \
...
```
During the submitting phase, Spark is going to call `HadoopFSDelegationTokenProvider` to fetch tokens from all declared NameNodes one by one, in **indeterminate** order (`HadoopFSDelegationTokenProvider.hadoopFSsToAccess` process and return a `Set[FileSystem]`), so the order may not respect the user declared order in `spark.kerberos.access.hadoopFileSystems`.
If the order is [`nn-1`, `nn-2`, `nn-3`], then we are going to request a token from `nn-1` successfully, but fail for `nn-2` with the below error, the left `nn-3` is going to be skipped. But such failure WON'T block the whole submitting process, the Spark app is going to be submitted with only `nn-1` token.
```
2024-01-03 12:41:36 [WARN] [main] org.apache.spark.deploy.security.HadoopFSDelegationTokenProvider#94 - Failed to get token from service hadoopfs
org.apache.hadoop.ipc.RemoteException: <Some Error Message>
at org.apache.hadoop.ipc.Client.getRpcResponse(Client.java:1507) ~[hadoop-common-2.9.2.2.jar:?]
...
at org.apache.hadoop.hdfs.DistributedFileSystem.addDelegationTokens(DistributedFileSystem.java:2604) ~[hadoop-hdfs-client-2.9.2.2.jar:?]
at org.apache.spark.deploy.security.HadoopFSDelegationTokenProvider.$anonfun$fetchDelegationTokens$1(HadoopFSDelegationTokenProvider.scala:122) ~[spark-core_2.12-3.3.1.27.jar:3.3.1.27]
at scala.collection.immutable.HashSet$HashSet1.foreach(HashSet.scala:335) ~[scala-library-2.12.15.jar:?]
at scala.collection.immutable.HashSet$HashTrieSet.foreach(HashSet.scala:1111) ~[scala-library-2.12.15.jar:?]
at scala.collection.immutable.HashSet$HashTrieSet.foreach(HashSet.scala:1111) ~[scala-library-2.12.15.jar:?]
at org.apache.spark.deploy.security.HadoopFSDelegationTokenProvider.fetchDelegationTokens(HadoopFSDelegationTokenProvider.scala:115) ~[spark-core_2.12-3.3.1.27.jar:3.3.1.27]
...
at org.apache.spark.deploy.security.HadoopDelegationTokenManager.obtainDelegationTokens(HadoopDelegationTokenManager.scala:146) ~[spark-core_2.12-3.3.1.27.jar:3.3.1.27]
at org.apache.spark.deploy.yarn.Client.setupSecurityToken(Client.scala:352) ~[spark-yarn_2.12-3.3.1.27.jar:3.3.1.27]
at org.apache.spark.deploy.yarn.Client.createContainerLaunchContext(Client.scala:1140) ~[spark-yarn_2.12-3.3.1.27.jar:3.3.1.27]
...
at org.apache.spark.deploy.SparkSubmit.main(SparkSubmit.scala) ~[spark-core_2.12-3.3.1.27.jar:3.3.1.27]
```
when the Spark app access `nn-2` and `nn-3`, it will fail with `o.a.h.security.AccessControlException: Client cannot authenticate via:[TOKEN, KERBEROS]`
Things become worse if the FS order is [`nn-3`, `nn-2`, `nn-1`], the Spark app will be submitted to YARN with only `nn-3` token, it even has no chance to allow NodeManager to upload aggregated logs after the application exit because it requires the app to provide a token to access `nn-1`.
the log from NodeManager
```
2024-01-03 08:08:14,028 [3173570620] - WARN [NM ContainerManager dispatcher:Client$Connection1$772] - Exception encountered while connecting to the server
Caused by: org.apache.hadoop.security.AccessControlException: Client cannot authenticate via:[TOKEN, KERBEROS]
at org.apache.hadoop.security.SaslRpcClient.selectSaslClient(SaslRpcClient.java:179)
at org.apache.hadoop.security.SaslRpcClient.saslConnect(SaslRpcClient.java:392)
...
at org.apache.hadoop.hdfs.DistributedFileSystem.getFileStatus(DistributedFileSystem.java:1768)
...
at org.apache.hadoop.yarn.server.nodemanager.containermanager.logaggregation.LogAggregationService.createAppDir(LogAggregationService.java:404)
at org.apache.hadoop.yarn.server.nodemanager.containermanager.logaggregation.LogAggregationService.initAppAggregator(LogAggregationService.java:273)
...
```
Without the logs, we even don't know what happened.
Eventually, due to the **indeterminate** order of NameNodes to request tokens, such a Job sometimes submitted successfully and sometimes failed without logs.
<img width="1903" alt="image" src="https://github.com/apache/spark/assets/26535726/7ca5c871-6399-4eae-b689-d6d741c1c373">
### Does this PR introduce _any_ user-facing change?
Yes, when the user configures `spark.kerberos.access.hadoopFileSystems` to access multiple Kerberized HDFS, and one or more NameNodes have issues, tokens are always fetched from the rest health NameNodes after this patch.
### How was this patch tested?
Tested in internal Kerberized cluster.
### Was this patch authored or co-authored using generative AI tooling?
No.
Closes apache#45030 from pan3793/SPARK-46977.
Authored-by: Cheng Pan <chengpan@apache.org>
Signed-off-by: Dongjoon Hyun <dhyun@apple.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What changes were proposed in this pull request?
This PR enhances the
HadoopFSDelegationTokenProviderto tolerate failures when fetching tokens from multiple locations.It is backport from 4.0 version.