[jira] [Commented] (HADOOP-14237) S3A Support Shared Instance Profile Credentials Across All Hadoop Nodes

["Steve Loughran (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/HADOOP-14237?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16558788#comment-16558788 ] 

Steve Loughran commented on HADOOP-14237:
-----------------------------------------

Revisiting this

* what was the use case which triggered this. You were trying to log in and things were failing?
* And this was in EC2, so you were using the IAM role aut, which does an HTTP Get?
* Can we have a stack trace?

We could address this by having our own subclass of the {{InstanceProfileCredentialsProvider}} whose getCredentials retries on whatever error gets raised by the service. That'd have to be a very different retry policy from {{S3ARetryPolicy}}, which tries to reconnect on network/connection refused. We will want failfast there. 

All I really need to know is the error raised & error text, and we can recover from failures here with retry & backoff.

Looks like com.amazonaws.retry.RetryUtils has a predicate to see if an exception is for throttling. If we use that in translateException, it'll work



> S3A Support Shared Instance Profile Credentials Across All Hadoop Nodes
> -----------------------------------------------------------------------
>
>                 Key: HADOOP-14237
>                 URL: https://issues.apache.org/jira/browse/HADOOP-14237
>             Project: Hadoop Common
>          Issue Type: Sub-task
>          Components: fs/s3
>    Affects Versions: 2.8.0, 3.0.0-alpha1, 3.0.0-alpha2, 2.8.1
>         Environment: EC2, AWS
>            Reporter: Kazuyuki Tanimura
>            Assignee: Kazuyuki Tanimura
>            Priority: Major
>
> When I run a large Hadoop cluster on EC2 instances with IAM Role, it fails getting the instance profile credentials, eventually all jobs on the cluster fail. Since a number of S3A clients (all mappers and reducers) try to get the credentials, the AWS credential endpoint starts responding 5xx and 4xx error codes.
> SharedInstanceProfileCredentialsProvider.java is sort of trying to solve it, but it still does not share the credentials with other EC2 nodes / JVM processes.
> This issue prevents users from creating Hadoop clusters on EC2



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)

---------------------------------------------------------------------
To unsubscribe, e-mail: common-issues-unsubscribe@hadoop.apache.org
For additional commands, e-mail: common-issues-help@hadoop.apache.org