[
https://issues.apache.org/jira/browse/HDFS-11804?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16032164#comment-16032164
]
Xiao Chen commented on HDFS-11804:
----------------------------------
Thanks [~shahrs87] for proposing the idea and a demonstrative patch. +1 on
adding client side retries.
Some comments:
- Should we also treat AuthenticationException as non-retry?
- Do we want retry policy's {{maxRetries}} also be configurable? Any reason
hard-code that to 0?
- generateEncryptedKey, reencryptEncryptedKey I think are idempotent.
addDelegationTokens should be Idempotent (because it's really
getDelegationTokens, so similar to hdfs's {{ClientProtocol}} IMHO.
- retry: m2c is we retry immediately to give each server a chance. If
multiplier > 1 (say, numFailovers > providers.length), then we retry with
delays. (Related, I think KMS's exceptions are less complicated than NN
exception, mostly because KMSCP's {{validateResponse}} throws just IOExceptions
in many cases.)
> KMS client needs retry logic
> ----------------------------
>
> Key: HDFS-11804
> URL: https://issues.apache.org/jira/browse/HDFS-11804
> Project: Hadoop HDFS
> Issue Type: Bug
> Affects Versions: 2.6.0
> Reporter: Rushabh S Shah
> Assignee: Rushabh S Shah
> Attachments: HDFS-11804-trunk-1.patch, HDFS-11804-trunk.patch
>
>
> The kms client appears to have no retry logic – at all. It's completely
> decoupled from the ipc retry logic. This has major impacts if the KMS is
> unreachable for any reason, including but not limited to network connection
> issues, timeouts, the +restart during an upgrade+.
> This has some major ramifications:
> # Jobs may fail to submit, although oozie resubmit logic should mask it
> # Non-oozie launchers may experience higher rates if they do not already have
> retry logic.
> # Tasks reading EZ files will fail, probably be masked by framework reattempts
> # EZ file creation fails after creating a 0-length file – client receives
> EDEK in the create response, then fails when decrypting the EDEK
> # Bulk hadoop fs copies, and maybe distcp, will prematurely fail
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
---------------------------------------------------------------------
To unsubscribe, e-mail: hdfs-issues-unsubscr...@hadoop.apache.org
For additional commands, e-mail: hdfs-issues-h...@hadoop.apache.org
