[jira] [Commented] (HADOOP-13884) s3a create(overwrite=true) to only look for dir/ and list entries, not file
[
https://issues.apache.org/jira/browse/HADOOP-13884?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16837281#comment-16837281
]
Steve Loughran commented on HADOOP-13884:
-
Revisiting:
#. sometimes that negative entry can last minutes rather than a few seconds.
# we also do existence checks on rename, so the same risk can occur there: you
copy to a destination and the new file isn't immediately visible.
Proposed: a new operation checkIsNotDirectory(key): boolean. which returns true
if there is either nothing at the far end or a file, that is "no directory
entry"
we can have a sequence of
* S3Guard: check the store to make sure it's not a directory entry.
* S3: HEAD key + "/"; LIST key.
*
> s3a create(overwrite=true) to only look for dir/ and list entries, not file
> ---
>
> Key: HADOOP-13884
> URL: https://issues.apache.org/jira/browse/HADOOP-13884
> Project: Hadoop Common
> Issue Type: Sub-task
> Components: fs/s3
>Affects Versions: 2.9.0
>Reporter: Steve Loughran
>Priority: Minor
>
> before doing a create(), s3a does a getFileStatus() to make sure there isn't
> a directory there, and, if overwrite=false, that there isn't a file.
> Because S3 caches negative HEAD/GET requests, if there isn't a file, then
> even after the PUT, a later GET/HEAD may return 404; we are generating create
> consistency where none need exist.
> when overwrite=true we don't care whether the file exists or not, only that
> the path isn't a directory. So we can just do the HEAD path +"/' and the LIST
> calls, skipping the {{HEAD path}}. This will save an HTTP round trip of a few
> hundred millis, and ensure that there's no 404 cached in the S3 front end for
> later callers
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
-
To unsubscribe, e-mail: common-issues-unsubscr...@hadoop.apache.org
For additional commands, e-mail: common-issues-h...@hadoop.apache.org
[jira] [Commented] (HADOOP-13884) s3a create(overwrite=true) to only look for dir/ and list entries, not file
[
https://issues.apache.org/jira/browse/HADOOP-13884?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16360726#comment-16360726
]
Steve Loughran commented on HADOOP-13884:
-
* Moot for S3Guard
* For raw S3, that initial HEAD $path is faster than doing a list, so if you
are overwriting a file you are doing a HEAD $path/ and a LIST path (two calls,
one slower), and there'll be no negative caching anyway.
We'd be increasing the cost of overwriting an existing file, but reducing the
cost of creating a new file and risk of caching 404, so in some codepaths
things would get better, some would get worse
The safest way to handle this would be add a hint in file creation to say "this
create doesn't expect the dest to exist" vs "this create expects a file at the
dest" or even "don't bother with any exists checks at all", which is what is
secretly done in the commit code via WriteOperationsHelper offering a put() API
call.
> s3a create(overwrite=true) to only look for dir/ and list entries, not file
> ---
>
> Key: HADOOP-13884
> URL: https://issues.apache.org/jira/browse/HADOOP-13884
> Project: Hadoop Common
> Issue Type: Sub-task
> Components: fs/s3
>Affects Versions: 2.9.0
>Reporter: Steve Loughran
>Priority: Minor
>
> before doing a create(), s3a does a getFileStatus() to make sure there isn't
> a directory there, and, if overwrite=false, that there isn't a file.
> Because S3 caches negative HEAD/GET requests, if there isn't a file, then
> even after the PUT, a later GET/HEAD may return 404; we are generating create
> consistency where none need exist.
> when overwrite=true we don't care whether the file exists or not, only that
> the path isn't a directory. So we can just do the HEAD path +"/' and the LIST
> calls, skipping the {{HEAD path}}. This will save an HTTP round trip of a few
> hundred millis, and ensure that there's no 404 cached in the S3 front end for
> later callers
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
-
To unsubscribe, e-mail: common-issues-unsubscr...@hadoop.apache.org
For additional commands, e-mail: common-issues-h...@hadoop.apache.org
