[
https://issues.apache.org/jira/browse/HADOOP-16279?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16838753#comment-16838753
]
Gabor Bota edited comment on HADOOP-16279 at 5/13/19 5:56 PM:
--------------------------------------------------------------
Thank you for reviewing this [~fabbri]!
With HADOOP-15999 we changed a bit how authoritative mode works. In
{{S3AFileSystem#innerGetFileStatus}} if {{allowAuthoritative}} is false then it
will send a HEAD request - so query S3 for the {{S3AFileStatus}} and compare
the mod_time. If it's newer than we update the MS and return the newer metadata.
With the current issue, I want to add the behavior to query S3 if a defined TTL
for the metadata is expired in the MS. In my PR I use the same TTL for
authoritative directories and all other entries. This is a simplification
because you don't have to deal with 2 different parameters for expiry.
----
_| why we need more prune() functions added to the MS interface_
That prune is for removing expired entries from the ddbms. It uses
{{last_updated}} for expiry rather than {{mod_time}}.
The main reason to add this is that for some operation (e.g directory rename
with lots of files ([~ste...@apache.org])) it can significantly improve
performance if we remove all expired tombstone entries before the operation.
Maybe a better solution for this to add an enum to the current prune's
parameter to select the internal prune query for the MSs?
_| So the patch made sense except the naming and description of the
configuration parameter_
The {{MetadataStore should be authoritative to use this configuration knob.}}
should be removed from the configuration. It was my mistake that I left it
there, and not intended.
_| LocalMetadataStore use of guava Cache meant the work was already done there?_
This is an expiry on a different level. I don't use the DB's expiry in dynamo,
so I think we should do the same for {{LocalMetadataStore}}.
_| smarter logic that allows you set a policy for handling S3 versus MS
conflicts_
So basically what you mean is to add a conflict resolution algorithm when an
entry is expired? That would be interesting, I can think of three solutions:
* skip checking the expiry, so TTL metadata expiry is turned off
* return null on expiry from the MS, so we will get the fresh metadata from S3
and update the MS
* check if mod_time is higher on MS vs S3, and then return the newer one
----
I think this expiry and authoritative directory listing could still live next
to each other:
* The dir listing can be authoritative if we get the full listing, but not
otherwise. This is good, so we don't have to query a dir listing, the
performance improvement is there
* The dir can lose authoritativeness in certain circumstances (like in
{{org.apache.hadoop.fs.s3a.s3guard.LocalMetadataStore#removeAuthoritativeFromParent}})
* {{org.apache.hadoop.fs.s3a.s3guard.S3Guard#listChildrenWithTtl}} will only
mean anything if a dir listing's auth is expired *during* an operation (and not
start of a request), which is very unlikely.
* but if we add policies like "{{TTL metadata expiry is turned off}}", then the
expiry of auth dirs will still have a meaning, and we don't have to add another
config parameter.
was (Author: gabor.bota):
Thank you for reviewing this [~fabbri]!
With HADOOP-15999 we changed a bit how authoritative mode works. In
{{S3AFileSystem#innerGetFileStatus}} if {{allowAuthoritative}} is false then it
will send a HEAD request - so query S3 for the {{S3AFileStatus}} and compare
the mod_time. If it's newer than we update the MS and return the newer metadata.
With the current issue, I want to add the behavior to query S3 if a defined TTL
for the metadata is expired in the MS. In my PR I use the same TTL for
authoritative directories and all other entries. This is a simplification
because you don't have to deal with 2 different parameters for expiry.
----
_| why we need more prune() functions added to the MS interface_
That prune is for removing expired entries from the ddbms. It uses
{{last_updated}} for expiry rather than {{mod_time}}.
The main reason to add this is that for some operation (e.g directory rename
with lots of files ([~ste...@apache.org])) it can significantly improve
performance if we remove all expired tombstone entries before the operation.
Maybe a better solution for this to add an enum to the parameter to select the
internal prune query for the MSs?
_| So the patch made sense except the naming and description of the
configuration parameter_
The {{MetadataStore should be authoritative to use this configuration knob.}}
should be removed from the configuration. It was my mistake that I left it
there, and not intended.
_| LocalMetadataStore use of guava Cache meant the work was already done there?_
This is an expiry on a different level. I don't use the DB's expiry in dynamo,
so I think we should do the same for {{LocalMetadataStore}}.
_| smarter logic that allows you set a policy for handling S3 versus MS
conflicts_
So basically what you mean is to add a conflict resolution algorithm when an
entry is expired? That would be interesting, I can think of three solutions:
* skip checking the expiry, so TTL metadata expiry is turned off
* return null on expiry from the MS, so we will get the fresh metadata from S3
and update the MS
* check if mod_time is higher on MS vs S3, and then return the newer one
----
I think this expiry and authoritative directory listing could still live next
to each other:
* The dir listing can be authoritative if we get the full listing, but not
otherwise. This is good, so we don't have to query a dir listing, the
performance improvement is there
* The dir can lose authoritativeness in certain circumstances (like in
{{org.apache.hadoop.fs.s3a.s3guard.LocalMetadataStore#removeAuthoritativeFromParent}})
* {{org.apache.hadoop.fs.s3a.s3guard.S3Guard#listChildrenWithTtl}} will only
mean anything if a dir listing's auth is expired *during* an operation (and not
start of a request), which is very unlikely.
* but if we add policies like "{{TTL metadata expiry is turned off}}", then the
expiry of auth dirs will still have a meaning, and we don't have to add another
config parameter.
> S3Guard: Implement time-based (TTL) expiry for entries (and tombstones)
> -----------------------------------------------------------------------
>
> Key: HADOOP-16279
> URL: https://issues.apache.org/jira/browse/HADOOP-16279
> Project: Hadoop Common
> Issue Type: Sub-task
> Components: fs/s3
> Reporter: Gabor Bota
> Assignee: Gabor Bota
> Priority: Major
>
> In HADOOP-15621 we implemented TTL for Authoritative Directory Listings and
> added {{ExpirableMetadata}}. {{DDBPathMetadata}} extends {{PathMetadata}}
> extends {{ExpirableMetadata}}, so all metadata entries in ddb can expire, but
> the implementation is not done yet.
> To complete this feature the following should be done:
> * Add new tests for metadata entry and tombstone expiry to {{ITestS3GuardTtl}}
> * Implement metadata entry and tombstone expiry
> I would like to start a debate on whether we need to use separate expiry
> times for entries and tombstones. My +1 on not using separate settings - so
> only one config name and value.
> ----
> Notes:
> * In HADOOP-13649 the metadata TTL is implemented in LocalMetadataStore,
> using an existing feature in guava's cache implementation. Expiry is set with
> {{fs.s3a.s3guard.local.ttl}}.
> * LocalMetadataStore's TTL and this TTL is different. That TTL is using the
> guava cache's internal solution for the TTL of these entries. This is an
> S3AFileSystem level solution in S3Guard, a layer above all metadata store.
> * This is not the same, and not using the [DDB's TTL
> feature|https://docs.aws.amazon.com/amazondynamodb/latest/developerguide/TTL.html].
> We need a different behavior than what ddb promises: [cleaning once a day
> with a background
> job|https://docs.aws.amazon.com/amazondynamodb/latest/developerguide/howitworks-ttl.html]
> is not usable for this feature - although it can be used as a general
> cleanup solution separately and independently from S3Guard.
> * Use the same ttl for entries and authoritative directory listing
> * All entries can be expired. Then the returned metadata from the MS will be
> null.
> * Add two new methods pruneExpiredTtl() and pruneExpiredTtl(String keyPrefix)
> to MetadataStore interface. These methods will delete all expired metadata
> from the ms.
> * Use last_updated field in ms for both file metadata and authoritative
> directory expiry.
--
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
