jclouds-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Andrew Gaul <g...@apache.org>
Subject Re: Question about JCLOUDS-658 Jira issue
Date Wed, 12 Nov 2014 08:34:21 GMT
On Mon, Nov 10, 2014 at 05:45:10PM -0800, Gabriel Lavoie wrote:
>     I'm currently reviewing the jclouds filesystem blobstore API. I
> noticed that user metadata doesn't get saved with the 1.8.1 version,
> but that a comment exists in the code regarding Java 7 and NIO
> filestore attributes. I also found out with the JCLOUDS-658 that
> jclouds 2.0 (unreleased) fixes the issue by using the NIO filestore
> attributes.
> 
> Regarding this issue and resolution:
> - Is the filesystem API considered "production" safe or only suggested for
> testing/debugging?
> - I'm not sure about the solution of using the filesystem metadata store
> to store metadata. Many (most?) filesystem archival/backup/explorer are
> not aware of this metadata and it may get lost without the user knowing
> about it. I would consider using this for testing, but never in a
> production environment.
>      - Could an alternate way of storing metadata be implemented, for
> example in a properties file named <id>.metadata stored alongside
> with the object file?
> - I could implement a file metadata storage by wrapping
> FilesystemStorageStrategyImpl, but I don't think this is a good idea to
> wrap it as I have to re-implement a few classes to have the dependency
> injection work correctly.

The filesystem blobstore has traditionally been used for testing
purposes and has a few caveats for production use.  However, it should
perform well enough for applications with a small number of objects
(tens of thousands) and small object sizes (gigabytes).

My primary concern for production use is that a filesystem is not a
blobstore; a single node lacks the performance and reliability that a
multi-node Swift installation provides.  While Swift uses a filesystem
to stores its blob data, it uses XFS with several tuning parameters and
hashes object names into directory shards to avoid scalability limits.

The filesystem blobstore has a few known issues; we will gladly accept
patches for these:

  * scalability: returning a paginated list of objects creates the
    entire list in-memory
  * compatibility: lacks multi-part upload, some listing options not
    supported
  * error paths: writing partial objects or concurrently overwriting
    objects has different semantics than real object stores, especially
    on Windows
  * performance: excessive system calls on various code paths
  * metadata: 1.8.1 does not support user metadata and other attributes.
    The approach for future releases only works on Linux and Windows.
    Current JDK on Mac OS X does not support xattr (JDK-8030048).

Using file extended attributes is the "right" way to store object
metadata and allows other tools to interact with these objects instead
of using a special format that only jclouds understands.  However, I
appreciate the practical reasons to prefer another approach; could you
explore this issue further?  Every solution has a trade-offs, for
example storing in <id>.metadata file requires extra work when listing
objects.  You may want to consider using a second directory for metadata
instead.  Finally you might want to look at Jimfs which creates an
in-memory filesystem provider; perhaps you could create an on-disk
provider which stores extended attributes in some other way?

-- 
Andrew Gaul
http://gaul.org/

Mime
View raw message