qpid-users mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Chuck Rolke <cro...@redhat.com>
Subject Re: Linearstore directory structure and handling structure change
Date Wed, 12 Nov 2014 14:41:18 GMT
Hi Kim,

Let me preface my remarks with an IANASU (I am not a store user) declaration.

I think the best strategy is for the old and new data schemes to have no overlap.
  * If a user starts new code before migrating any data then 
    he gets NO data but all his data is complete and intact in the old format/layout.
  * The user may migrate the data to the new scheme and start again and get all the
    stored data.
  * Users in-the-know will migrate the data during the upgrade using a tool.

A useful feature to add to the upgrade tool is for it to emit the exact sequence of 
instructions the operator would need in order to downgrade the data to the original format/layout.

  * This would work for either the manual upgrade or for an automatic upgrade.
  * If for any reason the user must go back to the previous store version there is an authoritative
list of steps to be taken.
  * Disclaim upgrading, running for a while in the new format possibly adding things, and
then downgrading using the emitted downgrade steps. The user is on his own now unless you
think you could help him with a downgrade tool.

-Chuck



----- Original Message -----
> From: "Kim van der Riet" <kim.vdriet@redhat.com>
> To: users@qpid.apache.org
> Sent: Tuesday, November 11, 2014 1:52:27 PM
> Subject: Linearstore directory structure and handling structure change
> 
> In order to prepare to implement QPID-5671 (Add ability to use disk
> partitions and select per-queue EFPs), it has been necessary to change
> the directory structure of the store. The journal files themselves have
> not changed, only the structure. The details of the old and new
> structures are given below for informational purposes, but are probably
> secondary to this discussion.
> 
> The question at hand is how to handle a directory structure change
> across versions, and in particular, if and how users should upgrade.
> There has been some verbal discussion on this issue, but I thought it
> would be of benefit to condense these and allow all interested parties
> to comment.
> 
> If a change in directory structure is such that the new version of
> lineartsore cannot read the old version store (and visa versa), then how
> should the store/broker handle cases where a start/restart encounters an
> old store? Here are some possibilities:
> 
> 1. Add code to the store which upgrades the old layout to the new layout
> if the old layout is encountered as it recovers. This provides a
> seamless experience for those upgrading, and can be accompanied if
> necessary by loud log entries detailing the upgrade action. However, if
> a user downgrades again, then older code will misread the new structure,
> and will require the store to be truncated or downgraded by hand.
> 
> 2. Stop the broker if the old layout is encountered. It is up to the
> user to upgrade the layout (either by hand or by utility) or delete the
> old store prior to restarting the broker with a new version.
> 
> 3. Create mutually exclusive directory structures for versions. This
> means that a newer version will never encounter the contents of an older
> version, and visa versa. If a user wants to upgrade a version, they will
> need to move the old store files into the new structure by hand or
> utility. This means that the broker will always start after an upgrade,
> but the user may "lose" existing persistent queues and messages unless
> they perform an upgrade.
> 
> Of course, some combinations of the above are also possible.
> 
> It should be also be born in mind that for the current changes to the
> store:
> 
> 1. Linearstore is not yet the "official" linux store, and being
> "experimental" is not built by default. Legacystore is still the current
> default store. A change to the directory structure at this stage should
> not be serious. These questions do affect future possible changes,
> though.
> 
> 2. Changes to the format of the individual journal files are handled
> separately. Journal files contain a version number which is checked
> prior to decoding and if it is incorrect will cause the store to stop.
> 
> 3. There is currently no direct means to detect a layout change other
> than by errors encountered as the store attempts to read the directory.
> This could be addressed by some kind of file or directory layout/naming
> convention which makes the version clear. This could be done in a way
> that is somewhat similar to the idea #3 above.
> 
> I would welcome ideas/feedback on this question.
> 
> 
> File layout for linearstore:
> ============================
> The following is for informational purposes, and illustrates the above
> issue.
> 
> Background:
> 
> In order to allow empty file pools (EFPs) to be established on different
> media, and for queues to be able to choose which partition to use
> depending on their performance requirements, the directory structure of
> linearstore must be changed. For example, queues with high throughput
> and low latency requirements may be established on expensive solid state
> media, while low throughput non-critical queues can be directed to use
> regular rotating magnetic media.
> 
> The current journal layout moves journal files into a directory within
> the "qls/jrnl/<queue_name>" directory. However, this limits the files to
> the partition on which the qls directory exists.
> 
> OLD LAYOUT
> 
> qls
>   +-dat <BDB database files>
>   +-jrnl
>   |   +-queue_0
>   |   |   +-file_1.jrnl
>   |   |   +-file_2.jrnl
>   |   |   ...
>   |   |   +-file_n.jrnl
>   |   +-queue_1
>   |   ...
>   |   +-queue_n
>   +-tpl <contains journal files for transaction boundaries>
>   +-p001 <EFP partition 1>
>       +-efp
>           +-2048k <Default pool of empty 2048k journal files>
> 
> Partition p001 is the default partition and is located on the same
> partition as the store-directory (or its default). Additional partitions
> (p002, p003, etc.) may be created and used as mount points for other
> physical disk partitions. Each of these partitions would contain empty
> file pools (EFPs) containing empty files of various sizes. Currently
> only one size (2048k) is in use, but other sizes may be used in the
> future.
> 
> To solve the single partition limitation, the journal files are now
> moved from the EFP directory into a "in_use" subdirectory on the same
> partition and a symlink is established in the qls/jrnl/<queue_name>
> directory to the in-use file.
> 
> An addition "tidy-up" change is to remove the "efp" directory under each
> partition, as it is superfluous and serves no purpose. Each EFP size
> directory is now directly under the partition directory:
> 
> NEW LAYOUT
> 
> qls
>   +-dat
>   +-jrnl
>   |   +-queue_0
>   |   |   +-symlink to file_1.jrnl
>   |   |   +-symlink to file_2.jrnl
>   |   |   ...
>   |   |   +-symlink to file_n.jrnl
>   |   +-queue_1
>   |   ...
>   |   +-queue_n
>   +-tpl <contains symlinks to files in in_use dir in the partitions>
>   +-p001
>   |   +-2048k
>   |   |   +-in_use <contains all files in use by all queues using this
>   |   |   partition>
>   |   |       +-file_1.jrnl
>   |   |       +-file_2.jrnl
>   |   |       ...
>   |   |       +-file_n.jrnl
>   |   |   +-returned <contains files returned from use, but not yet cleaned
>   |   |   up for re-use>
>   |   +-32768k
>   |   |   +-in_use
>   |   |   +-returned
>   |   ...
>   |   +-size_n <other possible EFP sizes>
>   |       +-in_use
>   |       +-returned
>   +-p002
>   |   +-<layout as for p001>
>   ...
>   +-pnnn
>       +-<layout as for p001>
> 
> The "returned" directory is intended to be a resting place for used
> files which have not yet been cleaned up or overwritten for re-use. The
> idea is that because cleaning up and overwriting are relatively
> time-consuming actions, an external process or internal worker thread
> can perform this function on a lower priority rather than the threads
> which are handling message persistence. The action of moving a file from
> one directory to another is relatively cheap. Currently the
> cleanup/overwriting is being done by the store itself on the persistence
> thread, but this opens the way to changing the way used files are
> handled at a later time.
>   
> 
> 
> ---------------------------------------------------------------------
> To unsubscribe, e-mail: users-unsubscribe@qpid.apache.org
> For additional commands, e-mail: users-help@qpid.apache.org
> 
> 

---------------------------------------------------------------------
To unsubscribe, e-mail: users-unsubscribe@qpid.apache.org
For additional commands, e-mail: users-help@qpid.apache.org


Mime
View raw message