[
https://issues.apache.org/jira/browse/JENA-327?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13459725#comment-13459725
]
Simon Helsen commented on JENA-327:
-----------------------------------
first of all, we currently do a backup by copying in the file system. Our
initial impression was that this would be faster than any other approach
especially given that the restore has to be fast as well. I have no numbers
right now, but I think for now, we want to stick to this unless we have
evidence the n-quad dump is fast enough on extremely large stores.
So if I understand you well, you are saying that a WRITE transaction will
guarantee that nothing changes on disk even if async writeback is introduced?
That sounds good and would serve our purpose. Why would using the read
mechanism be safer long-term?
The 3rd option is not clear to me. Are you saying there is API (not internal
interfaces) we can use to a) hold up everything and b) manually flush the
journal?
> TDB Tx transaction lock to permit backups
> -----------------------------------------
>
> Key: JENA-327
> URL: https://issues.apache.org/jira/browse/JENA-327
> Project: Apache Jena
> Issue Type: Improvement
> Components: TDB
> Affects Versions: TDB 0.9.4
> Reporter: Simon Helsen
>
> With large repositories, it is important to be able to create backups once in
> a while. This is because recreating an rdf store with millions of triples can
> be forbiddingly expensive. Moreover, it should be possible to take those
> backups while still allowing read activity on the store as in many cases, a
> complete shutdown is usually not possible. Before the introduction of tx, it
> was relatively straightforward to provide the right locks on the client-side
> to safely suspend any disk activity for a period of time enough to make a
> backup of the index.
> However, since tx, things have become slightly more complicated because TDB
> Tx touches the disk at other times than when performing write/sync
> activities. Right now, because of some understanding of how TDB Tx is
> implemented, it is still possible for clients to avoid disk activities to
> implement a backup process, but this dependency on TDB Tx implementation
> details is not very good. Moreover, we anticipate that in the future, the
> merging process from the journal into the main index may become entirely
> asynchornous for performance reasons. The moment that happens, client have no
> control anymore as to when the disk is being touched.
> For this reason, we are requesting the following feature: a "backup" lock (by
> lack of a better name). Its semantics is that when the lock is taken, TDB Tx
> guarantees that no disk activity takes place and if necessary pauses
> activities. In other words, no write transaction should be able to complete
> and read transactions will not attempt to merge the journal. The idea would
> be that regular read activities can still continue. The API could be as
> simple as something like this:
> try {
> dataset.begin(ReadWrite.BACKUP) ;
> <do whatever is necessary to backup the index>
> } finally {
> dataset.end()
> }
> As for the implementation, we suspect you currently have locks in place which
> could be used to guarantee this behavior. E.g. could
> txn.getBaseDataset().getLock().enterCriticalSection(Lock.WRITE) be sufficient?
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira
