Re: svn commit: r934008 - /subversion/trunk/subversion/libsvn_subr/sqlite.c

[Philip Martin <ph...@wandisco.com>]

philip@apache.org writes:

> Author: philip
> Date: Wed Apr 14 16:35:11 2010
> New Revision: 934008
>
> URL: http://svn.apache.org/viewvc?rev=934008&view=rev
> Log:
> * subversion/libsvn_subr/sqlite.c
>   (svn_sqlite__hotcopy): Use the SQLite backup interface if available.
>
> Modified:
>     subversion/trunk/subversion/libsvn_subr/sqlite.c

The backup method works but is measurably slower in my tests than than
the plain copy under a lock; the benefit should be better concurrency.

I don't know which is better: the faster hotcopy that blocks writers
or the slower hotcopy that allows writers to progress?  At the moment
it's a compile time decision based on the version of SQLite but that's
not a good way to decide whether we should be using this feature.  I
think it's too obscure to make it configurable on a per-repository
basis so I don't really know what we should do.

-- 
Philip

Re: svn commit: r934008 - /subversion/trunk/subversion/libsvn_subr/sqlite.c

["C. Michael Pilato" <cm...@collab.net>]

Philip Martin wrote:
> philip@apache.org writes:
> 
>> Author: philip
>> Date: Wed Apr 14 16:35:11 2010
>> New Revision: 934008
>>
>> URL: http://svn.apache.org/viewvc?rev=934008&view=rev
>> Log:
>> * subversion/libsvn_subr/sqlite.c
>>   (svn_sqlite__hotcopy): Use the SQLite backup interface if available.
>>
>> Modified:
>>     subversion/trunk/subversion/libsvn_subr/sqlite.c
> 
> The backup method works but is measurably slower in my tests than than
> the plain copy under a lock; the benefit should be better concurrency.
> 
> I don't know which is better: the faster hotcopy that blocks writers
> or the slower hotcopy that allows writers to progress?  At the moment
> it's a compile time decision based on the version of SQLite but that's
> not a good way to decide whether we should be using this feature.  I
> think it's too obscure to make it configurable on a per-repository
> basis so I don't really know what we should do.

The nature of a "hot copy" is generally such that the copied thing isn't
"unavailable" by most definitions thereof.  Not sure that write-locking the
repository really counts as "hot".  :-)

-- 
C. Michael Pilato <cm...@collab.net>
CollabNet   <>   www.collab.net   <>   Distributed Development On Demand

Re: svn commit: r934008 - /subversion/trunk/subversion/libsvn_subr/sqlite.c

[Greg Stein <gs...@gmail.com>]

On Wed, Apr 14, 2010 at 12:43, Philip Martin <ph...@wandisco.com> wrote:
> philip@apache.org writes:
>
>> Author: philip
>> Date: Wed Apr 14 16:35:11 2010
>> New Revision: 934008
>>
>> URL: http://svn.apache.org/viewvc?rev=934008&view=rev
>> Log:
>> * subversion/libsvn_subr/sqlite.c
>>   (svn_sqlite__hotcopy): Use the SQLite backup interface if available.
>>
>> Modified:
>>     subversion/trunk/subversion/libsvn_subr/sqlite.c
>
> The backup method works but is measurably slower in my tests than than
> the plain copy under a lock; the benefit should be better concurrency.
>
> I don't know which is better: the faster hotcopy that blocks writers
> or the slower hotcopy that allows writers to progress?  At the moment
> it's a compile time decision based on the version of SQLite but that's
> not a good way to decide whether we should be using this feature.  I
> think it's too obscure to make it configurable on a per-repository
> basis so I don't really know what we should do.

To me, "hot copy" means that you are trying to continue operation.
Thus, the concurrency is needed.

I don't see that a hot copy needs to be blazing fast. Just concurrent.

Cheers,
-g