subversion-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Philip Martin <philip.mar...@wandisco.com>
Subject Re: What stands between us and branching 1.7?
Date Fri, 07 Jan 2011 09:32:56 GMT
Stefan Sperling <stsp@elego.de> writes:

> On Thu, Jan 06, 2011 at 11:20:40PM +0000, Philip Martin wrote:
>> "C. Michael Pilato" <cmpilato@collab.net> writes:
>> 
>> >   What, exactly, stands in the way of us branching for 1.7 stabilization?
>> 
>> Performance, particulary on network disks, is still a concern.  If this
>> requires using fewer, bigger transactions then we really want to do that
>> before we branch.
>> 
>> The biggest wcng feature that needs work is that revert doesn't really
>> work on tree changes.  Some of the recursive reverts go through invalid
>> wcng database states before reaching a valid final state (so an
>> interrupt would be bad), and some of the non-recursive reverts leave the
>> database in an invalid state.
>> 
>> A related issue: we are bad at detecting invalid wcng database states.
>> I suppose this could be considered a good thing as it allows the client
>> to muddle through in some cases, but if we produced hard errors then
>> perhaps we would already have fixed the code that produces invalid
>> databases.

One feature I missed:

Update should put in all the base nodes when skipping a tree conflict.
Strictly speaking we could have 1.7 behave like 1.6 but that would be
missing a major benefit of centralised metadata.  It's possible that
merge should also be putting in more nodes when skipping.

>> 
>> There are areas that would benefit from review:
>> 
>>   - Actual-only nodes, i.e. certain types of tree conflicts.  I hacked
>>     read_info to get something working, but the API was never really
>>     intended for actual-only nodes.
>>   - Granularity of transactions
>>   - Use of workqueues
>
> What do you think about the query-per-node performance problems?
> I'm surprised you don't mention them.

My first paragraph, "fewer, bigger transactions" covers that.  It needs
more than just review.

>> There are small issues that need work.  We could fix these after
>> branching but obviously it's more efficient to do it before branching:
>> 
>>   - timestamps don't self-repair when no mods are detected
>>   - cleanup doesn't fix timestamps
>>   - wc-to-wc copy breaks timestamps
>>   - revert working not-present
>>   - delete a child in a replace needs to reset some database columns
>>   - operations like mv/rm on a tree with an actual-only node
>>   - upgrade 1.6 wc that contains replaced directory (as produced by merge)
>>   - the redirect regression tests fail using serf
>>   - XFAILs
>
> Wow, nice list. Can you file issues for these points and describe them in a
> little more detail in those issues? That might give people who aren't already
> waist-deep in wc-ng something to chew on.

Will do.

-- 
Philip

Mime
View raw message