On Jun 14, 2009, at 6:45 PM, Mark Hammond wrote:
On 15/06/2009 7:49 AM, Damien Katz wrote:
On Jun 10, 2009, at 7:56 PM, Mark Hammond wrote:
On 4/06/2009 3:27 AM, Damien Katz wrote:
I was once again reminded we need to add proper Windows support for
CouchDB when Mark Hammond provided a very nice patch to provide a
windows version of a couchdb unix script. (Thanks Mark!)
Thanks! Sorry I was a bit slow to see this thread.
But we also have some broken tests. Other than the issue Mark's
patch
fixes, I think it's only compaction that's still broken.
compaction and deletion are broken. I've opened a bug on the delete
problem at https://issues.apache.org/jira/browse/COUCHDB-326, as
alisdair seemed to think that couch could have a simple change
made to
allow 'delete' to work without changing how the files are opened.
> I think the
cleanest way to fix that is to change the erlang VM to open db
files
using flags to give them unix like semantics.
I've learnt something about Windows file-systems today :) My
experiments show that when using FILE_SHARE_DELETE:
* A file can be deleted while it is in use - but another file of
that
name can not be created until the handles are closed.
* A file can however be renamed while it is in use, thereby allowing
the original name to be re-opened.
If we consider the 'delete' case, then FILE_SHARE_DELETE may still
end
up with a 'race' if an attempt to delete the DB is quickly
followed by
a request to create it - the delete should succeed, but the create
will not.
If we consider the 'compact' case, I *suspect* that couch will
need to
rely on the ability to rename and replace rather than a simple
delete
- so that might work fine.
I'll have a bit of a dig around in the erlang sources and see what I
find, but of you have any other thoughts or pointers in the
meantime,
please let me know.
Cheers,
Mark
Perhaps all files that get deleted (via explicit deletion or due to
compaction) should first be renamed or moved to special trash dir
just
before being actually deleted. The file should be given a random
name to
avoid conflicting with previous deletions.
I don't see how this will fix things for delete without also
changing the open mode of the file - if the file is still opened by
one of the couchdb processes, the rename to a temp file will still
fail. If we change couch such that the couch process is terminated
before deleting there should be no need to also change the file
mode, nor to use a temp dir. If we do change the file mode, there
should be no need to use a temp dir either.
IOW, using a temp dir doesn't change the fact that either (a) all
handles must be closed before rename/delete or (b) we change the
open mode of the file. Unless of course I'm missing something.
Is it possible that for the delete case we can do as suggested in
COUCHDB-326 - ie, after we signal the couchdb processes to
terminate, actually wait for termination to complete?
I understand compaction may be trickier still when it comes to
swapping the original and compacted files over, but the same basic
invariants apply - either we investigate changing the mode of the
file, or we take steps to ensure the relevant couch processes have
terminated at that time.
My apologies if I'm missing something obvious,
Sorry, I wasn't clear, but I'm assuming we'll need to change the
erlang vm to use the proper flags.
What I meant was I'll get the Erlang side of ready, someone needs to
patch and build the Windows Erlang vm and test if it works. Then we'll
merge the Erlang portion of the code to trunk and see about getting
the patch upstream to erlang vm, and in the meantime those who want to
build on windows will need to patch Erlang themselves.
-Damien
