Yep. Db#db.update_seq is not always the same as Db#db.committed_update_seq when delayed_commits
are on.
Adam
On Apr 12, 2010, at 8:54 AM, Paul Davis wrote:
> An idle curiosity, is it ever possible to replicate something that has
> been written to disk before a header is flushed?
>
>
> On Mon, Apr 12, 2010 at 8:46 AM, Adam Kocoloski <kocolosk@apache.org> wrote:
>> Yep, your analysis is dead-on, and is a more complete solution than what I propose.
Best,
>>
>> Adam
>>
>> On Apr 12, 2010, at 4:51 AM, Robert Newson wrote:
>>
>>> Would it be safer to have a low- and high- watermark for the
>>> update_seq in memory? What I mean is that the db writer will never
>>> write out an update_seq that is N higher than the last committed one;
>>> if it is forced to do so, to permit a write, it then fsync's and
>>> resets high_seq to last_committed_seq. This way you can genuinely
>>> ensure that you don't reuse an update_seq. In practice we could allow
>>> a large delta, one that is larger than the number of fsyncs we expect
>>> to manage in the commit interval.
>>>
>>> Your idea to just bump the update_seq "significantly" mostly pans out
>>> (I know a system that does precisely this) but it would be a data loss
>>> scenario if when it doesn't pan out.
>>>
>>> B.
>>>
>>> On Mon, Apr 12, 2010 at 3:54 AM, Adam Kocoloski
>>> <adam.kocoloski@gmail.com> wrote:
>>>> Currently a DB update_seq can be reused if there's a power failure before
the header is sync'ed to disk. This adds some extra complexity and overhead to the replicator,
which must confirm before saving a checkpoint that the source update_seq it is recording will
not be reused later. It does this by issuing an ensure_full_commit call to the source DB,
which may be a pretty expensive operation if the source has a constant write load.
>>>>
>>>> Should we try to fix that? One way to do so would be start at a significantly
higher update_seq than the committed one whenever the DB is opened after an "unclean" shutdown;
that is, one where the DB header is not the last term stored in the file. Although, I suppose
that's not an ironclad test for data loss -- it might be the case that none of the lost updates
were written to the file. I suppose we could "bump" the update_seq on every startup.
>>>>
>>>> Adam
>>>>
>>>>
>>
>>
|