Re: Creation of an empty table is not fsync'd at checkpoint

Hi,

On 2022-01-28 00:39:22 +0200, Heikki Linnakangas wrote:
> On 28/01/2022 00:11, Thomas Munro wrote:
> > ... but we still never synchronize "base/5".  According to our
> > project's reading of the POSIX tea leaves we should be doing that to
> > nail down the directory entry.
> 
> Really? 'base/5' is fsync'd by initdb, when it's created. I didn't think we
> try to fsync() the directory, when a new file is created in it.

I've not heard of concrete reports of it being needed (whereas the directory
fsync being needed after a rename() is pretty easy to be reproduce). There's
some technical reasons why it'd make sense for it to only be really needed for
things after the initial file creation, but I'm not sure it's a good idea to
rely on it.

From the filesystem POV a file doesn't necessarily know which directory it is
in, and it can be several at once due to hardlinks. Scheduling all the
directories entries to be durable only really is feasible if all metadata
operations are globally ordered - which we don't really want for scalability
reasons.

But when initially creating a file, it's always in the context of a
directory. I assume that most journalled filesystem is going to have the
creation of the directory entry and of the file itself be related journalling
operations. But I wouldn't bet it's all, and theoretically we claim to be
usable on non-journalled filesystems as well...

Greetings,

Andres Freund