Re: O_DSYNC broken on MacOS X?

On Thu, Sep 30, 2010 at 5:02 PM, Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> wrote:
> Robert Haas <robertmhaas(at)gmail(dot)com> writes:
>> Oh, I missed that.  Actually, I wasn't really so concerned with
>> whether his benchmark is correct.  I *am* concerned about being broken
>> out of the box on MacOS X.
>
> Actually, the problem with OSX is that OSX is broken out of the box,
> at least by that standard.  The system's normal configuration is that
> fsync() does nothing, so it's hardly surprising that O_DSYNC is no
> better.  You have to use wal_sync_method = fsync_writethrough to get
> actual bits-to-the-platter behavior.
>
> I'm not sure whether we should select fsync_writethrough as the default
> on OSX.  We don't make an equivalent attempt to prevent OS or storage
> malfeasance on other Unixoid platforms --- in fact, I'd say OSX is a bit
> ahead of the game in that you *can* force writethrough without resorting
> to arcane hacks with hdparm or some such.
>
> We could definitely stand to be a bit more verbose about documenting
> the platform-specific issues in this area.

I think some documentation is definitely in order, at the least. It's
certainly astonishing that the default settings aren't crash-safe.
I'd really like to understand how this shakes out on different
plaforms.

Whether we should try to work around them is a trickier question, but
I'm somewhat inclined to say yes. If we're trying to have the system
be performant in the default config, turning off synchronous_commit
would be saner than failing to make use of a system call which we know
absolutely for sure to be necessary to avoid the possibility of
database corruption.

--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise Postgres Company