Re: Volatile write caches on macOS and Windows, redux - Mailing list pgsql-hackers

From Nathan Bossart
Subject Re: Volatile write caches on macOS and Windows, redux
Date
Msg-id Zl3hD7iS9Ho9xWMI@nathan
Whole thread Raw
In response to Re: Volatile write caches on macOS and Windows, redux  (Peter Eisentraut <peter@eisentraut.org>)
Responses Re: Volatile write caches on macOS and Windows, redux
List pgsql-hackers
On Wed, May 29, 2024 at 06:49:57AM -0700, Peter Eisentraut wrote:
> On 25.05.24 04:01, Jelte Fennema-Nio wrote:
>> Is this the only reason why you're suggesting adding fsync=full,
>> instead of simply always setting F_FULLFSYNC when fsync=true on MacOS.
>> If so, I'm not sure we really gain anything by this tri-state. I think
>> people either care about data loss on power loss, or they don't. I
>> doubt many people want his third intermediate option, which afaict
>> basically means lose data on powerloss less often than fsync=false but
>> still lose data most of the time.
> 
> I agree, two states should be enough.  It could basically just be
> 
> pg_fsync(int fd)
> {
> #if macos
>     fcntl(fd, F_FULLFSYNC);
> #else
>     fsync(fd);
> #endif
> }

IIUC with this approach, anyone who is using a file system that fails
fcntl(F_FULLSYNC) with ENOSUPP would have to turn fsync off.  That might be
the right thing to do since having a third option that sends the data to
the disk cache but doesn't provide any real guarantees if you lose power
may not be worth much.  However, if such a file system _did_ provide such
guarantees with just fsync(), then it would be unfortunate to force people
to turn fsync off.  But this could very well all be hypothetical, for all I
know...  In any case, I agree that we should probably use F_FULLFSYNC by
default on macOS.

-- 
nathan



pgsql-hackers by date:

Previous
From: jian he
Date:
Subject: SQL/JSON query functions context_item doc entry and type requirement
Next
From: Pavel Borisov
Date:
Subject: Re: POC: GROUP BY optimization