From: Johannes Sixt <j6t@kdbg.org>
Date: Sun, 16 Jul 2017 10:10:29 GMT
Subject: Re: reftable: new ref storage format
Message-ID: <4c2b03e6-cd29-6dec-e816-ce2c847355c4@kdbg.org>
In-Reply-To: <20170716100321.3jj4pdz7jqoa3dr4@sigill.intra.peff.net>

Am 16.07.2017 um 12:03 schrieb Jeff King:
> On Sun, Jul 16, 2017 at 10:07:57AM +0200, Johannes Sixt wrote:
> 
>> Am 14.07.2017 um 22:08 schrieb Jeff King:
>>> The implementation on this doesn't seem overly complex. My main concerns
>>> are what we're asking from the filesystem in terms of atomicity, and
>>> what possible races there are.
>>
>> One of the failure modes is that on Windows a file cannot be deleted while
>> it is open in any process. It can happen that a compacting updater wants to
>> remove a reftable file that is still open in a reader.
> 
> Good point. I think the explicit pointers I mentioned are an improvement
> there, because a compacting updater _can_ leave the file in place if the
> delete fails (and later, another compaction can clean up cruft that was
> left).

Yes, I think so, too. The pointers make things so much simpler.

> I assume that's more or less how pack deletion works on Windows.

Correct.

-- Hannes