From: Jeff King <peff@peff.net>
Date: Thu, 20 Apr 2017 20:14:43 GMT
Subject: Re: [PATCH] repack: respect gc.pid lock
Message-ID: <20170420201443.ee4tgoymzpfvl4jq@sigill.intra.peff.net>
In-Reply-To: <7e31f4ed5c0f4c31b2870fb58cf7110e@exmbdft7.ad.twosigma.com>

On Thu, Apr 20, 2017 at 08:10:24PM +0000, David Turner wrote:

> > Is "-a" or "-A" the key factor? Are there current callers who prefer the current
> > behavior of "possibly duplicate some work, but never report failure" versus "do
> > not duplicate work, but sometimes fail due to lock contention"?
> 
> One problem with failing is that it can leave a temp pack behind.

Yeah. IMHO we should probably treat failed object and pack writes as
normal tempfiles and remove them (but possibly respect a "debug mode"
that leaves them around). But that's another patch entirely.

> I think the correct fix is to change the default code.packedGitLimit on 64-bit 
> machines to 32 terabytes (2**45 bytes).  That's because on modern Intel 
> processors, there are 48 bits of address space actually available, but the kernel 
> is going to probably reserve a few bits.  My machine claims to have 2**46 bytes 
> of virtual address space available.  It's also several times bigger than any 
> repo that I know of or can easily imagine.
> 
> Does that seem reasonable to you?

Yes, it does.

-Peff