Re: reftable [v4]: new ref storage format

From: Shawn Pearce <spearce@spearce.org>
Date: Aug 2, 2017, 15:17 UTC
Message-ID: <CAJo=hJu1rud5pEZ93HDty1qyaCOHmwn89aEvPFe2ER0JD1ExwQ@mail.gmail.com>
In-Reply-To: <20170802092846.u4lyiogvvl7ezdfq@sigill.intra.peff.net>

On Wed, Aug 2, 2017 at 2:28 AM, Jeff King <peff@peff.net> wrote:

Show 19 quoted lines

> On Tue, Aug 01, 2017 at 07:38:37PM -0700, Shawn Pearce wrote:
>
>> > OBJS blocks can also be
>> > unbounded in size if very many references point at the same object,
>> > thought that is perhaps only a theoretical problem.
>>
>> Gah, I missed that in reftable. The block id pointer list could cause
>> a single object id to exceed what fits in a block, and that will cause
>> the writer to fail unless its caller sets the block size larger. I
>> basically assumed this overflow condition is very unlikely, as its not
>> common to have a huge number of refs pointing to the same object.
>
> It's actually quite common for us, as we have big shared-object repos
> that contain a copy of the refs of all of their child repos (for
> reachability during packing, etc). So tags, where the value is the same
> in each fork, you have one ref per fork pointing to it.
>
> Just peeking at torvalds/linux, we have some objects with ~35K refs
> pointing to them (e.g., the v2.6.11 tag).

Oy. I'll bet that every occurrence winds up in its own block due to the layout of the namespace, and so the obj block list needs 35k varint pointers. That requires a larger block size if it has any chance of fitting into the reftable format.

Another option is disable the obj table for these shared-object repos. Its an optional part of the format and can be omitted if the reader isn't likely to need to lookup by SHA-1, or is willing to pay the brute force cost of scanning every ref.