Re: reftable: new ref storage format

From: Shawn Pearce <spearce@spearce.org>
Date: Jul 14, 2017, 00:27 UTC
Message-ID: <CAJo=hJv7kaT3m6k1nz1-tGuVAMmgnrS0dcfycGfE3PyXjG3xRA@mail.gmail.com>
In-Reply-To: <20170713203533.vcfyf5iei46g4tcf@sigill.intra.peff.net>

On Thu, Jul 13, 2017 at 1:35 PM, Jeff King <peff@peff.net> wrote:

Show 6 quoted lines

> On Thu, Jul 13, 2017 at 12:56:54PM -0700, Stefan Beller wrote:
>
> I agree that a full binary search of a reftable is harder because of the
> prefix compression (it may still be possible by scanning backwards, but
> I think there are ambiguities when you land in the middle of a record,
> since there's no unambiguous end-of-record character).

Its impossible to safely binary search this reftable format using a naive divide byte count in half and find record boundary approach. I actually did design an earlier version of reftable that was safe to use this approach for its binary search within blocks, and wound up discarding it. It was slower and more complex implementation than the format I shared with the list.

> But I don't think
> it matters. If you binary-search to a constant-sized block, then a
> linear scan of the block is acceptable.

Depends on the block size. :)

> Not that I'm recommending just gzipping the whole packed-refs file. It
> ruins the fast-lookup.

As I just mentioned elsewhere in the thread:

  src file    65306185
  gzip        28338906
  reftable  28782292

The reftable format (for 64k block, 256 restart) is within spitting distance (432 KiB) of a default level gzip of packed-refs. We can get fast-lookup, and OK compression.

Show 5 quoted lines

> We _could_ consider gzipping individual blocks of
> a reftable (or any structure that allows you to search to a
> constant-sized block and do a linear search from there). But given that
> they're in the same ballpark, I'm happy with whatever ends up the
> simplest to code and debug. ;)

This does help to shrink the file, e.g. it drops from 28M to 23M.

It makes it more CPU costly to access a block, as we have to inflate that to walk through the records. It also messes with alignment. When you touch a block, that may be straddling two virtual memory pages in your kernel/filesystem.

I'm not sure those penalties are worth the additional 16% reduction in size.

Show 11 quoted lines

>> When Shawn presented the proposal, a couple of colleagues here
>> were as excited as I was, but the daring question is, why Shawn
>> did not give the whole thing in BNF format from top down:
>>
>>   initial-block
>>   content-blocks*
>>   (index-block)
>>   footer
>
> Yeah, I agree it took me a bit to figure out what was going on. A
> high-level overview of the format would have been nice.

Noted, I've added this to my writeup.