Re: GIN improvements part 1: additional information

On 12/09/2013 11:34 AM, Alexander Korotkov wrote:

On Mon, Dec 9, 2013 at 1:18 PM, Heikki Linnakangas
<hlinnakangas@vmware.com>wrote:

Even if we use varbyte encoding, I wonder if it would be better to treat
block + offset number as a single 48-bit integer, rather than encode them
separately. That would allow the delta of two items on the same page to be
stored as a single byte, rather than two bytes. Naturally it would be a
loss on other values, but would be nice to see some kind of an analysis on
that. I suspect it might make the code simpler, too.

Yeah, I had that idea, but I thought it's not a better option. Will try to
do some analysis.

The more I think about that, the more convinced I am that it's a good idea. I don't think it will ever compress worse than the current approach of treating block and offset numbers separately, and, although I haven't actually tested it, I doubt it's any slower. About the same amount of arithmetic is required in both versions.

Attached is a version that does that. Plus some other minor cleanup.

(we should still investigate using a completely different algorithm, though)