On Mon, 2012-05-14 at 19:20 +0200, Joel Reardon wrote:
> The long long is divided as follows:
> 32 bits for the (KSA-relative) LEB number, 32 bits for the offset in the
> leb where the key is found. So its the same as the lnum/offs for the
> current one. Theres substancial compression though, that is available,
> since theres likely not more than 2^^32 LEBS for the KSA and the number of
> bits needed for key offset is LEB_SHIFT - 4.
> 
> Is 32 bits sufficient to address all keys:
> one key per datanode means 4096 * 2^32 = 2^44, so only 16 TB available
> for 32 bit key addresses.
> 
> Though there is similar waste for lnum/offs as well. Perhaps zbranches can
> be stored as a u8[] and demarshalled with bit-op macros when needed for
> computations.

OK, thanks for explanation. Why not to then store 2x32-bit fields
instead, which is consistent with the current style? Why "long long"?

-- 
Best Regards,
Artem Bityutskiy