Guys, why we're still on alpha stage, maybe we could fix multisegments too?
Many (i.e 7 or 8) years ago we used following schema for multisegment keys
without padding
Suppose we have keys A and B and want to make a multisegment key out of it.
The resulting key would be
f(A) 0x00 f(B)
0x00 serves as separator and f() is a transformation that converts
0x00=>0x01 0x00
0x01=>0x01 0x01
Any other byte remains unchanged.0x00s at the end can be compressed, so we
get efficient key is there are only/many NULLs.
I do not think the schema is much more complicated than RUN length and
padding.
Vlad
And while we're on it why not to fix integer representation;) Doubles are
strange creation by mathematician, exact longlongs would be really nice,
not?
> -----Original Message-----
> From: Kevin.Lewis@stripped [mailto:Kevin.Lewis@stripped]
> Sent: Tuesday, February 17, 2009 11:55 PM
> To: Lars-Erik Bjørk
> Cc: Jim Starkey; Vladislav Vaintroub; 'FalconDev'
> Subject: Re: Patch for bug#42208
>
> According to the blog link forwearded by Mark, Oracle customers don not
> like that zero length strings (which ar equal to each other) are
> automatically converted to NULLs. Both suggestions take care of that
> in
> Falcon. So this is the most inportant thing; make a zero length string
> equal to 0x00 length 1.
>
> The question is whether to keep adding 0x00 to other lengths of binary
> zero strings. Jim says it does not matter to anyone but QA that 0x00
> and 0x0000 sort separately. And Vlad points out that even if we did
> this for single field keys, it would not sort them differently for
> multisegment key since we always pad them to a RUN length. I think
> that
> if it does not cause any extra difficulties or comlexity in the code,
> why not keep QA happy for single segment keys.
>
> And I still am unclear why this little change in index order should
> cause us to change the ODS format while still in the alpha stage. What
> is the downside of a new engine that starts converting zero length
> strings into 0x00? New entries will be added to the index after the
> NULLS. Older zero length strings would be mixed up with the NULLS and
> may not be found for direct searched until the index is rebuilt. We
> can
> document that as a bug fix in the index, which it is. Nobodies critical
> data is depending on us finding all zero length strings.
>
> Kevin
>
> Lars-Erik Bjørk wrote:
> > Ok, so we probably don't want to do the caching after all then? Does
> > anyone else have an opinion on how to proceed on this? Do we agree on
> > any best approach?
> >
> > /Lars-Erik
> >
> > Jim Starkey wrote:
> >> Vladislav Vaintroub wrote:
> >>> Hi Lars-Erik,
> >>> I wonder if adding 0x00 to the (binary) string values that already
> start
> >>> with 0x00 would not be less works that modifying index walker etc.
> This
> >>> looks like huge amount of work you have done (good) but I wonder if
> >>> there is
> >>> a good reason for it. Assuming (binary) strings that start with
> 0x00 are
> >>> really seldom, prepending 0x00 to a key after a check is not going
> to
> >>> be an
> >>> expensive operation. And that makes NULL *really* different from
> >>> other index
> >>> values. And that allows maybe in some distant future index-only
> >>> access, so
> >>> you can answer "is null/is not null" without extra accessing the
> >>> record and
> >>> this is a real performance advantage.
> >>>
> >>>
> >> Why do you want to do that? Is the following sufficient:
> >>
> >> 1. A null is represented as either a zero length key or a missing
> >> segment in a multi-segment key. This collates lowest.
> >> 2. A zero length binary key is represented by a single byte of
> zero.
> >> 3. A binary key with a single zero byte is indistinquishable from
> a
> >> zero length (but non-null) key
> >> 4. A binary key with a leading zero byte and a subsequent non-zero
> >> byte will collate about #2 and #3.
> >>
> >> I don't think we really care about the ordering of a non-null, zero
> >> length key and and all zero binary key. I don't think anyone else
> >> should, either.
> >>
> >
> >
