passing parameters to CREATE INDEX

Simon Riggs <simon@2ndquadrant.com> writes:

Just wanted to make clear to Hackers that the gates are now open to
include other parameters for CREATE INDEX, as originally requested here:
http://archives.postgresql.org/pgsql-hackers/2005-09/msg00851.php

Just to follow up on the discussion of that thread: what's been
implemented is a way to store arbitrary name=value strings in an index's
pg_class entry, and to make these available in a pre-parsed form through
the index relcache entry. However you'd have to be cautious about using
the values directly for any fundamental index structure decisions,
because ALTER INDEX will just change them without giving you an
opportunity to modify the index in response. So depending on what you
are doing, you might need to store the "real" values in the index
metapage, and set those values from the reloptions parameters only at
ambuild() time. This would mean that ALTER INDEX + REINDEX would be the
procedure needed to change the structure of an existing index. OTOH,
if you can tolerate on-the-fly changes of a parameter, then using it
directly from the rd_options struct would be reasonable.

Also: as of CVS tip ginoptions() accepts FILLFACTOR but nothing is done
with it. Can you do something useful with FILLFACTOR in GIN?

regards, tom lane

http://archives.postgresql.org/pgsql-hackers/2005-09/msg00851.php

Just to follow up on the discussion of that thread: what's been
implemented is a way to store arbitrary name=value strings in an index's
pg_class entry, and to make these available in a pre-parsed form through
the index relcache entry. However you'd have to be cautious about using
the values directly for any fundamental index structure decisions,
because ALTER INDEX will just change them without giving you an
opportunity to modify the index in response. So depending on what you
are doing, you might need to store the "real" values in the index
metapage, and set those values from the reloptions parameters only at
ambuild() time. This would mean that ALTER INDEX + REINDEX would be the

I see. There is one more problem: pluggable parameters for index. For example,
the parameter needed for tsearch2 (size of signature) isn't useful for others
modules/opclasses. Another issue, GiST (and GIN too) doesn't have metapage at
all for now, it's not a problem, but until now it wasn't needed. I think, we may
can add to pg_opclass's definition method/parameter name and create some API
(may be, index specific) to propagate parameter's to module's interface
functions to index.

Also: as of CVS tip ginoptions() accepts FILLFACTOR but nothing is done
with it. Can you do something useful with FILLFACTOR in GIN?

Now GIN is nested B-Tree: B-tree for entries (lexemes for tsearch2) and B-Tree
for ItemPointers per entry if entry is popular enough. So fillfactor may be used
as usual.
Small advertising :) : http://www.sigaev.ru/gin/GinStructure.pdf

--
Teodor Sigaev E-mail: teodor@sigaev.ru
WWW: http://www.sigaev.ru/

Teodor Sigaev <teodor@sigaev.ru> writes:

can add to pg_opclass's definition method/parameter name and create some API
(may be, index specific) to propagate parameter's to module's interface
functions to index.

Huh? You can get them from the index's Relation structure. I don't
think there's anything missing in the API. About all you need is an
extended struct definition for rd_options, and to provide your own code
substituting for default_reloptions(). An index AM can do both of those
locally to itself.

regards, tom lane

Teodor Sigaev <teodor@sigaev.ru> wrote:

pluggable parameters for index.
I think, we may
can add to pg_opclass's definition method/parameter name and create some API
(may be, index specific) to propagate parameter's to module's interface
functions to index.

How abount adding a new option hander to GiST/GIN support functions?

Presently, amoptions() are defined as
bytea *amoptions (ArrayType *reloptions, bool validate).
If there is a support function like this form in operator classes,
we can propagate options. Eventually, calling sequence will be
DefineIndex() -> amoptions() -> tsvectoroptions().

We also need to change gist/ginoptions() in order to determine what operator
class is used. This is a bit messy problem. Options are parsed before support
functions are initialized, so that I don't know in what form we should pass the
operator class to amoptions().

Regards,
---
ITAGAKI Takahiro
NTT Open Source Software Center