Berkeley DB... - Mailing list pgsql-hackers

From Mikheev, Vadim
Subject Berkeley DB...
Date
Msg-id 8F4C99C66D04D4118F580090272A7A23018BF5@SECTORBASE1
Whole thread Raw
Responses Re: Berkeley DB...  ("Michael A. Olson" <mao@sleepycat.com>)
List pgsql-hackers
Well, I've read SDB code/doc for a few hours...

1. Using RECNO db for heap.
For relational DB over-writing smgr means ability to re-use space after
DELETE/UPDATE operations (without vacuum -:)). RECNO (btree by nature,
with record number as key) will not give us this ability. To insert record
into RECNO db one has either to provide "put" method with record number
(where to store new record) or specify DB_APPEND in flags, to add new record
to the end of db (without space re-using). So, the problem (one of two base
problems of over-writing smgr for us) "where to store new tuple" (ie - where
in data file there is free space for new tuple) is not resolved.
=> we can't use SDB smgr: there are no required features - space re-using
and MVCC support.

2. SDB' btree-s support only one key, but we have multi-key btree-s...

3. How can we implement gist, rtree AND (multi-key) BTREE access methods
using btree and hash access methods provided by SDB?!

1,2,3 => we have to preserve our access methods (and ability to add new!).

Now, about WAL. What is WAL? WAL *mostly* is set of functions to 
write/read log (90% implemented) + *access method specific* redo/undo
functions... to be implemented anyway, because of conclusion above.

Comments?

Vadim


pgsql-hackers by date:

Previous
From: Tom Lane
Date:
Subject: Re: More Performance
Next
From: Bruce Momjian
Date:
Subject: Re: More Performance