Subject | Re[3]: [Firebird-Architect] blobs causes table fragmentation. TEST+LONG |
---|---|
Author | Dmitry Kuzmenko |
Post date | 2004-10-05T07:52:59Z |
Hello, Dmitry!
Tuesday, October 5, 2004, 11:28:11 AM, you wrote:
DK> Ok, I did that test. Full test description follows (LONG).
Resume:
1. blobs that fit at data page affects performance when
they are not read.
2. database restoring with bigger/smaller page size
can downgrade/upgrade performance, because new
page size will allow (or not) to fit some blobs on data page
This is bad because we got ununderstandable behavior for
users who think about performance gain increasing page size.
3. optimizer behavior is predictable for that test.
if it calculates cardinality by dividing average data pages
by record size in rdb$format, it will have bad cardinality
for tables b and c (and wrong record count if it is ever needed).
This is Ok, but table D shows advantage since it's cardinality
will be close to table A.
4. gstat ouput can confuse user. average record size
is reported independently of blobs stored on data page.
Maybe it is OK, but need to be commented somewhere.
--
Dmitri Kouzmenko, www.ibase.ru, (095) 953-13-34
Tuesday, October 5, 2004, 11:28:11 AM, you wrote:
DK> Ok, I did that test. Full test description follows (LONG).
Resume:
1. blobs that fit at data page affects performance when
they are not read.
2. database restoring with bigger/smaller page size
can downgrade/upgrade performance, because new
page size will allow (or not) to fit some blobs on data page
This is bad because we got ununderstandable behavior for
users who think about performance gain increasing page size.
3. optimizer behavior is predictable for that test.
if it calculates cardinality by dividing average data pages
by record size in rdb$format, it will have bad cardinality
for tables b and c (and wrong record count if it is ever needed).
This is Ok, but table D shows advantage since it's cardinality
will be close to table A.
4. gstat ouput can confuse user. average record size
is reported independently of blobs stored on data page.
Maybe it is OK, but need to be commented somewhere.
--
Dmitri Kouzmenko, www.ibase.ru, (095) 953-13-34