why this 2-column index not used

I have a composite index on two columns in a table. However, the
index is not used in a query that restricts the 2nd column to a
constant. If both columns are linked with columns in other join
tables, the index will be used.

To illustrate it with an example, I have a query like this:

select s.ticker, p.quantity
from stock s, positions p
where s.type_id = 4
and s.stock_id = p.stock_id

There is an index on stock: (stock_id, type_id). In the table stocks,
there are about 100 different type_ids evenly distributed, and ten of
thousand different stock_ids (each stock_id could map to every one of
the 100 type_ids).

From the plan, the above query does not use the index, but the
following query does use the index:

select s.ticker, p.quantity
from stock s, positions p
where s.type_id = p.pos_type_id
and s.stock_id = p.stock_id

The only difference is that in the first query, type_id is constant 4,
while in the second it is linked with another column in the second
table.

The stats are good. Is there anything else that might have caused the
above?

Even if I add another index stock (stock_id), the first query still
doesn't use any index.

However, if the query is like this:

select s.ticker, p.quantity
from stock s, positions p
where s.stock_id = p.stock_id

The index on stock (stock) will be used (of course). If the query is
like this:

select s.ticker, p.quantity
from stock s, positions p
where s.stock_id = 101
s.type_id = 4

then the index on stock (stock_id, type_id) is used.

BTW, there is a foreign key in stock that links the stock_id field to
the positions (stock_id).

Anybody can explain the above?

"Henry J." wrote in message
news:02e9b6d6-229f-459a-b8f7-0097b19ba2b1-at-m45g2000hsb.googlegroups.com...
>I have a composite index on two columns in a table. However, the
> index is not used in a query that restricts the 2nd column to a
> constant. If both columns are linked with columns in other join
> tables, the index will be used.
>
> To illustrate it with an example, I have a query like this:
>
> select s.ticker, p.quantity
> from stock s, positions p
> where s.type_id = 4
> and s.stock_id = p.stock_id
>
> There is an index on stock: (stock_id, type_id). In the table stocks,
> there are about 100 different type_ids evenly distributed, and ten of
> thousand different stock_ids (each stock_id could map to every one of
> the 100 type_ids).
>
> From the plan, the above query does not use the index, but the
> following query does use the index:
>
> select s.ticker, p.quantity
> from stock s, positions p
> where s.type_id = p.pos_type_id
> and s.stock_id = p.stock_id
>
> The only difference is that in the first query, type_id is constant 4,
> while in the second it is linked with another column in the second
> table.
>
> The stats are good. Is there anything else that might have caused the
> above?

You are confusing the use of an index with automatically getting better
performance than not using an index. Sometimes using an index will result in
better performance, and sometimes not. If DB2 determines that it is faster
to not use an index, then it will not use one. This typically happens when
DB2 will have read every data page (rows are stored in pages) to satisfy the
query.

Keep in mind that you are returning all rows from the table (you don't
qualify which type_id or which stock_id's you want), so it is safe to assume
that DB2 will have to read not just every page, but in fact every row, to
satisfy the query.

Assuming you did a runstats as follows:
runstats on table with distribution on key columns and indexes
all
then DB2 almost always makes the correct decision.

On Sep 13, 12:34*am, "Mark A" wrote:
> "Henry J." wrote in message
>
> news:02e9b6d6-229f-459a-b8f7-0097b19ba2b1-at-m45g2000hsb.googlegroups.com...
>
>
>
> >I have a composite index on two columns in a table. *However, the
> > index is not used in a query that restricts the 2nd column to a
> > constant. *If both columns are linked with columns in other join
> > tables, the index will be used.
>
> > To illustrate it with an example, I have a query like this:
>
> > select s.ticker, p.quantity
> > from stock s, positions p
> > * * *where s.type_id = 4
> > * * * * * * * and s.stock_id = p.stock_id
>
> > There is an index on stock: (stock_id, type_id). *In the table stocks,
> > there are about 100 different type_ids evenly distributed, and ten of
> > thousand different stock_ids (each stock_id could map to every one of
> > the 100 type_ids).
>
> > From the plan, the above query does not use the index, but the
> > following query does use the index:
>
> > select s.ticker, p.quantity
> > from stock s, positions p
> > * * *where s.type_id = p.pos_type_id
> > * * * * * * * and s.stock_id = p.stock_id
>
> > The only difference is that in the first query, type_id is constant 4,
> > while in the second it is linked with another column in the second
> > table.
>
> > The stats are good. *Is there anything else that might have caused the
> > above?
>
> You are confusing the use of an index with automatically getting better
> performance than not using an index. Sometimes using an index will resultin
> better performance, and sometimes not. If DB2 determines that it is faster
> to not use an index, then it will not use one. This typically happens when
> DB2 will have read every data page (rows are stored in pages) to satisfy the
> query.
>
> Keep in mind that you are returning all rows from the table (you don't
> qualify which type_id or which stock_id's you want), so it is safe to assume
> that DB2 will have to read not just every page, but in fact every row, to
> satisfy the query.
>
> Assuming you did a runstats as follows:
> runstats on table with distribution on key columns and indexes
> all
> then DB2 almost always makes the correct decision.

My first query is not to return all data? It's only returning on
s.type_id = 4.

Actually my questions arised when I changed the indexes on the table
stock and the same queries run more than twice as long as before.
Then I found the main index is not used at all. Since the tables are
huge (40 mil rows), I don't see why DB2 decides it's faster not to use
the index.

Maybe I should create an index on stock (type_id)? Actually it
eliminates a table scan on stock. But why it is not using (stock_id,
type_id) even though both columns are used in where clause.

I know I must have some mis-conceptions. I just don't know which ones.

On Sep 13, 10:07*am, "Henry J." wrote:
> On Sep 13, 12:34*am, "Mark A" wrote:
>
>
>
> > "Henry J." wrote in message
>
> >news:02e9b6d6-229f-459a-b8f7-0097b19ba2b1-at-m45g2000hsb.googlegroups.com....
>
> > >I have a composite index on two columns in a table. *However, the
> > > index is not used in a query that restricts the 2nd column to a
> > > constant. *If both columns are linked with columns in other join
> > > tables, the index will be used.
>
> > > To illustrate it with an example, I have a query like this:
>
> > > select s.ticker, p.quantity
> > > from stock s, positions p
> > > * * *where s.type_id = 4
> > > * * * * * * * and s.stock_id = p.stock_id
>
> > > There is an index on stock: (stock_id, type_id). *In the table stocks,
> > > there are about 100 different type_ids evenly distributed, and ten of
> > > thousand different stock_ids (each stock_id could map to every one of
> > > the 100 type_ids).
>
> > > From the plan, the above query does not use the index, but the
> > > following query does use the index:
>
> > > select s.ticker, p.quantity
> > > from stock s, positions p
> > > * * *where s.type_id = p.pos_type_id
> > > * * * * * * * and s.stock_id = p.stock_id
>
> > > The only difference is that in the first query, type_id is constant 4,
> > > while in the second it is linked with another column in the second
> > > table.
>
> > > The stats are good. *Is there anything else that might have caused the
> > > above?
>
> > You are confusing the use of an index with automatically getting better
> > performance than not using an index. Sometimes using an index will result in
> > better performance, and sometimes not. If DB2 determines that it is faster
> > to not use an index, then it will not use one. This typically happens when
> > DB2 will have read every data page (rows are stored in pages) to satisfy the
> > query.
>
> > Keep in mind that you are returning all rows from the table (you don't
> > qualify which type_id or which stock_id's you want), so it is safe to assume
> > that DB2 will have to read not just every page, but in fact every row, to
> > satisfy the query.
>
> > Assuming you did a runstats as follows:
> > runstats on table with distribution on key columns and indexes
> > all
> > then DB2 almost always makes the correct decision.
>
> My first query is not to return all data? *It's only returning on
> s.type_id = 4.
>
> Actually my questions arised when I changed the indexes on the table
> stock and the same queries run more than twice as long as before.
> Then I found the main index is not used at all. * Since the tables are
> huge (40 mil rows), I don't see why DB2 decides it's faster not to use
> the index.
>
> Maybe I should create an index on stock (type_id)? *Actually it
> eliminates a table scan on stock. * But why it is not using (stock_id,
> type_id) even though both columns are used in where clause.
>
> I know I must have some mis-conceptions. *I just don't know which ones.

You should create index (type_id, stock_id) on stock.

Yonglei

On Sep 14, 10:51*am, yongl...@gmail.com wrote:
> On Sep 13, 10:07*am, "Henry J." wrote:
>
>
>
> > On Sep 13, 12:34*am, "Mark A" wrote:
>
> > > "Henry J." wrote in message
>
> > >news:02e9b6d6-229f-459a-b8f7-0097b19ba2b1-at-m45g2000hsb.googlegroups.com....
>
> > > >I have a composite index on two columns in a table. *However, the
> > > > index is not used in a query that restricts the 2nd column to a
> > > > constant. *If both columns are linked with columns in other join
> > > > tables, the index will be used.
>
> > > > To illustrate it with an example, I have a query like this:
>
> > > > select s.ticker, p.quantity
> > > > from stock s, positions p
> > > > * * *where s.type_id = 4
> > > > * * * * * * * and s.stock_id = p.stock_id
>
> > > > There is an index on stock: (stock_id, type_id). *In the table stocks,
> > > > there are about 100 different type_ids evenly distributed, and ten of
> > > > thousand different stock_ids (each stock_id could map to every one of
> > > > the 100 type_ids).
>
> > > > From the plan, the above query does not use the index, but the
> > > > following query does use the index:
>
> > > > select s.ticker, p.quantity
> > > > from stock s, positions p
> > > > * * *where s.type_id = p.pos_type_id
> > > > * * * * * * * and s.stock_id = p.stock_id
>
> > > > The only difference is that in the first query, type_id is constant4,
> > > > while in the second it is linked with another column in the second
> > > > table.
>
> > > > The stats are good. *Is there anything else that might have caused the
> > > > above?
>
> > > You are confusing the use of an index with automatically getting better
> > > performance than not using an index. Sometimes using an index will result in
> > > better performance, and sometimes not. If DB2 determines that it is faster
> > > to not use an index, then it will not use one. This typically happenswhen
> > > DB2 will have read every data page (rows are stored in pages) to satisfy the
> > > query.
>
> > > Keep in mind that you are returning all rows from the table (you don't
> > > qualify which type_id or which stock_id's you want), so it is safe toassume
> > > that DB2 will have to read not just every page, but in fact every row, to
> > > satisfy the query.
>
> > > Assuming you did a runstats as follows:
> > > runstats on table with distribution on key columns and indexes
> > > all
> > > then DB2 almost always makes the correct decision.
>
> > My first query is not to return all data? *It's only returning on
> > s.type_id = 4.
>
> > Actually my questions arised when I changed the indexes on the table
> > stock and the same queries run more than twice as long as before.
> > Then I found the main index is not used at all. * Since the tables are
> > huge (40 mil rows), I don't see why DB2 decides it's faster not to use
> > the index.
>
> > Maybe I should create an index on stock (type_id)? *Actually it
> > eliminates a table scan on stock. * But why it is not using (stock_id,
> > type_id) even though both columns are used in where clause.
>
> > I know I must have some mis-conceptions. *I just don't know which ones.
>
> You should create index (type_id, stock_id) on stock.
>
> Yonglei

I think you are right -- if I create an index stock (type_id,
stock_id), the index will be used and no more table scan is done.

I created the index on (stock_id, type_id) because stock_id is much
more selective. Can't the DB2 optimizer figure out that this index
should be used?

BTW, as (stock_id, type_id) was the only index on the table, I'm able
to get rid of the table scan by making the table volatile.

I know DB2 optimizer is good, so what do I miss?

On Sep 14, 3:35*pm, "Henry J." wrote:
> On Sep 14, 10:51*am, yongl...@gmail.com wrote:
>
>
>
> > On Sep 13, 10:07*am, "Henry J." wrote:
>
> > > On Sep 13, 12:34*am, "Mark A" wrote:
>
> > > > "Henry J." wrote in message
>
> > > >news:02e9b6d6-229f-459a-b8f7-0097b19ba2b1-at-m45g2000hsb.googlegroups.com...
>
> > > > >I have a composite index on two columns in a table. *However, the
> > > > > index is not used in a query that restricts the 2nd column to a
> > > > > constant. *If both columns are linked with columns in other join
> > > > > tables, the index will be used.
>
> > > > > To illustrate it with an example, I have a query like this:
>
> > > > > select s.ticker, p.quantity
> > > > > from stock s, positions p
> > > > > * * *where s.type_id = 4
> > > > > * * * * * * * and s.stock_id = p.stock_id
>
> > > > > There is an index on stock: (stock_id, type_id). *In the table stocks,
> > > > > there are about 100 different type_ids evenly distributed, and ten of
> > > > > thousand different stock_ids (each stock_id could map to every one of
> > > > > the 100 type_ids).
>
> > > > > From the plan, the above query does not use the index, but the
> > > > > following query does use the index:
>
> > > > > select s.ticker, p.quantity
> > > > > from stock s, positions p
> > > > > * * *where s.type_id = p.pos_type_id
> > > > > * * * * * * * and s.stock_id = p.stock_id
>
> > > > > The only difference is that in the first query, type_id is constant 4,
> > > > > while in the second it is linked with another column in the second
> > > > > table.
>
> > > > > The stats are good. *Is there anything else that might have caused the
> > > > > above?
>
> > > > You are confusing the use of an index with automatically getting better
> > > > performance than not using an index. Sometimes using an index will result in
> > > > better performance, and sometimes not. If DB2 determines that it isfaster
> > > > to not use an index, then it will not use one. This typically happens when
> > > > DB2 will have read every data page (rows are stored in pages) to satisfy the
> > > > query.
>
> > > > Keep in mind that you are returning all rows from the table (you don't
> > > > qualify which type_id or which stock_id's you want), so it is safe to assume
> > > > that DB2 will have to read not just every page, but in fact every row, to
> > > > satisfy the query.
>
> > > > Assuming you did a runstats as follows:
> > > > runstats on table with distribution on key columns andindexes
> > > > all
> > > > then DB2 almost always makes the correct decision.
>
> > > My first query is not to return all data? *It's only returning on
> > > s.type_id = 4.
>
> > > Actually my questions arised when I changed the indexes on the table
> > > stock and the same queries run more than twice as long as before.
> > > Then I found the main index is not used at all. * Since the tables are
> > > huge (40 mil rows), I don't see why DB2 decides it's faster not to use
> > > the index.
>
> > > Maybe I should create an index on stock (type_id)? *Actually it
> > > eliminates a table scan on stock. * But why it is not using (stock_id,
> > > type_id) even though both columns are used in where clause.
>
> > > I know I must have some mis-conceptions. *I just don't know which ones.
>
> > You should create index (type_id, stock_id) on stock.
>
> > Yonglei
>
> I think you are right -- if I create an index stock (type_id,
> stock_id), the index will be used and no more table scan is done.
>
> I created the index on (stock_id, type_id) because stock_id is much
> more selective. * Can't the DB2 optimizer figure out that this index
> should be used?
>
> BTW, as (stock_id, type_id) was the only index on the table, I'm able
> to get rid of the table scan by making the table volatile.
>
> I know DB2 optimizer is good, so what do I miss?

For index1 (stock_id, type_id), DB2 needs to scan the whole index and
keep keys with type_id=4, while for index2 (type_id, stock_id), it
only needs to scan portion of index where type_id=4.

That's huge difference.