Thanks Tanel,
I performed many tests considering all the scenarios and appears that this
fetch size has serious implications at the database level which basically
undermines the prefetch and batching capabilities at the IO level.
found the following cases where side affects of low fetch sizes in terms of
buffer gets and latches required are minimal to none.
- SORT ORDER BY to sort the results in which case the rows obtained
from the SQL execution are initially stored in the session’s PGA
- Use of Aggregate functions such as group by, avg, sum, count,
distinct etc.,
- most operations that results in storing the results in the PGA.
- Effect of reacquiring the latches decreases with the increase in
the row size but a low FETCHSIZE eventually results in more IOPS on server
or inefficient prefetching or batched access.
- Since the index leaf blocks can accommodate more entries, there is a
high chance that the index leaf block will be accessed multiple times
(provided that an indexed column value result in multiple rows).
Considering 90-10 and 50-50 splits, rebuilding an index which had more
50-50 leaf block splits will elevate the number of fetches as the index
leaf block now accommodates more rows. In such cases, it is better not to
build the index or increase the PCTFREE of an index should help, but
remember that increasing the PCTFREE reduces the efficiency of the buffer
cache as the blocks contain more free space.
- Index only scans are also affected since prefetching cannot be done for
range based queries as the next leaf block can be any location, unless an
index rebuild.
- Also found that if the application uses asynchronous programming or calls
other APIs for each result that they obtain causes a mess at the buffer
cache level especially if the blocks gets modified in the process and this
magnifies if the table is small mutating table where updates are more
frequent... we can never detect this unless we explicitly ask them (and the
session ends up waiting on sql*net message from client).
- short tables in which the clustering factor is incorrect (calculation
where the subsequent ROWIDs point to alternate blocks) in this case things
get even more ugly...
ORDER BY on the indexed column is also affected as the sorting step is
avoided... I was also looking at the ORM solutions such as hibernate which
generate queries, in these cases, if the application is heavily dependent
on such solutions, we have no control over the order by except increasing
the fetchsize to a very high value, but this takes much time if the
application uses FIRST_ROWS and most environments have a shared network
between the database and application servers...
Looks like there is more to this index range/skip scan.... have to do many
more tests to see any other aspects are also effected....
Thanks,
Vishnu
On Sun, Dec 29, 2019 at 2:11 AM Tanel Poder <tanel@xxxxxxxxxxxxxx> wrote:
Lower fetch size means less data is retrieved/sent back per *database
call*. Buffer pins (and held latches) must be released at the end of
every call to avoid some idle session holding these locks for too long. So
if you only ask for 5 rows per fetch, you end up re-visiting the same block
to get another 5 much more likely compared to fetching 500 rows every time.
Also you can try to just add an ORDER BY to the end of your test query
(and make sure that the ORDER BY operation actually shows up in the plan)
and the LIO/CBC latch get difference with different array sizes should
mostly go away as the ORDER BY step in the plan would use max internal
batch sizes to populate the query resultset in the sortarea in PGA/TEMP and
you'd fetch the results after that from there without needing CBC latches
or LIOs.
Tanel
On Wed, Dec 25, 2019 at 7:56 PM Vishnu Potukanuma <
vishnupotukanuma@xxxxxxxxx> wrote:
Got it,
looks like it is the default/expected behaviour...
since Oracle doesn't use asynchronous calls which is evident from the
fact that it uses only single thread per process model (server processes),
the logic flow must shift at point in time or another to a different
function to transmit the data (which can take more time and is dependent on
various factors such as network delays etc) to the client and returns to
the previous call where it has to mandatorily touch or read the blocks to
continue processing from where it left and this is completely dependent on
the fetchsize, lower the fetchsize more buffegets and latchgets..
looks like this is not the case all the time ...
Thanks,
Vishnu
On Wed, Dec 25, 2019 at 4:43 AM Vishnu Potukanuma <
vishnupotukanuma@xxxxxxxxx> wrote:
Hi,
The situation is as follows:
create table randomload(roll number, mark1 number, mark2 number, mark3
number, mark4 number);
load the table with random data;
then create an different table so that the clustering factor on mark4
column will be close to table blocks.
create table randomload2 as select * from randomload order by mark4;
create index idx on randomload2(mark4);
exec dbms_stats.gather_table_stats('VISHNU','RANDOMLOAD', CASCADE=>
TRUE);
now carefully consider any value from mark4 (indexed column) that
retrieves more than 500 rows (example) eventually we will find that 100
rows also causes high latching activity...
select * from randomload2 where mark4= 100;
see the buffergets for this SQL and also the carefully monitor the latch
activity on each of these latches holding these data blocks (CBC) latches.
i used a query such as the following:
select sum(gets) from v$latch_children where addr in
('000000006C9A4160','000000006CA29DA8','000000006CA37068','000000006CAFEF20','000000006CB52450','000000006CB52450','000000006CB52450','000000006CB91E28','000000006CB91EF0');
it is very interesting that the latch activity or buffer gets for a SQL
statement is dependent on the array size or fetch size..
average latch gets or buffer gets will always be higher when the fetch
size is small, even 100, the database repeatedly accesses the data blocks,
or latches...
testing with set fetchsize 20, 30, 60, 100, 200.
we will see different buffer gets per execution each time and latch
activity is very high...
even though the number of blocks accessed will be less (physical blocks
in my case only 6) since average row length is less and the mark4 column
had only 3200 distinct values).
Can someone please tell me
Is this expected? or I am missing any thing?
Thanks
Vishnu
