Sequences in Oracle 10g RAC

Just recently I got a call from a developer. He had a table with a primary key populated by a sequence, a timestamp column with the current date and some other columns. He had a specific set of data that, when ordered by the primary key had out of order timestamps. He was puzzled how this could be. This is a RAC database and the sequence was created with the default values.

Not only the sequences cache was the default of 20, but it was “noordered”. Being “noordered” Oracle will not guarantee the order in which numbers are generated.

Example of “noorder” sequence in 10g RAC:

Session 1 on node-A: nextval -> 101
Session 2 on node-A: nextval -> 102
Session 1 on node-B: nextval -> 121
Session 1 on node-B: nextval -> 122
Session 1 on node-A: nextval -> 103
Session 1 on node-A: nextval -> 104

The sequence cache is in the shared pool, therefore sessions on the same node can share the cached entry, but sessions on different nodes cannot. I wonder why Oracle doesn’t make “ordered” the default for sequences.

So I explained to the developer how sequences work in RAC and how each node has its own “cache”.

We changed the sequence to “ordered” and increased the cache to 1000. Now selecting on either node gets the next number as he expected. I warned him that there would be some performance implications due to cluster synchronization. Him been a responsive developer, asked me what would be the impact, so I tested it out.

How does RAC synchronize sequences?

In Oracle 10g RAC, if you specify the “ordered” clause for a sequence, then a global lock is allocated by the node when you access the sequence.

This lock acquisition happens only at the first sequence access for the node (A), and subsequent uses of the sequence do not wait on this lock. If another node (B) selects from that sequence, it requests the same global lock and once acquired it returns the sequences next value.

The wait event associated with this activity is recorded as “events in waitclass Other” when looked in gv$system_event. So much for event groups, it couldn’t be more obscure. That view shows overall statistics for the session.

However if you look in the gv$session_wait_history it shows as “DFS lock handle” with the “p1” parameter been the object_id of the sequence. This second view has a sample of the last 10 wait events for a session.

In a SQL_TRACE with waitevents (10046 trace) it will be a “DFS lock handle” but in AWR or statspack reports it will be “events in waitclass Other”. So much for consistency.

How does that change our example?

Session 1 on node-A: nextval -> 101 (DFS Lock handle) (CR read)
Session 2 on node-A: nextval -> 102
Session 1 on node-B: nextval -> 103 (DFS Lock handle)
Session 1 on node-B: nextval -> 104
Session 1 on node-A: nextval -> 105 (DFS Lock handle)
Session 1 on node-A: nextval -> 106
(more selects)
Session 1 on node-A: nextval -> 998
Session 1 on node-B: nextval -> 999 (DFS Lock handle)
Session 1 on node-B: nextval -> 1000 (CR read)

The cache size also has some RAC synchronization implications. When the cached entries for the sequence are exhausted, the sequence object needs to be updated. This usually causes a remote CR (current read) over the interconnect for the block that has the specific sequence object. So a bit more activity here.

Test case:

create sequence test_rac;
declare
  dummy number;
begin
for i in 1..50000 loop
  select test_rac.nextval into dummy from dual;
end loop;
end;
/

Results:

50 000 loops with cache = 20 (default)
1 node = 5 seconds
2 nodes at same time = 14 seconds
2 nodes at same time ordered = 30 seconds

50 000 loops with cache = 1000

1 node = 1.5 seconds
2 nodes at same time = 1.8 seconds
2 nodes at same time ordered = 20 seconds

With a smaller cache, the “noordered” still has as significant impact as every 10 fetches (cache 20 divided by 2 nodes fetching) it has to synchronize between the 2 nodes

The conclusion

By default sequences in 10g RAC are created without ordering. Beware of using applications that rely on sequences to be ordered and using it in a RAC environment.

Consider changing all user sequences to “ordered” as a precaution and increasing the cache size.

The default cache value is still very low and even not-ordered sequences will cause contention in a highly-active sequence even in non-RAC and causing an additional block exchange every 20 values in RAC.

For high volume insert operations where ordering is not performed on the value returned from the sequence, consider leaving the sequence “noordered” but increasing the cache size significantly.

Either way, the sequence parameters should be reviewed, as chances are, the defaults are not what you need.

I remember reading somewhere that in Oracle 9i the “ordered” clause in RAC was equivalent to “nochache”. I can’t imagine how bad that would be in concurrent selects from the same sequence.

It would be interesting if someone running 9i RAC performs the test case and I would appreciate if you post the results in the comments.