Sequences – where did the view go?

I was pottering around with the sequence cache recently, investigating a few things like exactly when it refreshes (such as if it is flushed or aged from the shared pool – pin it!) and I was monitoring the next value using V$_SEQUENCE, like I have been since, erm, Oracle 8i in 1999 or summit.

I happened to log into an Oracle 12.1.0.2 DB for a check-test and it was empty! My first reaction was, how come there are no sequences cached in the shared pool, shortly followed by No! The view doesn’t work any more!

Right, it’s a v$ view. (Actually, it’s a V$_ view, of which the only other one I know is V_$_lock). It is based upon c-memory-structure-plus-helper x$ “tables” (warning, these tables required tuning/diagnostic licensing packs to query!), so let’s query the underlying object view: x$kglob.

What used to happen pre-Oracle 12c:

SYS @ orcl > select * from v$version;
BANNER
--------------------------------------------------------------------------------
Oracle Database 11g Enterprise Edition Release 11.2.0.2.0 - Production

select test_seq.nextval from dual;
NEXTVAL
----------
40067

select sequence_owner,sequence_name,nextvalue,cache_size from v$_sequences where sequence_name = 'TEST_SEQ'
/
SEQUENCE_OWNER                 SEQUENCE_NAME                  NEXTVALUE  CACHE_SIZE
------------------------------ ------------------------------ ---------- ----------
NEIL                           TEST_SEQ                            40068         20

select KGLNAOBJ,
KGLOBTN0 next_val,
KGLOBTN4 cache_size,
kglhdkmk kept
from x$kglob
where KGLOBTYP=6
and KGLNAOBJ = 'TEST_SEQ'
/
KGLNAOBJ             NEXT_VAL   CACHE_SIZE KEPT
-------------------- ---------- ---------- ----------
TEST_SEQ                  40068         20          0

And now what happened in Oracle 12:

SQL> select * from v$version;
BANNER CON_ID
-------------------------------------------------------------------------------- ----------
Oracle Database 12c Enterprise Edition Release 12.1.0.2.0 - 64bit Production 0

select test_seq.nextval from dual;
NEXTVAL
----------
3

select sequence_owner,sequence_name,nextvalue,cache_size from v$_sequences where sequence_name = 'TEST_SEQ';

no rows selected

select KGLNAOBJ,
KGLOBTN0 next_val,
KGLOBTN4 cache_size,
kglhdkmk kept
from x$kglob
where KGLOBTYP=6
and KGLNAOBJ = 'TEST_SEQ';

KGLNAOBJ NEXT_VAL CACHE_SIZE KEPT
-------------------- ---------- ---------- ----------
TEST_SEQ 0

So the values are no longer exposed in Oracle 12c. NEXT_VAL and CACHE_SIZE are empty.

Does anybody know where to get this value? It's no longer in any column in X$KGLOB. Have Oracle hidden it in some weird container mode I don't understand yet? Can anybody help, or am I going to have to attach the SGA with a C program and step through a bunch of linked lists?

Sequences – a quick guide and an unexpected quirk

Sequences on Oracle databases are simple, but there are some basic truths about sequences which need to be understood if you are going to use them, especially on RAC:

  1. Sequences will have “holes” in them, even if you specify NOCACHE (e.g. if you rollback your transaction).
  2. Don’t specify NOCACHE. It doesn’t scale.
  3. Sequences will NOT necessarily be, erm, sequential for your transactions. That will depend upon the COMMIT order and if you are using RAC.
  4. The higher the CACHE value, the bigger the holes but the more scalable the sequence will be become. You will lose the entire set of unused cache values across instance restart.
  5. Oracle stores the CACHE value on EACH RAC node, so for a cache of 100,000, NODE 1 gets 1-100,000, NODE 2 get 100,001-200,000 to play with, and so on.
  6. If you are using RAC, and you have a heavy workload on one node and a low number of transactions on another node, there may be a significant disparity between the NEXTVAL on each node. I have seen this disparity manifest in the tens of millions.
  7. There is a global cache lock which must be passed from node to node when getting a new cache of sequence numbers on that node. It is lightweight, but it is a serialisation point. If your CACHE value is too low for your throughput, you will get contention on this lock.
  8. Same as the serialisation point if you specify ORDERED, so you get the sequences advancing in step (rather than each node having its own cache) on multiple RAC nodes. Don’t do this either.
  9. The point if sequences is to give you a unique value for a (surrogate) Primary Key. You can rely upon sequences to be unique. Well. Mostly. As long as they do not CYCLE.
  10. If you use sequences for a PK, you will probably end up with a right-hand index with hot-block contention during inserts, causing you to implement something suboptimal, like hash partitioning or reverse-key indexes to get around your design flaw.

Sequences on RAC can behave in unexpected ways. The below example is a little contrived but shows a potentially unexpected behaviour caused by CYCLEing your sequences:

NODE 1 NODE 2
create sequence seq_1
start     with 1
increment by   1
maxvalue       10
cache          2
cycle;

Sequence created.

NODE1 :select seq_1.nextval from dual;
NEXTVAL: 1
NODE2 :select seq_1.nextval from dual;
NEXTVAL
3 <- Gets next value above cache on node 1
NODE1 : /
NEXTVAL: 2
NODE1 : /
NEXTVAL: 5 (skipping 3 and 4 on Node 2)
6
7
8
9
10
1
2
3  <- erm. Hang on. Node 2 has this!
4  <- and this!!!
5
6
7
8
9
10
1
NODE2 : /
NEXTVAL
4 <- Reading from the local cache
NODE2 : /
NEXTVAL
3 <- And get a new cache value

So, if you are using sequences and you can’t guarantee sequence and you can’t guarantee no gaps and, as this demo shows, you can’t guarantee uniqueness, and they cause hot block right hand index problems, tell me again why you are using them?

Because they are fast and you don’t CYCLE them.

However, sequences are becoming the new evil in the database, right behind triggers.

%d bloggers like this: