This whole area is one that has to be tested at scale (and old and new
versions of Oracle may behave differently), with some concurrency; and you
may find that non-unique and unique indexes behave differently. And there
are 4 different ways to do the cleanup, so there's a potential for a lot of
testing.
I was suprised by Andy's comment about locking problems - though if you
have a unique index with a row in one partition and drop the partition and
then start a cleanup, an insert of the same unique key into another
partition may see some odd locking problems (which may depend on whether
the cleanup gets to that value first or the new insert gets there first) as
the insert will want to re-use the index entry and update the rowid
"carried data" component. This is the sort of thing that doesn't get
described in the manuals and can be very hard to determine in small tests;
and it's the sort of thing that customers discover and report to Oracle and
they fix in the next release and create patches for.
The biggest problem for DBAs is that there are some things that you'll only
see as problems when you're operating at a scale that's larger than
Oracle's internal tests and do something is different from Oracle's test
cases.
Regards
Jonathan Lewis
On Sat, 8 Jan 2022 at 17:30, Lok P <loknath.73@xxxxxxxxx> wrote:
Thank you very much Jonathan and Andy.
Thanks for correcting me. So basically in case of a unique constraint ,
the index has to be global index or global partitioned index but it can't
be a local index because that will be then unique for each day/partition
but won't be at the table lebel. So considering that fact , we were
lucky in our case as in most scenarios we were good with the local unique
index which is not exactly unique at the table level.
But it's true that we can't avoid the global index for a scenario in which
the partition key(say date column) does not necessarily have to be part of
the uniqueness. And in those cases we have to come across the index
maintenance overhead stuff while dropping/truncating/purging data from the
partition table.
So is there any way to get this work on a 24/7 system? Say as we move
ahead, we want to drop/purge one daily range partition from the history
each day and in that case the global index(or say primary key index) has to
be rebuilt after the drop partition immediately(but that can cause issue if
it get unusable+rebuild in a 24/7 system) or has to pass through a delayed
maintenance phase through 'update indexes' syntax. So in this case even as
Andy mentioned there are chances of blocking while the maintenance is
happening, so is there a real online way of achieving this global index
requirement without making it unusable for big partitioned tables?
Or it has to be then partitioned (say e.g hash partitioned through the
same PK column and then it would not block others sessions during the
delayed cleanup?
On Sat, Jan 8, 2022 at 3:59 PM Jonathan Lewis <jlewisoracle@xxxxxxxxx>
wrote:
Forget about adding the partition key to indexes that are currently
unique in order to make them local.
The uniqueness requirement is about the CONSTRAINT, not the index. If
you add the partition key to the unique constraing the original unique
condition is no longer unique so you could get duplicates spread across
different partitions.
e.g. if order_id is considered to be unique but order_date is the
partition key, then
a) the index on (order_id) can be unique, but has to be global (or
globally partitioned).
or
b) the index on (order_id, order_date) can be unique and local, but only
if you now allow duplicates of order_id.
or
c) the index on (order_id, order_data) can be non-unique and local, but
if you create a unique constraint on order_id Oracle will create a global
unique index on (order_id)
In this example (where you are not likely to execute a query "where
order_id between X and Y") a damage-limiting exercise is to create the
index as globally hash partitioned so that "order_id = X" requires only one
index partition to be visitied and the index partitions are reduced in size
hence easier to handle when doing house-keeping. (e.g. when rebuilding).
Regards
Jonathan Lewis
On Sat, 8 Jan 2022 at 06:52, Lok P <loknath.73@xxxxxxxxx> wrote:
Thank You So much Andy.
When you said "*And worthy of note (though kind of obvious), if the
clean jobs aren't able to complete then those segments are continuously
growing*.", Do you mean that if the delayed index maintenance/cleanup
job is not getting completed but still we keep on dropping partitions in
subsequent days. This can cause the global index to grow substantially and
then it can impact the performance of the queries those are using this
index access path in them?
As you mentioned , there are cases in which this index cleanup job can
block the other sessions , which looks scary on a 24/7 system. And even
for unique indexes we are being asked/forced to have the partition key
included in the index keys so as to make it unique and local. So ,overall
It means we are no better even on the higher/recent versions of Oracle for
considering creation of global indexes on partitioned tables. Correct me if
I'm wrong.
On Sat, Jan 8, 2022 at 2:38 AM Andy Klock <andy@xxxxxxxxxxxxx> wrote:
Hi Lok,
‐‐‐‐‐‐‐ Original Message ‐‐‐‐‐‐‐
On Friday, January 7th, 2022 at 3:35 PM, Lok P <loknath.73@xxxxxxxxx>
wrote:
So my question was , how feasible its is in reality , if its reliable
one and now we can think of creating global indexes without hesitation now
on 12.1 onward where-ever needed?
Global indexes are still feasible and sometimes necessary (unique
constraints, for example), but because of the maintenance you speak of,
local indexes would be preferred where appropriate.
Is this 12.1 version, deferred global index maintenance going to take
same time as it used to happen during 11.2 version with immediate 'update
indexes' option or its faster? Or any other downside, say if during this
auto deferred maintenance it will block any DML/SELECT statements etc?
Partition DROPS and TRUNCATES are measurably faster now that Oracle is
deferring the cleaning up of orphaned entries, but your concerns are still
very valid. Depending on size, I've had cases where cleaning wasn't able to
complete within the maintenance window, even with parallelism enabled.
There were also cases where sessions were being blocked by the clean job
and those segments had to be deferred and index rebuilds had to be done
during downtime windows. And worthy of note (though kind of obvious), if
the clean jobs aren't able to complete then those segments are continuously
growing.
Andy K