Thank you Lothar.
We do have big partition tables having historical partitions compressed
with 'query high'. So as you mentioned, ideally they should all be by
default stored as in columnar format and it should not be required to scan
blocks in terms of row store. So then why am I not seeing the statistics ''
'cell physical IO bytes saved by columnar cache'. or is it DB or cell
server version dependent? We have some databases still on 11.2.0.4
version, not sure if that can be a factor.
Can you please shed some more light on what exact statistics can be queried
from dba_hist_active_sess_history for sql_ids to link those, if any
significant IO can be reduced using such technique?
On Wed, Sep 22, 2021 at 1:18 PM Lothar Flatz <l.flatz@xxxxxxxxxx> wrote:
In Addition once you use hybrid columnar compression (HCC) the data is
stored in columnar format. (Am I mistaken? I am irritated by the fact that
nobody mentioned that simple fact.)
In an ETL Job there are phases where data is checked for valid keys and
additional data is added (e.g. from dimension tables) to simplify queering.
In such phases columnar storage is most likely to help. You often use only
a few columns of a table. You can consider using HCC on selected tables.
Columnar storage is likely to help less when you are doing aggregation of
some sort. (Although nothing is sure unless you check it.)
Some queries on DBA_HIST_ACTIVE_SESS_HISTORY should tell you were the time
is spend.
Regards
Lothar
Am 21.09.2021 um 23:41 schrieb Tanel Poder:
Exadata storage cells (starting from cellsrv v12.1.2.1.0 / Jan 2015) can
use fully columnar flash cache for greatly speeding up reads (the cache is
fully columnar, not hybrid like the datafile storage).
You can look into slides 17-20 in this presentation (from 2015), some
things may have changed/improved by now:
https://tanelpoder.com/2015/03/24/oracle-exadata-performance-latest-improvements-and-less-known-features/
This would speed up queries doing lots of scanning *if* your current SQL
performance bottleneck is about reading too many datafile blocks (and not
somewhere else, like having a large fact-fact hash join spill to temp).
This columnar flash caching won't speed up writes & large data loads.
So, if the goal is to speed up your ETL processing, you should first
measure (if you haven't done so already) where the response time of your
ETL jobs is spent... and see if it's in *smart scan data retrieval*,
where the storage cells can't somehow keep up with the data ingest demand
of the DB layer and they waif for disk reads a lot (but I'd say it's
unlikely... depending on how your flash cache allowance is set up. You
might already be benefitting from the columnar flash cache, the slides have
metrics that show how much).
Otherwise, just to reduce I/O (which kind of I/O - data retrieval for
scans or data processing for GROUP BY/HASH JOINs?), you could review your
top SQLs' execution plans and see if a better partitionings,
subpartitioning (and maybe even Exadata-aware indexing) scheme would help
to allow doing less I/O. I would look into fancier things like attribute
clustering (and zone maps) after done with the basics first.
--
Tanel Poder
https://tanelpoder.com
On Tue, Sep 21, 2021 at 2:48 PM Lok P <loknath.73@xxxxxxxxx> wrote:
Hello Listers, We have oracle exadata databases which are performing
mostly warehousing or batch type data processing. And few are hybrid i.e
combination of both oltp+warehousing/analytics types processing. ETL jobs
run on a few of these which move/read/write billions rows daily.
The databases are 50TB to ~150TB in size currently. Few architecture
team members suggested evaluating if we can use columnar database type of
offering for IO reduction and thus better performance considering the
future growth. As per my understanding , Oracle stores data in row format
only, so is there any other offering from Oracle for columnar datastore
format or columnar databases and we should evaluate that? Or is there any
clustering kind of technique which can be evaluated which will help reduce
IO? Want to understand experts' views here on this.
Regards
Lok
