After it finally finishes parsing, did the plan end up serial?
There is a specific bug where the children don't get the same plan as the
coordinator, and they try again and again until some limit or timeout is
reached and then you get the one original plan running serially. (I was
puzzled why the hash wasn't just passed instead of re-parsing, but that's
another issue unexplored.)
Unfortunately all my notes on this were behind a secure area (can't take
pictures or forward the email, even if you could hack it you agreed not to
try) and the customer throws all that away every 90 days and flatlines your
PC hard drive to re-use and discards your host linux server files at the end
of the project so I can't look back and type you a specific number or
combination of numbers.
Pantloads of cursor: pin S wait on X event were a feature of this bug. That
was however 12.1.x. I'm not sure whether the patch to fix this or turning
off adaptive plans or both were required to make the bad behavior go away. I
don't know if this can happen on 11.x.
Graham's colleague Mihajlo Tekic probably has accurate notes on this because
he helped solve it and had seen it before and Oracle was magically allowed
to transport their notes wherever they liked.
Twas brillig.
Possibly this behavior is already well-known, I just hadn't seen it before.
mwf
From: oracle-l-bounce@xxxxxxxxxxxxx [mailto:oracle-l-bounce@xxxxxxxxxxxxx]
On Behalf Of Dominic Brooks
Sent: Tuesday, April 17, 2018 5:14 AM
To: Stefan Knecht; oracle-l-freelists
Subject: RE: Crazy dynamic sampling?
It's not the subsequent mutexes that I'm bothered about.
More about why the session at the head of the queue is hard parsing for
minutes or hours.
_____
From: Stefan Knecht <knecht.stefan@xxxxxxxxx>
Sent: Tuesday, April 17, 2018 10:09:35 AM
To: Dominic Brooks; oracle-l-freelists
Subject: Re: Crazy dynamic sampling?
Not seen that on 11.2.0.4.
What mutex and location exactly are the sessions waiting on? You can hit
x$mutex_sleep from the p2 in v$session for the cursor: pin S wait on X
event. Some more details on how to do thatare here: WAITEVENT: "cursor: pin
S wait on X" Reference Note (Doc ID 1298015.1)
Then I'd search MoS for bugs related to that mutex location.
On Tue, Apr 17, 2018 at 3:55 PM, Dominic Brooks <dombrooks@xxxxxxxxxxx>
wrote:
Good question - sorry for missing that out.
11.2.0.4
_____
From: Stefan Knecht <knecht.stefan@xxxxxxxxx>
Sent: Tuesday, April 17, 2018 9:54:57 AM
To: Dominic Brooks; oracle-l-freelists
Subject: Re: Crazy dynamic sampling?
Are you on 12.1.0.2 ?
On Tue, Apr 17, 2018 at 3:49 PM, Dominic Brooks <dombrooks@xxxxxxxxxxx>
wrote:
In addition to invisible indexes not being so invisible, I hash partitioned
some global PK indexes to reduce insert hot block contention.
These two things - invisible local indexes with some unusable subpartitions
and hash partitioned PK indexes - are the only things which changed.
I've now got a few handfuls of statements which, in addition to some of them
doing table expansion thanks to the not-so-invisible invisible indexes, are
doing crazy long hard parsing - like 30 to 40 minutes. And that causes
knock-on effects to other sessions in the same workflow for the same sql /
objects with "cursor: pin S wait on X" and "library cache lock"
The likely culprit is dynamic sampling. I can see the sessions doing lots of
IO against different table partitions as part of the hard parse.
Otherwise the health of the database is no less healthy than it was before
and the SQL is the same SQL as it was before.
Whilst I wait for an optimizer trace file and some other dumps/traces, any
wild theories as to why just hash partitioning a global PK might cause this?
_____
From: Dominic Brooks <dombrooks@xxxxxxxxxxx>
Sent: Tuesday, April 10, 2018 11:09:03 AM
To: oracle-l@xxxxxxxxxxxxx
Subject: Re: Invisible indexes and table expansion
Probably just covered by bug 16544878.
Sent from my iPhone
On 10 Apr 2018, at 10:50, Dominic Brooks <dombrooks@xxxxxxxxxxx> wrote:reproduce. And it's difficult for me to get an optimiser trace file in the
Just working through a problem, gathering evidence and trying to
table expansion - VW_TE_2.
In an 11.2.0.4 environment, I've got a whole bunch of queries now using
partitions as usable - something which table expansion was designed to help
Recent change was to add three local indexes, each with a subset of
to cause table expansion - but then can't be used.
But they are invisible indexes.
So my theory is the invisibility is limited - ie they are visible enough
Anyone looked into something similar?
Cheers
Dominic
Sent from my iPhone
