Re: Same query with no plan change or volume but runs for hours vs minutes

From: Jonathan Lewis <jlewisoracle_at_gmail.com>
Date: Wed, 29 Nov 2023 22:38:37 +0000
Message-ID: <CAGtsp8kf57+y8sO4GNxGWFSfxGSJQjknrr-aujqd7vRoiDqtpA_at_mail.gmail.com>



I see I missed an M after the 9.8, but you obviously realised what I meant.

The update MIGHT explain why some of the partitions are much larger than others - but it would mean the number of rows updated, or the change in row length due to the update was quite variable. To explain: if you have the default pctfree = 10 and the update added 22 bytes to every row in the table then maybe one row in 20 would migrate, and that would be about the right number to give you the extra 500K blocks. But if some days the update was only added about 20 bytes then maybe none of the rows would migrate; or if you added 50 bytes to about 6 rows in every block then you might see the migration and growth, but if you added 48 bytes, or if one day the inserted bytes averaged a couple of bytes less, then you wouldn't see the migration and growth.

The "single partition target" makes this more puzzling; and I don't think the "DML restart" applies in this case - your original Monitor showed a single START for the insert, and I think (but not 100% certain) that you would see starts=3 for the insert if it had restarted.

There was one detail in the original Monitor that I avoided mentioning initially - but it may be the answer to all the oddities: One of the samples reported from ASH during the Load Table Conventional was:

enq: FB - contention (1)

That's "format blocks" - which means your session is waiting for someone else to format a set of blocks in the same segment, which means someone is inserting rows into the same segment at the same time. The fact that you're also reporting "gc current grant busy" etc. is also a little suspect because no-one should be competing with you on that object. (Though it's possible that these waits relate to dictionary objects or file headers if someone else is doing equally high volume on the other node.)

Regards
Jonathan Lewis

On Wed, 29 Nov 2023 at 21:26, Lok P <loknath.73_at_gmail.com> wrote:

> Thank you.
>
> Yet to validate many of the items suggested here but below is the answer
> to a few of questions from Jonathan..
>
>
> *Does the insert go into a single partition, or is it mostly one
> partition with a small percentage elsewhere? *
> INSERTs from this query will go to one partition because the value for
> partition key i.e. the date column is hard coded in the SELECT list of this
> INSERT query as a literal. And thus each day the sql_id changes but the
> plan remains the same for the insert query. We do update two of the
> columns(those are not part of this PK index though) of this table.
>
>
> *Is the primary key index local or global?*
> The primary key index is local.
>
> *Given your row size the 33M rows = 9.8 blocks is about right, so the
> excess 500K blocks might be a pointer to why the space allocation is
> behaving so badly. (But that might be cause and effect reversed). Have you
> cross-checked the occasions when the insert was slow with the partitions
> which are too large?*
>
> No, I see multiple partitions intermittently showing those higher numbers
> of blocks but the slow execution is not observed for all of those days
> holding a higher number of blocks. So basically the slowness pattern is not
> exactly matching for the days/partitions holding a higher number of blocks
> in dba_tab_partitions.
>
> On Wed, Nov 29, 2023 at 9:44 PM Rajeev Prabhakar <rprabha01_at_gmail.com>
> wrote:
>
>> One more note worth a look - 2057107.1
>>
>> On Nov 29, 2023 at 10:43 AM, <Rajeev Prabhakar <rprabha01_at_gmail.com>>
>> wrote:
>>
>> Lok,
>>
>> If you haven’t already, can you check if
>> Oracle note 1948087.1 is of any help/
>> relevance to your case ?
>>
>> Rajeev
>>
>>
>> On Nov 29, 2023 at 9:48 AM, <Mark W. Farnham <mwf_at_rsiz.com>> wrote:
>>
>> That all makes sense.
>>
>>
>>
>> And recreating the index in the hopes that the error was an unlikely
>> event, but which should make the problem go away.
>>
>>
>>
>> To the scenario possibilities I will add a delete operation that was
>> interrupted in an unfriendly way before the commit. Hmm, I wonder if an
>> aborted insert could burn space it shouldn’t permanently burn, leaving the
>> index leaf marked with plenty of space available, but the free space in the
>> block somehow trashed.
>>
>>
>>
>> All the “fixes” that come to mind would destroy this as a research case,
>> and if something in their operational schedule triggers an otherwise
>> unusual event, the usual thing that happens when problems that go away
>> unexplained will probably happen.
>>
>>
>>
>> Thanks for saving me from drilling into the actual details.
>>
>>
>>
>> It certainly would be tedious to dump the blocks of the index looking for
>> perversity, and then tying anything odd to a past event is likely to be an
>> anecdotal search. “What unusual thing did all y’all do after the last
>> ‘good’ run (that might have cleaned some things up) before the next ‘bad’
>> run?”
>>
>>
>>
>> Happy Holidays.
>>
>>
>>
>> *From:* oracle-l-bounce_at_freelists.org [mailto:
>> oracle-l-bounce_at_freelists.org] *On Behalf Of *Jonathan Lewis
>> *Sent:* Wednesday, November 29, 2023 4:45 AM
>> *To:* Oracle L
>> *Subject:* Re: Same query with no plan change or volume but runs for
>> hours vs minutes
>>
>>
>>
>> Mark,
>>
>>
>>
>> I'm still trying to work out a consistent story of what the names of the
>> statistics are trying to tell us, but it seems likely that part (possibly a
>> major part) of the problem relates to Oracle trying to find usable blocks
>> in the index segment. One very clear indicator is that the fast run doesn't
>> show a value for "ASSM gsp:Optimized index block rejects" whlie the slow
>> run reports 438M
>>
>>
>>
>> Other observations from the slow stats:
>>
>> ASSM gsp:reject db = 622315351
>>
>> ASSM cbk:blocks examined = 184M which is very close to: "reject db -
>> optimizer index blocks reject"
>>
>>
>>
>> I think this tells us that "optimized" means I checked a "bit" so I
>> didn't need to check the block it references.
>>
>>
>>
>> So when we see
>>
>> ASSM gsp:L2 bitmaps examined = 151,000
>>
>> ASSM gsp:Optimized reject l1 = 1,150
>>
>> ASSM gsp:L1 bitmaps examined = 42M
>>
>>
>>
>> I think that's telling us that there are lots of L1 bitmaps that should
>> have been flagged in the L2 as "no space in my data blocks", but haven't.
>>
>> Somewhere in the code path we have something is failing to set bits in
>> the L1.
>>
>>
>>
>> The trouble is, I don't think it's the insert that's doing it wrong.
>> Check these three stats:
>>
>> leaf node splits 4,602
>>
>> branch node splits 8
>>
>> ASSM cbk:blocks marked full 4,610
>>
>>
>>
>> We know that Oracle had to allocate (or re-use) 4,610 index leaf blocks
>> because of the block splits - and it has immediately marked those blocks as
>> "full" (index blocks are only ever "unformatted" "newly formatted",
>> "reusable (FS2)" or "in use (FULL)" as far as the bitmaps are concerned).
>>
>>
>>
>> The only ideas I have at present are:
>>
>> a) an index coalesce or a shrink space has taken place, freed up no (or
>> very few) blocks, but erroneously marked the leaf blocks as reusable.
>>
>> b) something happened to cause two copies of the code to run, and one
>> eventually failed or stuck; or the PL/SQL block controlling the insert
>> rolled back and restarted, and the internal code left lots of blocks in the
>> wrong state for some reason.
>>
>>
>>
>> Regards
>>
>> Jonathan Lewis
>>
>>
>>
>>
>>
>>
>>
>> On Tue, 28 Nov 2023 at 23:20, Mark W. Farnham <mwf_at_rsiz.com> wrote:
>>
>> “So ideally it should be a brand new segment without any fragmentation
>> etc.”
>>
>>
>>
>> ? Does that apply to both the table partition and the index?
>>
>>
>>
>> (A quick skim of what JL wrote makes me thing that index space available
>> for new inserts is the problem, but I really just glanced.)
>>
>>
>>
>> ? Do you insert rows with key values and lots of attributes null and then
>> immediately update the null attributes to the real values?
>>
>>
>>
>> (If you’re doing that, then do it someplace else (with a huge pctfree)
>> and copy the fully formed rows into the real destination. This interim
>> location does not need to be indexed, let the secondary insert into the
>> “real” destination handle any relational integrity issues, and if your
>> initial insert is really noisy with violations possibly test index it for
>> uniqueness after the initial insert.)
>>
>>
>>
>> ? What is the size of the initial extent of the new partition compared to
>> the amount of a batch being stuffed in?
>>
>>
>>
>> IF the answer to Dominic’s question is that your function is not direct
>> insert friendly, then think about putting the fully formed rows into an
>> interim location unindexed and then use those fully formed rows as a source
>> for direct inserts into the “real” destination.
>>
>>
>>
>> Good luck.
>>
>>
>>
>> mwf
>>
>>
>>
>> *From:* oracle-l-bounce_at_freelists.org [mailto:
>> oracle-l-bounce_at_freelists.org] *On Behalf Of *Dominic Brooks
>> *Sent:* Tuesday, November 28, 2023 4:24 PM
>> *To:* loknath.73_at_gmail.com
>> *Cc:* jlewisoracle_at_gmail.com; Oracle L
>> *Subject:* Re: Same query with no plan change or volume but runs for
>> hours vs minutes
>>
>>
>>
>> Is there a good reason for not using direct path inserts?
>>
>>
>>
>> Sent from my iPhone
>>
>>
>>
>> On 28 Nov 2023, at 21:03, Lok P <loknath.73_at_gmail.com> wrote:
>>
>> 
>>
>> Thank you Jonathan.
>>
>> So, It means the time spent is really on doing INSERT only but not the
>> "reading/select" part of the query. While loading , it spends a lot of time
>> scanning blocks that are full and thus unable to insert rows in them.
>> However, onething want to share here, that the target partition key/date is
>> hard coded in this query, which means the data inserted into a fresh new
>> blank partition each day. So ideally it should be a brand new segment
>> without any fragmentation etc.
>>
>> *"When I asked about a "catastrophic" event, I wanted you to think
>> whether there was any large-scale activity you might have done in the
>> recent past.* "
>>
>> In the recent past this database moved from the older version of
>> exadata to the latest X9 and also the database was moved to multitenant
>> PDB. How can that be related? But then why do we see the slowness on
>> certain days only but not daily when this data load happens?
>>
>>
>>
>> Also we see some high "block" count for almost the same amount num_rows,.
>> I.e for the same ~30million rows , some partitions showing the blocks as
>> ~1million and some are showing ~1.5million. Is this expected?.
>>
>>
>>
>> PARTITION_NAME NUM_ROWS BLOCKS AVG_ROW_LEN COMPRESSION
>> PBDRA_DTL_11272023_P 33734616 1487844 197 DISABLED
>> PBDRA_DTL_11262023_P 33469912 980070 198 DISABLED
>> PBDRA_DTL_11252023_P 33527112 994892 198 DISABLED
>> PBDRA_DTL_11242023_P 33948522 991827 198 DISABLED
>> PBDRA_DTL_11232023_P 32667974 993492 198 DISABLED
>> PBDRA_DTL_11222023_P 34653004 1529017 198 DISABLED
>> PBDRA_DTL_11212023_P 33233838 968435 198 DISABLED
>> PBDRA_DTL_11202023_P 32870086 1472631 198 DISABLED
>> PBDRA_DTL_11192023_P 32485084 947193 198 DISABLED
>>
>>
>>
>> Yet to find the space usage percentage from the new segments/partitions.
>> Hope i can use the scripts in the blog below for the same.
>>
>>
>>
>> https://jonathanlewis.wordpress.com/2016/09/11/space-usage/
>>
>>
>>
>> On Tue, Nov 28, 2023 at 11:38 PM Jonathan Lewis <jlewisoracle_at_gmail.com>
>> wrote:
>>
>>
>>
>> an index on (large sequence value, date_value) should store about 310 to
>> 320 index entries per leaf block when full.
>>
>> So the leaf node splits suggest you've inserted roughly 4602 * 310 = 1.4M
>> rows in the bad case and 40928 * 310 = 12.7M rows in the good case.
>>
>> These figures are in line with the execute count (etc.) of 1424230
>> and 12752800. This suggests that virtually every row that gets through the
>> other join predicates executes the select statement in fun1 and survives
>> the predicate check.
>>
>>
>>
>> The "index range scans" statistic match the "execution count" statistic
>> in both cases, and the various figures for consistent gets suggest that the
>> access path is the same in both cases (and is index only-only). So you can
>> discount something odd happening with the function.
>>
>>
>>
>> Take the execute count as the number of rows inserted
>>
>> Take "ASSM gsp: get free block" - "leaf node splits" - "branch node
>> splits" as the number of new blocks requested for the table.
>>
>> Divide "rows inserted" by "new table blocks" and the answer is about
>> 34.5: multiply that by 203 (avg_row_len + 5 for the row overhead) and the
>> answer is about 7,000 - so the figures about data size etc. hang together
>> quite well, and make it reasonable to look for the problem in "other
>> activity">
>>
>>
>>
>> So in the bad case the session requests some ASSM space for inserting
>> table rows 41,500 times (46,150 - 4,600). To find that space is examines
>>
>> To find that space Oracle examines 42.6M Level 1 bitmap blocks: (ASSM
>> gsp:L1 bitmaps examined 42601388)
>>
>> and 184M data blocks: (ASSM cbk:blocks examined 184273393)
>>
>> Of these(!), Oracle reports 633M rejected (ASSM gsp:reject db
>> 622315351)
>>
>>
>>
>> I think the inconsistency in the examined/rejected tells us that there's
>> an error somewhere in how Oracle is totalling up different subsets, and
>> there are lots of ways you can kick around all the examine/reject figures
>> to try and guess what they mean and how they've been summed incorrectly.
>>
>>
>>
>> What I do think, though, is that the 184M blocks examined, and the 42M L1
>> bitmap blocks example are where your code is burning its excess CPU in the
>> slow case -- checking blocks which aren't flagged as FULL, but can't accept
>> the data that needs to be inserted. In passing 184M + 42M = 226M, which is
>> most of the session logical I/O.
>>
>>
>>
>>
>>
>> When I asked about a "catastrophic" event, I wanted you to think whether
>> there was any large-scale activity you might have done in the recent past.
>> If I had wanted to know if you had done a big delete I would have asked if
>> you had done a big delete. We need to find a reason why there are lots of
>> blocks that seem to be marked as "space available" in the L1 bitmap, but
>> aren't usable once Oracle gets to the block. (At some point in the past
>> Oracle Corp. added code for a session to record a list of blocks that it
>> had visited unsuccessfully "a few" times so that it stopped visiting them,
>> but that may have changed since the initial implementation.)
>>
>>
>>
>> You could use the dbms_space.space_usage procedure to check the state of
>> the blocks in one of the target partitions to see if there are lots of
>> blocks that between 75 and 10% full. This might be a bit of information to
>> pass to Oracle.
>>
>>
>>
>>
>>
>> Regards
>>
>> Jonathan Lewis
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>> There exists only one composite (TXN_ID, WOD_DT) primary key index on
>> this table-PBDRA_DTL. This table is daily range partitioned on a date
>> column. And also the sequence -PBDRAD_SEQ does fed into the leading
>> column(TXN_ID) of the PK index.
>>
>>
>>
>> The AVG_ROW_LEN i see in dba_tab_partition for last few of the partitions
>> are showing ~198 and similar for others too.
>>
>> PARTITION_NAME NUM_ROWS BLOCKS AVG_ROW_LEN COMPRESSION
>> PBDRA_DTL_11272023_P 33734616 1487844 197 DISABLED
>> PBDRA_DTL_11262023_P 33469912 980070 198 DISABLED
>> PBDRA_DTL_11252023_P 33527112 994892 198 DISABLED
>> PBDRA_DTL_11242023_P 33948522 991827 198 DISABLED
>> PBDRA_DTL_11232023_P 32667974 993492 198 DISABLED
>> PBDRA_DTL_11222023_P 34653004 1529017 198 DISABLED
>> PBDRA_DTL_11212023_P 33233838 968435 198 DISABLED
>> PBDRA_DTL_11202023_P 32870086 1472631 198 DISABLED
>> PBDRA_DTL_11192023_P 32485084 947193 198 DISABLED
>>
>> Select list doesnt have any plsql function.
>>
>> Today I have captured the sql monitor and related stats from v$sesstat
>> for a ~5minutes delta period for a quick/good run in which the complete
>> query finished in ~15minutes. I have posted the sql monitor and stats in
>> same github link, The session stats i have posted as comments to the same
>> github linc.
>>
>> https://gist.github.com/oraclelearner/722a77ef5ebce23396d82e2835958303
>>
>> When you mentioned "*some catastrophic activity on the base table might
>> cause the ASSM related stats*", are you pointing to delete or anything?
>> Because we normally dont performa DELETE but just use partition DROP for
>> data purge on these tables. Below is the stats specific to ASSM i do
>> observe , and those i have not noticed in past frequently. And also as we
>> encounter this slow run intermittently, so not sure if we are encountering
>> any bugs related to ASSM here?
>>
>>
>>
>>

--
http://www.freelists.org/webpage/oracle-l
Received on Wed Nov 29 2023 - 23:38:37 CET

Original text of this message