RE: Reorganize Big Table with BLOB column
Date: Sat, 6 Mar 2021 05:09:27 -0500
Message-ID: <018b01d71270$cb447f60$61cd7e20$_at_rsiz.com>
re: How can you have "nearly no I/O" but suffer from write complete waits? The two symptoms seem contradictory. What is the O/S saying about it?
IF your check for redo (undo applied) is true and massive AND there is a lot of block relocation of rows where the original first place are direct read and then it spins row at a time in the PGA reconstructing the correct vintage row image over and over and over again. Then a ton of all in cache work could be being done constructing the final image of the block you’re trying to write. That wait time probably shouldn’t be accounted for as write complete wait, leaving us with your “The two symptoms seem contradictory, ” unless there is an accounting problem. Or something else that hasn’t come to mind yet.
A bulk process that can be multi-threaded (meaning parallel at the independent jobs layer) is to carve up the space holding the object and read one row from each block (or otherwise drag your otherwise probably PGA reads through the SGA for permanent cleaning of the block whilst only having to process one row per block (and possibly only one column of one row per block).
Has anyone tried my suggestion to see if everything except the BLOB can be done quickly followed by ripping through that to copy the blob into a separate tablespace as out of line storage? If you’re going to have to row-by-row clean the blocks via SGA otherwise to get the context index to complete, I think that would be a faster way to get an untangled “don’t need no stinking undo applied to get read consistency” structure and then build the context index.
Likewise I endorse JL’s partition if you can – there is probably an n log n operation in there somewhere and your locally partitioned idea knocks “n” down to about “n/(number of partitions)” or just the largest partition if it his very uneven.
Head scratcher.
mwf
From: oracle-l-bounce_at_freelists.org [mailto:oracle-l-bounce_at_freelists.org] On Behalf Of Jonathan Lewis
Sent: Saturday, March 06, 2021 3:53 AM
To: Lothar Flatz
Cc: Oracle L
Subject: Re: Reorganize Big Table with BLOB column
-- http://www.freelists.org/webpage/oracle-lReceived on Sat Mar 06 2021 - 11:09:27 CET