Pre-loading Oracle text in memory with Oracle 12c

Question

There is a white paper from Roger Ford that explains how to load the Oracle index in memory : http://www.oracle.com/technetwork/database/enterprise-edition/mem-load-082296.html
In our application, Oracle 12c, we are indexing a big XML field (which is stored as XMLType with storage secure file) with the PATH_SECTION_GROUP. If I don't load the I table (DR$..$I) into memory using the technique explained in the white paper then I cannot have decent performance (and especially not predictable performance, it looks like if the blocks from the TOKEN_INFO columns are not memory then performance can fall sharply)
But after migrating to oracle 12c, I got a different problem, which I can reproduce: when I create the index it is relatively small (as seen with ctx_report.index_size) and by applying the technique from the whitepaper, I can pin the DR$ I table into memory. But as soon as I do a ctx_ddl.optimize_index('Index','REBUILD') the size becomes much bigger and I can't pin the index in memory. Not sure if it is bug or not.
What I found as work-around is to build the index with the following storage options:
ctx_ddl.create_preference('TEST_STO','BASIC_STORAGE');
ctx_ddl.set_attribute ('TEST_STO', 'BIG_IO', 'YES' );
ctx_ddl.set_attribute ('TEST_STO', 'SEPARATE_OFFSETS', 'NO' );
so that the token_info column will be stored in a secure file. Then I can change the storage of that column to put it in the keep buffer cache, and write a procedure to read the LOB so that it will be loaded in the keep cache. The size of the LOB column is more or less the same as when creating the index without the BIG_IO option but it remains constant even after a ctx_dll.optimize_index. The procedure to read the LOB and to load it into the cache is very similar to the loaddollarR procedure from the white paper.
Because of the SDATA section, there is a new DR table (S table) and an IOT on top of it. This is not documented in the white paper (the white paper was written for Oracle 10g). In my case this DR$ S table is much used, and the IOT also, but putting it in the keep cache is not as important as the token_info column of the DR I table. A final note: doing SEPARATE_OFFSETS = 'YES' was very bad in my case, the combined size of the two columns is much bigger than having only the TOKEN_INFO column and both columns are read.
Here is an example on how to reproduce the problem with the size increasing when doing ctx_optimize
1. create the table
drop table test;
CREATE TABLE test
(ID NUMBER(9,0) NOT NULL ENABLE,
XML_DATA XMLTYPE
XMLTYPE COLUMN XML_DATA STORE AS SECUREFILE BINARY XML (tablespace users disable storage in row);
2. insert a few records
insert into test values(1,'<Book><TITLE>Tale of Two Cities</TITLE>It was the best of times.<Author NAME="Charles Dickens"> Born in England in the town, Stratford_Upon_Avon </Author></Book>');
insert into test values(2,'<BOOK><TITLE>The House of Mirth</TITLE>Written in 1905<Author NAME="Edith Wharton"> Wharton was born to George Frederic Jones and Lucretia Stevens Rhinelander in New York City.</Author></BOOK>');
insert into test values(3,'<BOOK><TITLE>Age of innocence</TITLE>She got a prize for it.<Author NAME="Edith Wharton"> Wharton was born to George Frederic Jones and Lucretia Stevens Rhinelander in New York City.</Author></BOOK>');
3. create the text index
drop index i_test;
  exec ctx_ddl.create_section_group('TEST_SGP','PATH_SECTION_GROUP');
begin
  CTX_DDL.ADD_SDATA_SECTION(group_name => 'TEST_SGP', 
                            section_name => 'SData_02',
                            tag => 'SData_02',
                            datatype => 'varchar2');
end;
exec ctx_ddl.create_preference('TEST_STO','BASIC_STORAGE');
exec  ctx_ddl.set_attribute('TEST_STO','I_TABLE_CLAUSE','tablespace USERS storage (initial 64K)');
exec  ctx_ddl.set_attribute('TEST_STO','I_INDEX_CLAUSE','tablespace USERS storage (initial 64K) compress 2');
exec  ctx_ddl.set_attribute ('TEST_STO', 'BIG_IO', 'NO' );
exec  ctx_ddl.set_attribute ('TEST_STO', 'SEPARATE_OFFSETS', 'NO' );
create index I_TEST
  on TEST (XML_DATA)
  indextype is ctxsys.context
  parameters('
    section group   "TEST_SGP"
    storage         "TEST_STO"
  ') parallel 2;
4. check the index size
select ctx_report.index_size('I_TEST') from dual;
it says :
TOTALS FOR INDEX TEST.I_TEST
TOTAL BLOCKS ALLOCATED:                                                104
TOTAL BLOCKS USED:                                                      72
TOTAL BYTES ALLOCATED:                                 851,968 (832.00 KB)
TOTAL BYTES USED:                                      589,824 (576.00 KB)
4. optimize the index
exec ctx_ddl.optimize_index('I_TEST','REBUILD');
and now recompute the size, it says
TOTALS FOR INDEX TEST.I_TEST
TOTAL BLOCKS ALLOCATED:                                               1112
TOTAL BLOCKS USED:                                                    1080
TOTAL BYTES ALLOCATED:                                 9,109,504 (8.69 MB)
TOTAL BYTES USED:                                      8,847,360 (8.44 MB)
which shows that it went from 576KB to 8.44MB. With a big index the difference is not so big, but still from 14G to 19G.
5. Workaround: use the BIG_IO option, so that the token_info column of the DR$ I table will be stored in a secure file and the size will stay relatively small. Then you can load this column in the cache using a procedure similar to
alter table DR$I_TEST$I storage (buffer_pool keep);
alter table dr$i_test$i modify lob(token_info) (cache storage (buffer_pool keep));
rem: now we must read the lob so that it will be loaded in the keep buffer pool, use the prccedure below
create or replace procedure loadTokenInfo is
  type c_type is ref cursor;
  c2 c_type;
  s varchar2(2000);
  b blob;
  buff varchar2(100);
  siz number;
  off number;
  cntr number;
begin
    s := 'select token_info from  DR$i_test$I';
    open c2 for s;
    loop
       fetch c2 into b;
       exit when c2%notfound;
       siz := 10;
       off := 1;
       cntr := 0;
       if dbms_lob.getlength(b) > 0 then
         begin
           loop
             dbms_lob.read(b, siz, off, buff);
             cntr := cntr + 1;
             off := off + 4096;
           end loop;
         exception when no_data_found then
           if cntr > 0 then
             dbms_output.put_line('4K chunks fetched: '||cntr);
           end if;
         end;
       end if;
    end loop;
end;
Rgds, Pierre

Answer 1

I have been working a lot on that issue recently, I can give some more info.
First I totally agree with you, I don't like to use the keep_pool and I would love to avoid it. On the other hand, we have a specific use case : 90% of the activity in the DB is done by queuing and dbms_scheduler jobs where response time does not matter. All those processes are probably filling the buffer cache. We have a customer facing application that uses the text index to search the database : performance is critical for them.
What kind of performance do you have with your application ?
In my case, I have learned the hard way that having the index in memory (the DR$I table in fact) is the key : if it is not, then performance is poor. I find it reasonable to pin the DR$I table in memory and if you look at competitors this is what they do. With MongoDB they explicitly says that the index must be in memory. With elasticsearch, they use JVM's that are also in memory. And effectively, if you look at the awr report, you will see that Oracle is continuously accessing the DR$I table, there is a SQL similar to
SELECT /*+ DYNAMIC_SAMPLING(0) INDEX(i) */    
TOKEN_FIRST, TOKEN_LAST, TOKEN_COUNT, ROWID    
FROM DR$idxname$I
WHERE TOKEN_TEXT = :word AND TOKEN_TYPE = :wtype    
ORDER BY TOKEN_TEXT,  TOKEN_TYPE,  TOKEN_FIRST
which is continuously done.
I think that the algorithm used by Oracle to keep blocks in cache is too complex. A just realized that in 12.1.0.2 (was released last week) there is finally a "killer" functionality, the in-memory parameters, with which you can pin tables or columns in memory with compression, etc. this looks ideal for the text index, I hope that R. Ford will finally update his white paper :-)
But my other problem was that the optimize_index in REBUILD mode caused the DR$I table to double in size : it seems crazy that this was closed as not a bug but it was and I can't do anything about it. It is a bug in my opinion, because the create index command and "alter index rebuild" command both result in a much smaller index, so why would the guys that developped the optimize function (is it another team, using another algorithm ?) make the index two times bigger ?
And for that the track I have been following is to put the index in a 16K tablespace : in this case the space used by the index remains more or less flat (increases but much more reasonably). The difficulty here is to pin the index in memory because the trick of R. Ford was not working anymore.
What worked:
first set the keep_pool to zero and set the db_16k_cache_size to instead. Then change the storage preference to make sure that everything you want to cache (mostly the DR$I) table come in the tablespace with the non-standard block size of 16k.
Then comes the tricky part : the pre-loading of the data in the buffer cache. The problem is that with Oracle 12c, Oracle will use direct_path_read for FTS which basically means that it bypasses the cache and read directory from file to the PGA !!! There is an event to avoid that, I was lucky to find it on a blog (I can't remember which, sorry for the credit).
I ended-up doing that. the events to 10949 is to avoid the direct path reads issue.
alter session set events '10949 trace name context forever, level 1';
alter table DR#idxname0001$I cache;
alter table DR#idxname0002$I cache;
alter table DR#idxname0003$I cache;
SELECT /*+ FULL(ITAB) CACHE(ITAB) */ SUM(TOKEN_COUNT),  SUM(LENGTH(TOKEN_INFO)) FROM DR#idxname0001$I;
SELECT /*+ FULL(ITAB) CACHE(ITAB) */ SUM(TOKEN_COUNT),  SUM(LENGTH(TOKEN_INFO)) FROM DR#idxname0002$I;
SELECT /*+ FULL(ITAB) CACHE(ITAB) */ SUM(TOKEN_COUNT),  SUM(LENGTH(TOKEN_INFO)) FROM DR#idxname0003$I;
SELECT /*+ INDEX(ITAB) CACHE(ITAB) */  SUM(LENGTH(TOKEN_TEXT)) FROM DR#idxname0001$I ITAB;
SELECT /*+ INDEX(ITAB) CACHE(ITAB) */  SUM(LENGTH(TOKEN_TEXT)) FROM DR#idxname0002$I ITAB;
SELECT /*+ INDEX(ITAB) CACHE(ITAB) */  SUM(LENGTH(TOKEN_TEXT)) FROM DR#idxname0003$I ITAB;
It worked. With a big relief I expected to take some time out, but there was a last surprise. The command
exec ctx_ddl.optimize_index(idx_name=>'idxname',part_name=>'partname',optlevel=>'REBUILD');
gqve the following
ERROR at line 1:
ORA-20000: Oracle Text error:
DRG-50857: oracle error in drftoptrebxch
ORA-14097: column type or size mismatch in ALTER TABLE EXCHANGE PARTITION
ORA-06512: at "CTXSYS.DRUE", line 160
ORA-06512: at "CTXSYS.CTX_DDL", line 1141
ORA-06512: at line 1
Which is very much exactly described in a metalink note 1645634.1 but in the case of a non-partitioned index. The work-around given seemed very logical but it did not work in the case of a partitioned index. After experimenting, I found out that the bug occurs when the partitioned index is created with  dbms_pclxutil.build_part_index procedure (this enables  enables intra-partition parallelism in the index creation process). This is a very annoying and stupid bug, maybe there is a work-around, but did not find it on metalink
Other points of attention with the text index creation (stuff that surprised me at first !) ;
- if you use the dbms_pclxutil package, then the ctx_output logging does not work, because the index is created immediately and then populated in the background via dbms_jobs.
- this in combination with the fact that if you are on a RAC, you won't see any activity on the box can be very frightening : this is because oracle can choose to start the workers on the other node.
I understand much better how the text indexing works, I think it is a great technology which can scale via partitioning. But like always the design of the application is crucial, most of our problems come from the fact that we did not choose the right sectioning (we choosed PATH_SECTION_GROUP while XML_SECTION_GROUP is so much better IMO). Maybe later I can convince the dev to change the sectionining, especially because SDATA and MDATA section are not supported with PATCH_SECTION_GROUP (although it seems to work, even though we had one occurence of a bad result linked to the existence of SDATA in the index definition). Also the whole problematic of mixed structured/unstructured searches is completly tackled if one use XML_SECTION_GROUP with MDATA/SDATA (but of course the app was written for Oracle 10...)
Regards, Pierre