Problem with BULK COLLECT with million rows - Oracle 9.0.1.4

Question

We have a requirement where are supposed to load 58 millions of rows into a FACT Table in our DATA WAREHOUSE. We initially planned to use Oracle Warehouse Builder but due to performance reasons, decided to write custom code. We wrote a custome procedure which opens a simple cursor and reads all the 58 million rows from the SOURCE Table and in a loop processes the rows and inserts the records into a TARGET Table. The logic works fine but it took 20hrs to complete the load.
We then tried to leverage the BULK COLLECT and FORALL and PARALLEL options and modified our PL/SQL code completely to reflect these. Our code looks very simple.
1. We declared PL/SQL BINARY_INDEXed Tables to store the data in memory.
2. We used BULK COLLECT into FETCH the data.
3. We used FORALL statement while inserting the data.
We did not introduce any of our transformation logic yet.
We tried with the 600,000 records first and it completed in 1 min and 29 sec with no problems. We then doubled the no. of rows to 1.2 million and the program crashed with the following error:
ERROR at line 1:
ORA-04030: out of process memory when trying to allocate 16408 bytes (koh-kghu
call ,pmucalm coll)
ORA-06512: at "VVA.BULKLOAD", line 66
ORA-06512: at line 1
We got the same error even with 1 million rows.
We do have the following configuration:
SGA - 8.2 GB
PGA
- Aggregate Target - 3GB
- Current Allocated - 439444KB (439 MB)
- Maximum allocated - 2695753 KB (2.6 GB)
Temp Table Space - 60.9 GB (Total)
- 20 GB (Available approximately)
I think we do have more than enough memory to process the 1 million rows!!
Also, some times the same program results in the following error:
SQL> exec bulkload
BEGIN bulkload; END;
ERROR at line 1:
ORA-03113: end-of-file on communication channel
We did not even attempt the full load. Also, we are not using the PARALLEL option yet.
Are we hitting any bug here? Or PL/SQL is not capable of mass loads? I would appreciate any thoughts on this?
Thanks,
Haranadh
Following is the code:
set echo off
set timing on
create or replace procedure bulkload as
-- SOURCE --
TYPE src_cpd_dt IS TABLE OF ima_ama_acct.cpd_dt%TYPE;
TYPE src_acqr_ctry_cd IS TABLE OF ima_ama_acct.acqr_ctry_cd%TYPE;
TYPE src_acqr_pcr_ctry_cd IS TABLE OF ima_ama_acct.acqr_pcr_ctry_cd%TYPE;
TYPE src_issr_bin IS TABLE OF ima_ama_acct.issr_bin%TYPE;
TYPE src_mrch_locn_ref_id IS TABLE OF ima_ama_acct.mrch_locn_ref_id%TYPE;
TYPE src_ntwrk_id IS TABLE OF ima_ama_acct.ntwrk_id%TYPE;
TYPE src_stip_advc_cd IS TABLE OF ima_ama_acct.stip_advc_cd%TYPE;
TYPE src_authn_resp_cd IS TABLE OF ima_ama_acct.authn_resp_cd%TYPE;
TYPE src_authn_actvy_cd IS TABLE OF ima_ama_acct.authn_actvy_cd%TYPE;
TYPE src_resp_tm_id IS TABLE OF ima_ama_acct.resp_tm_id%TYPE;
TYPE src_mrch_ref_id IS TABLE OF ima_ama_acct.mrch_ref_id%TYPE;
TYPE src_issr_pcr IS TABLE OF ima_ama_acct.issr_pcr%TYPE;
TYPE src_issr_ctry_cd IS TABLE OF ima_ama_acct.issr_ctry_cd%TYPE;
TYPE src_acct_num IS TABLE OF ima_ama_acct.acct_num%TYPE;
TYPE src_tran_cnt IS TABLE OF ima_ama_acct.tran_cnt%TYPE;
TYPE src_usd_tran_amt IS TABLE OF ima_ama_acct.usd_tran_amt%TYPE;
src_cpd_dt_array src_cpd_dt;
src_acqr_ctry_cd_array      src_acqr_ctry_cd;
src_acqr_pcr_ctry_cd_array     src_acqr_pcr_ctry_cd;
src_issr_bin_array      src_issr_bin;
src_mrch_locn_ref_id_array     src_mrch_locn_ref_id;
src_ntwrk_id_array      src_ntwrk_id;
src_stip_advc_cd_array      src_stip_advc_cd;
src_authn_resp_cd_array      src_authn_resp_cd;
src_authn_actvy_cd_array      src_authn_actvy_cd;
src_resp_tm_id_array      src_resp_tm_id;
src_mrch_ref_id_array      src_mrch_ref_id;
src_issr_pcr_array      src_issr_pcr;
src_issr_ctry_cd_array      src_issr_ctry_cd;
src_acct_num_array      src_acct_num;
src_tran_cnt_array      src_tran_cnt;
src_usd_tran_amt_array      src_usd_tran_amt;
j number := 1;
CURSOR c1 IS
SELECT
cpd_dt,
acqr_ctry_cd ,
acqr_pcr_ctry_cd,
issr_bin,
mrch_locn_ref_id,
ntwrk_id,
stip_advc_cd,
authn_resp_cd,
authn_actvy_cd,
resp_tm_id,
mrch_ref_id,
issr_pcr,
issr_ctry_cd,
acct_num,
tran_cnt,
usd_tran_amt
FROM ima_ama_acct ima_ama_acct
ORDER BY issr_bin;
BEGIN
OPEN c1;
FETCH c1 bulk collect into
src_cpd_dt_array ,
src_acqr_ctry_cd_array ,
src_acqr_pcr_ctry_cd_array,
src_issr_bin_array ,
src_mrch_locn_ref_id_array,
src_ntwrk_id_array ,
src_stip_advc_cd_array ,
src_authn_resp_cd_array ,
src_authn_actvy_cd_array ,
src_resp_tm_id_array ,
src_mrch_ref_id_array ,
src_issr_pcr_array ,
src_issr_ctry_cd_array ,
src_acct_num_array ,
src_tran_cnt_array ,
src_usd_tran_amt_array ;
CLOSE C1;
FORALL j in 1 .. src_cpd_dt_array.count
INSERT INTO ima_dly_acct (
     CPD_DT,
     ACQR_CTRY_CD,
     ACQR_TIER_CD,
     ACQR_PCR_CTRY_CD,
     ACQR_PCR_TIER_CD,
     ISSR_BIN,
     OWNR_BUS_ID,
     USER_BUS_ID,
     MRCH_LOCN_REF_ID,
     NTWRK_ID,
     STIP_ADVC_CD,
     AUTHN_RESP_CD,
     AUTHN_ACTVY_CD,
     RESP_TM_ID,
     PROD_REF_ID,
     MRCH_REF_ID,
     ISSR_PCR,
     ISSR_CTRY_CD,
     ACCT_NUM,
     TRAN_CNT,
     USD_TRAN_AMT)
     VALUES (
     src_cpd_dt_array(j),
     src_acqr_ctry_cd_array(j),
     null,
     src_acqr_pcr_ctry_cd_array(j),
          null,
          src_issr_bin_array(j),
          null,
          null,
          src_mrch_locn_ref_id_array(j),
          src_ntwrk_id_array(j),
          src_stip_advc_cd_array(j),
          src_authn_resp_cd_array(j),
          src_authn_actvy_cd_array(j),
          src_resp_tm_id_array(j),
          null,
          src_mrch_ref_id_array(j),
          src_issr_pcr_array(j),
          src_issr_ctry_cd_array(j),
          src_acct_num_array(j),
          src_tran_cnt_array(j),
          src_usd_tran_amt_array(j));
COMMIT;
END bulkload;
SHOW ERRORS
-----------------------------------------------------------------------------

Answer 1

do you have a unique key available in the rows you are fetching?
It seems a cursor with 20 million rows that is as wide as all the columnsyou want to work with is a lot of memory for the server to use at once. You may be able to do this with parallel processing (dop over 8) and a lot of memory for the warehouse box (and the box you are extracting data from)...but is this the most efficient (and thereby fastest) way to do it?
What if you used a cursor to select a unique key only, and then during the cursor loop fetch each record, transform it, and insert it into the target?
Its a different way to do a lot at once, but it cuts down on the overall memory overhead for the process.
I know this isnt as elegant as a single insert to do it all at once, but sometimes trimming a process down so it takes less resources at any given moment is much faster than trying to do the whole thing at once.
My solution is probably biased by transaction systems, so I would be interested in what the data warehouse community thinks of this.
For example:
source table my_transactions (tx_seq_id number, tx_fact1 varchar2(10), tx_fact2 varchar2(20), tx_fact3 number, ...)
select a cursor of tx_seq_id only (even at 20 million rows this is not much)
you could then either use a for loop or even bulk collect into a plsql collection or table
then process individually like this:
procedure process_a_tx(p_tx_seq_id in number)
is
rTX my_transactions%rowtype;
begin
select * into rTX from my_transactions where tx_seq_id = p_tx_seq_id;
--modify values as needed
insert into my_target(a, b, c) values (rtx.fact_1, rtx.fact2, rtx.fact3);
commit;
exception
when others
rollback;
--write to a log or raise and exception
end process_a_tx;
procedure collect_tx
is
cursor tx is
select tx_seq_id from my_transactions;
begin
for rTx in cTx loop
process_a_tx(rtx.tx_seq_id);
end loop;
end collect_tx;