Quick and simple frequency analysis

I use this simple query quite often when exploring the data in a table in any Oracle database (from Oracle v8 onwards):

select q.*, 100 * ratio_to_report(c) over () rtr
from (select distinct v, count(*) over (partition by v) c from (
select MYCOLUMN v from MYTABLE
)) q order by c desc;

Just substitute the table name for “MYTABLE” and the column you’re interested in for “MYCOLUMN”. This gives a frequency analysis of values, e.g.:

V         C       RTR
========  ======  =============
INACTIVE  401001  92.9254049544
ACTIVE    30529   7.0745950455

V is the value from the column. C is the count of how many times that value appeared. RTR is the % ratio to the total. The first row indicates the most popular value.

If it’s a very large table and you want quicker results, you can run the analysis over a smaller sample easily, just by adding the SAMPLE keyword:


SPOD for a query

I have two queries that need to be executed by a PL/SQL program. Both of them are quite complex, and both of them have a large section which is identical – because they are different views of the same underlying source data.

One option is to expand out both queries in full, e.g.:

Query 1:

 SELECT <complicated expressions>
       <large complicated query>
      ), <other tables>
 WHERE <complicated predicates>;

Query 2:

 SELECT <different complicated expressions>
       <large complicated query>
      ), <other different tables>
 WHERE <different complicated predicates>;

I don’t like the fact that my <large complicated query> is repeated in full in both cursor definitions. I’d rather have one place where that subquery is defined, because it should remain the same for both queries, since they are supposed to be different views of the same underlying data.

Another option is to create a view on the <large complicated query>, and refer to that in both queries. This is a perfectly acceptable option, and one which I often use. The only downside is if there are any parameters that need to be “fed” to the view. One way is for the view to expose the parameter as a column in the view, and for the calling query to simply query it on that column. This is not always the most efficient method, however, depending on the complexity of the view and how well Oracle can “push down” the predicate into the view at execution time. Another solution to the parameter problem is to use a user-defined context as described here.

The other downside which I don’t like for this case is that the view moves the query away from the package – I’d prefer to have the definitions close together and maintained in one location.

The solution which I used in this case is a pipelined function. For example:

 FUNCTION large_complicated_query
   RETURN source_data_table_type
   rc source_data_type;
   FOR r IN (<large complicated query>) LOOP
     rc.col1 := r.col1;
     rc.col2 := r.col2;
     -- etc.
     PIPE ROW (rc);

Now, the two queries in my package can re-use it like this:

 SELECT <complicated expressions>
 FROM TABLE(my_package.large_complicated_query)
      ,<other tables>
 WHERE <complicated predicates>;

In the package spec I have:

 -- *** dev note: for internal use only ***
 TYPE source_data_type IS
   RECORD (col1 col1_data_type, etc....);
 TYPE source_data_type_table IS TABLE OF source_data_type;
 FUNCTION large_complicated_query
   RETURN source_data_table_type PIPELINED;
 -- *** ******************************* ***

Because the pipelined function is going to be called by SQL (in fact, two queries defined in the same package), its declaration must also be added to the package spec.

In the package body, I use private global variable(s) to hold the parameter for the large complicated query.

When the queries are run, the global variable(s) must first be set to the required parameter. The queries are run, then the global variables are cleared.

The pipelined function is deliberately not useful to other processes – if a developer tried to call it, they’d get no results because they can’t set the parameters (since they are declared as private globals).

A downside to this approach is that the optimizer will not be able to optimize the entire queries “as a whole” – it will execute the entire query in the pipelined function (at least, until the calling queries decide to stop fetching from it). For my case, however, this is not a problem. The entire process runs in less than a second – and this is 10 times faster than it needs to be. In other words, in this case maintainability is more important than performance.

There may be other ways to do this (in fact, I’m quite sure there are), but this way worked for me.

Psychology and Large Tables

Today a project manager asked me about a change to a query being implemented for a search function on our 11gR2 database. His concern was that the new query had a plan that involved several full table scans; whereas the old version used primarily index range scans.

The query used to look something like this:

SELECT score(1) AS score, a.*
      FROM entity_details ed
      JOIN associations a
      ON ed.entity_id = a.entity_id
      LEFT JOIN addresses ad
      ON ed.entity_id = ad.entity_id
     ) a
WHERE CONTAINS(a.entity_name,
  :criterion, 1) > 0
ORDER BY a.entity_name;

Its plan looked like this:

- SORT ORDER BY cost=52 card=13
  - NESTED LOOPS OUTER cost=51 card=13
    - NESTED LOOPS cost=31 card=12
      - TABLE ACCESS BY INDEX ROWID entity_details cost=7 card=12
        - DOMAIN INDEX entity_name_ic cost=4
      - TABLE ACCESS BY INDEX ROWID associations cost=2 card=1
        - INDEX RANGE SCAN asso_entity_i cost=1 card=1
    - TABLE ACCESS BY INDEX ROWID addresses cost=2 card=1
      - INDEX RANGE SCAN address_entity_fk_i cost=1 card=1

The new query had an additional predicate:

SELECT score(1) AS score, a.*
      FROM entity_details ed
      JOIN associations a ON ed.entity_id = a.entity_id
      LEFT JOIN addresses ad ON ed.entity_id = ad.entity_id
     ) a
WHERE CONTAINS(a.entity_name, :criterion, 1) > 0
OR a.entity_name LIKE '%' || UPPER(:criterion) || '%'
ORDER BY a.entity_name;

The query plan for the original query involved index range scans on all the tables; whereas the query plan for the new query involved full table scans on associations and addresses.

SELECT STATEMENT ALL_ROWS cost=348 card=1269
- SORT ORDER BY cost=348 card=1269
  - HASH JOIN OUTER cost=347 card=1269
    - HASH JOIN cost=244 card=1187
      - TABLE ACCESS BY INDEX ROWID entity_details cost=107 card=1187
          - BITMAP OR
              - SORT ORDER BY
                - INDEX RANGE SCAN entity_name_i cost=4
              - SORT ORDER BY
                - DOMAIN INDEX entity_name_ic cost=4
      - TABLE ACCESS FULL association cost=136 card=2351
    - TABLE ACCESS FULL addresses cost=102 card=18560

Initial testing in dev revealed no noticeable performance difference between the two queries, so he was just concerned about the impact of the full table scans on the system.

As you can see, the new plan was still using the domain index, as well as using the ordinary index on entity_name; concatenating the two sets of ROWIDs (BITMAP OR) and then accessing the table as before. Previously, the cardinality estimate for just the CONTAINS predicate was 12 (side note: I’m curious as to how predicates using context indexes are costed – anyone know any details or references?); now, the total cardinality estimate for entity_details is 1187. The cardinality estimate if we just did the LIKE predicate is 1176 (which is simply 5% of the number of rows in the table). The higher cardinality has pushed the rest of the query away from index accesses for association and addresses, towards hash joins and full table scans on those tables.

If I override the cardinality estimate with a lower figure, e.g.

SELECT /*+CARDINALITY(a.ed 50)*/ score(1) AS score, a.* ...

the query plan changes into a typical nested-loops-with-index-access one:

- SORT ORDER BY cost=295 card=53
  - NESTED LOOPS OUTER cost=294 card=53
    - NESTED LOOPS cost=207 card=50
      - TABLE ACCESS BY INDEX ROWID entity_details cost=107 card=50
          - BITMAP OR
              - SORT ORDER BY
                - INDEX RANGE SCAN entity_name_i cost=4
              - SORT ORDER BY
                - DOMAIN INDEX entity_name_ic cost=4
      - TABLE ACCESS BY INDEX ROWID association cost=2 card=1
        - INDEX RANGE SCAN asso_entity_i cost=1 card=1
    - TABLE ACCESS BY INDEX ROWID addresses cost=2 card=1
      - INDEX RANGE SCAN address_entity_fk_i cost=1 card=1

Substituting various cardinalities reveals the “tipping point” for this particular instance (with its particular set of statistics and optimizer parameters) to be around 50.

The cardinality estimates for the full table scans should be the major giveaway: they’re all less than 20,000. In fact, the total number of records in each of these tables does not exceed 25,000, and our dev instance has a full set of data.

I advised this manager to not worry about the new plan. These tables are unlikely to grow beyond 100,000 (they only hold records for about 25,000 associations, entered over the past 5 years) for the expected life of the product, and the entire tables fit in under 500 blocks. With a db_file_multiblock_read_count of 128, it’s likely that the first query of the day will load all the blocks of the tables into the buffer cache with just a dozen or so reads.

If this query were to use index range scans plus table accesses for each search, the performance would only become marginally better (and probably imperceptibly so), at the cost of slower queries on the occasions when users enter poor search criteria. Whereas, with full table scans, even with the worst search criteria, they will typically get their results in less than 5 seconds anyway.

We’re so accustomed to tables like “entities” and “addresses” having millions or tens of millions of rows, so instinctively recoil from full table scans on them; but, in this instance at least, to Oracle, these tables are tiny – for which full table scans are often better.

Infinite Query

This is the query that never ends,
It just goes on and on, my friends.
Some people started fetching not knowing what it was,
And now they can’t stop fetching forever just because…

This is the query that never ends,


CREATE FUNCTION row_generator
RETURN number_table_type
    FOR i IN 1..100 LOOP
      PIPE ROW (i);

SELECT * FROM TABLE(row_generator);

…inspired by…

Question: why can’t the optimizer do better with these?

I’m scratching my head over this one. I thought the cost-based optimizer would be smart enough to eliminate certain predicates, joins and sorts automatically, and pick a cheaper plan accordingly; but in my tests it doesn’t seem to. Can you shed any light on this?

First, the setup for my test case:

select * from v$version;
-- Oracle Database 11g Enterprise Edition
--     Release - 64bit Production
-- PL/SQL Release - Production
-- CORE  Production
-- TNS for Linux: Version - Production
-- NLSRTL Version - Production

create table parent_table
(parent_id number(12)     not null
,padding   varchar2(1000)

create table child_table
(child_id  number(12)     not null
,parent_id number(12)     not null
,padding   varchar2(1000)

-- I find I need at least 100,000 rows for the
-- differences of costs to become significant

insert into parent_table
select rownum, lpad('x',1000,'x')
from dual connect by level <= 100000;

insert into child_table
select rownum, rownum, lpad('x',1000,'x')
from dual connect by level <= 100000;
alter table parent_table add (
  constraint parent_pk primary key (parent_id) );
alter table child_table add (
  constraint child_pk primary key (child_id)
 ,constraint fk foreign key (parent_id)
  references parent_table (parent_id) );
create index child_table_parent_i on child_table (parent_id);
begin dbms_stats.gather_table_stats(ownname => USER
, tabname => 'PARENT_TABLE'
, estimate_percent => 100
, method_opt => 'for all columns size auto'
, cascade => TRUE); end;
begin dbms_stats.gather_table_stats(ownname => USER
, tabname => 'CHILD_TABLE'
, estimate_percent => 100
, method_opt => 'for all columns size auto'
, cascade => TRUE); end;

Case 1

explain plan for
select count(*)
from child_table;

-- index fast full scan (unique) on child_pk - cost 58 - great

explain plan for
select count(*)
from child_table
where parent_id is not null;

-- index fast full scan on child_table_parent_i - cost 62

Q.1: Why can’t this query with the NOT NULL predicate on a NOT NULL column eliminate the predicate – and use the pk index instead?

Case 2

explain plan for
select count(*)
from parent_table inner join child_table
on (parent_table.parent_id = child_table.parent_id);

-- index fast full scan (unique) on child_pk - cost 58 - great

explain plan for
select count(*)
from parent_table inner join child_table
on (parent_table.parent_id = child_table.parent_id)
order by parent_table.padding;

-- hash join
--    index fast full scan on child_table_parent_i
--    full table scan parent_table
-- - cost 9080

Q.2: The first query eliminated the join; why can’t it eliminate the ORDER BY as well, since it’s irrelevant to the results?

Handling unique constraint violations by Hibernate

A particular table in our system is a M:M link table between Bonds and Payments, imaginatively named BOND_PAYMENTS; and to make the Java devs’ jobs easier it has a surrogate key, BOND_PAYMENT_ID. Its structure, therefore, is basically:


This is a very simple design, quite common in relational database designs. There is a Primary key constraint on BOND_PAYMENT_ID, and we’ve also added a Unique constraint on (BOND_NUMBER, PAYMENT_ID) since it makes no sense to have more than one link between a Bond and a Payment.

The application allows a user to view all the Payments linked to a particular Bond; and it allows them to create new links, and delete existing links. Once they’ve made all their desired changes on the page, they hit “Save”, and Hibernate does its magic to run the required SQL on the database. Unfortunately, this was failing with ORA-00001: unique constraint violated.

Now, the way this page works is that it compares the old set of payments for the bond with the new target set, and Hibernate works out which records need to be deleted, which need to be inserted, and leaves the rest untouched. Unfortunately, in its infinite wisdom it does the INSERTs first, then it does the DELETEs. Apparently this order can’t be changed.

This is the cause of the unique constraint violation – if the user deletes a link to a payment, then changes their mind and re-inserts a link to the same payment, Hibernate quite happily tries to insert it then delete it. Since these inserts/deletes are running as separate SQL statements, Oracle validates the constraint immediately on the first insert.

We had only a few options:

  1. Make the constraint deferrable
  2. Remove the unique constraint

Option 2 was not very palatable, because the constraint provides excellent protection from nasty application bugs that might allow inconsistent data to be saved. We went with option 1.

ALTER TABLE bond_payments ADD
  CONSTRAINT bond_payment_uk UNIQUE (bond_number, payment_id)

This solved it – and no changes required to the application. If a bug in the application were to cause it to try to insert a duplicate row, it will fail with ORA-02091 (transaction rolled back) and ORA-00001 (unique constraint violated) when the session COMMITs.

The only downside is that the index created to police this constraint is now a non-unique index, so may be somewhat less efficient for queries. We decided this is not as great a detriment for this particular case.

If you know of any other options that we should have considered, let me know 🙂

Data dictionary quiz question

This is a totally unfair quiz question (for anyone who isn’t intimately acquainted with the Oracle data dictionary views). There, I’ve warned you. I would have got this one wrong, myself.

Which of the following queries (if any) will run without error?

SELECT * FROM dba_tab_privs WHERE owner = 'SCOTT';
SELECT * FROM all_tab_privs WHERE owner = 'SCOTT';
SELECT * FROM user_tab_privs WHERE owner = 'SCOTT';

Now, don’t go and try running these in your database until after you’ve written down what you think the answers are. That would be cheating 🙂





Ok, now to break the suspense for all the other readers (both of you), who couldn’t be bothered testing it out for yourself.

Statement 1. The query on DBA_TAB_PRIVS will succeed – assuming you have the necessary privileges on the DBA* views. This view does include the column OWNER.

Statement 3. The query on USER_TAB_PRIVS, unlike what I and several others might assume,  will succeed.  Many data dictionary views, such as USER_TABLES, omit the OWNER column – which makes sense, since it is expected that it will be simply the currently-logged-in user. But for user_tab_privs, this column is provided, for good reason – because the table you have a privilege on may very well be owned by another schema.

Statement 2. The query on ALL_TAB_PRIVS, in order to be consistent with DBA_TAB_PRIVS and USER_TAB_PRIVS, should have the OWNER column, by rights. But, just to make things interesting, the column is called TABLE_SCHEMA instead. So, my query would fail.

What’s this table? SYS_IOT_OVER_152769

I was writing some scripts to drop all the objects in a particular schema on an 11gR2 database, and in querying the USER_OBJECTS view, came across a whole lot of tables with names like this:


What in the world are these? As it turns out, these are overflow tables for Index Organized Tables.

The following query on USER_TABLES (or ALL_TABLES or DBA_TABLES) will reveal all:

SQL> SELECT table_name, iot_type, iot_name FROM USER_TABLES
     WHERE iot_type IS NOT NULL;
===================  ============  ==========================

The IOT_NAME reveals the table that owns the overflow table. The create command for TBMS_REF_ACCOUNT_TYPE was:

  COMMENTS               VARCHAR2(4000),
  VERSION_ID             NUMBER(12)        DEFAULT 1       NOT NULL,

This means that ACCOUNT_TYPE_CODE and DESCRIPTION will be kept in the index, since these are pretty much the only columns normally accessed; the rest, including the big comments field (which seems to be largely unused), will be stored in the overflow table if they are set.

Right. So I drop the REF table – that should take care of the overflow table, right? Wrong. The SYS_IOT_OVER table is still there! Ah – that’s because the REF table is sitting in the recyclebin. Purge it, and now the SYS_IOT_OVER table is gone. (Not that there was anything wrong with it, mind you – I just wanted to clean this schema out so I could recreate it.)

For more info: http://download.oracle.com/docs/cd/E11882_01/server.112/e16508/indexiot.htm#CNCPT911

A simple data ETL method – nothin’ but SQL

My client has decided to design and build a completely new replacement system for an aging system running on Oracle Forms 6i on Oracle 8. The new system will have a web frontend, backed by Hibernate (don’t get me started) on top of an Oracle 11gR1 database. Crucially, due to changes to business practices and legislation, the new system has been designed “from scratch”, including a new data model.

My task is to write the ETL scripts which will take the data from the legacy database (an Oracle 8i schema), transform it to meet the requirements of the new model, and load it. If you’re looking at building scripts to transform data from one system to another, the method I used might be helpful for you too.

Making it more complicated is their desire that the data move be executed in two stages – (1) before the switch-over, transform and load all “historical” data; (2) at go-live, transform and load all “current” data, as well as any modifications to “historical” data.

Since the fundamental business being supported by this system hasn’t changed very much, the old and new data models have a lot in common – the differences between them are not very complex. In addition, the data volume is not that great (coming from someone who’s worked with terabyte-scale schemas) – the biggest table only had 2 million rows. For these reasons, the purchase of any specialised ETL tools was not considered. Instead, they asked me to write the ETL as scripts that can just be run on the database.

These scripts must be re-runnable: they should be able to be run without modification to pick up any changes in the legacy data, and automatically work out how to merge the changes into the new schema.

The first step for me was to analyse both data models, and work out a mapping for the data between them. The team for the project had a very good idea of what the tables in the new model meant (since they had designed it), but there was no-one available to me to explain how the old data model worked. It was down to me to learn how the legacy data model worked – by exploring it at the database level, examining the source for the forms and reports, and in some cases by talking to the users.

The outcome of this analysis was two spreadsheets: one was a list of every table and column in the legacy database, and the other was a list of every table and column in the new database. For each table in the legacy database, I recorded which table (or tables) the data would be migrated to in the new schema, or an explanation if the data could be safely disregarded. For each table in the new schema, I recorded which table (or tables) in the legacy database would feed into it. In the end, eleven of the tables in the new schema would be loaded.

Then, for each table in the legacy and new schemas, I worked through each column, identifying what it meant, and how it would be mapped from the old to the new. In some cases, the mapping was very 1:1 – perhaps some column names were different, or code values different, but relatively simple. In other cases, the mapping would require a more complex transformation, prehaps based on multiple tables. For example, both systems had a table named “ADDRESS” which stored street or postal addresses; in the old system, this table was a child table to the “PARTY” table; so PARTY was 1:M to ADDRESS. In the new model, however, there was a master “ADDRESS” table which was intended to store any particular address once and only once; the relationship of PARTY to ADDRESS is M:M. De-duplication of addresses hasn’t come up yet but it’s going to be fun when it does 🙂

Thankfully, in no cases was the mapping so complicated that I couldn’t envisage how it could be done using relatively simple SQL.

Once the spreadsheets were filled, I was finally able to start coding!

In order to meet the requirements, my scripts must:

  1. INSERT rows in the new tables based on any data in the source that hasn’t already been created in the destination
  2. UPDATE rows in the new tables based on any data in the source that has already been inserted in the destination
  3. DELETE rows in the new tables where the source data has been deleted

Now, instead of writing a whole lot of INSERT, UPDATE and DELETE statements, I thought “surely MERGE would be both faster and better” – and in fact, that has turned out to be the case. By writing all the transformations as MERGE statements, I’ve satisfied all the criteria, while also making my code very easily modified, updated, fixed and rerun. If I discover a bug or a change in requirements, I simply change the way the column is transformed in the MERGE statement, and re-run the statement. It then takes care of working out whether to insert, update or delete each row.

My next step was to design the architecture for my custom ETL solution. I went to the dba with the following design, which was approved and created for me:

  1. create two new schemas on the new 11g database: LEGACY and MIGRATE
  2. take a snapshot of all data in the legacy database, and load it as tables in the LEGACY schema
  3. grant read-only on all tables in LEGACY to MIGRATE
  4. grant CRUD on all tables in the target schema to MIGRATE.

All my scripts will run as the MIGRATE user. They will read the data from the LEGACY schema (without modifying) and load it into intermediary tables in the MIGRATE schema. Each intermediary table takes the structure of a target table, but adds additional columns based on the legacy data. This means that I can always map from legacy data to new data, and vice versa.

For example, in the legacy database we have a table:

 par_id             NUMBER        PRIMARY KEY,
 par_domain         VARCHAR2(10)  NOT NULL,
 par_first_name     VARCHAR2(100) ,
 par_last_name      VARCHAR2(100),
 par_dob            DATE,
 par_business_name  VARCHAR2(250),
 created_by         VARCHAR2(30)  NOT NULL,
 creation_date      DATE          NOT NULL,
 last_updated_by    VARCHAR2(30),
 last_update_date   DATE)

In the new model, we have a new table that represents the same kind of information:

 party_id           NUMBER(9)     PRIMARY KEY,
 party_type_code    VARCHAR2(10)  NOT NULL,
 first_name         VARCHAR2(50),
 surname            VARCHAR2(100),
 date_of_birth      DATE,
 business_name      VARCHAR2(300),
 db_created_by      VARCHAR2(50)  NOT NULL,
 db_created_on      DATE          DEFAULT SYSDATE NOT NULL,
 db_modified_by     VARCHAR2(50),
 db_modified_on     DATE,
 version_id         NUMBER(12)    DEFAULT 1 NOT NULL)

This was the simplest transformation you could possibly think of – the mapping from one to the other is 1:1, and the columns almost mean the same thing.

The solution scripts start by creating an intermediary table:

 old_par_id         NUMBER        PRIMARY KEY,
 party_id           NUMBER(9)     NOT NULL,
 party_type_code    VARCHAR2(10)  NOT NULL,
 first_name         VARCHAR2(50),
 surname            VARCHAR2(100),
 date_of_birth      DATE,
 business_name      VARCHAR2(300),
 db_created_by      VARCHAR2(50),
 db_created_on      DATE,
 db_modified_by     VARCHAR2(50),
 db_modified_on     DATE,
 deleted            CHAR(1))

You’ll notice that the intermediary table has the same columns of the new table (except for VERSION_ID, which will just be 1), along with the minimum necessary to link each row back to the source data – the primary key from the source table, PAR_ID.

You might also notice that there is no unique constraint on PARTY_ID – this is because we needed to do some merging and de-duplication on the party info. I won’t go into that here, but the outcome is that for a single PARTY_ID might be mapped from more than one OLD_PAR_ID.

The second step is the E and T parts of “ETL”: I query the legacy table, transform the data right there in the query, and insert it into the intermediary table. However, since I want to be able to re-run this script as often as I want, I wrote this as a MERGE statement:

  SELECT par_id            AS old_par_id,
         par_id            AS party_id,
         CASE par_domain
           WHEN 'P' THEN 'PE' /*Person*/
           WHEN 'O' THEN 'BU' /*Business*/
         END               AS party_type_code,
         par_first_name    AS first_name,
         par_last_name     AS surname,
         par_dob           AS date_of_birth,
         par_business_name AS business_name,
         created_by        AS db_created_by,
         creation_date     AS db_created_on,
         last_updated_by   AS db_modified_by,
         last_update_date  AS db_modified_on
    SELECT null
    WHERE  d.old_par_id = s.par_id
    AND    (d.db_modified_on = s.last_update_date
            OR (d.db_modified_on IS NULL
               AND s.last_update_date IS NULL))
  ) src
ON (src.OLD_PAR_ID = dest.OLD_PAR_ID)
  party_id        = src.party_id        ,
  party_type_code = src.party_type_code ,
  first_name      = src.first_name      ,
  surname         = src.surname         ,
  date_of_birth   = src.date_of_birth   ,
  business_name   = src.business_name   ,
  db_created_by   = src.db_created_by   ,
  db_created_on   = src.db_created_on   ,
  db_modified_by  = src.db_modified_by  ,
  db_modified_on  = src.db_modified_on
  src.old_par_id      ,
  src.party_id        ,
  src.party_type_code ,
  src.first_name      ,
  src.surname         ,
  src.date_of_birth   ,
  src.business_name   ,
  src.db_created_by   ,
  src.db_created_on   ,
  src.db_modified_by  ,
  src.db_modified_on  ,
  NULL                );

You’ll notice that all the transformation logic happens right there in a single SELECT statement. This is an important part of how this system works – every transformation is defined in one place and one place only. If I need to change the logic for any column, all I have to do is update it in one place, and re-run the MERGE.

This is a simple example; for some of the tables, the SELECT statement is quite complex.

(Warning: you’ll note that I’ve omitted the column list from the INSERT clause; this can be dangerous if you’re not in complete control of the column order like I am for this particular table)

There is a follow-up UPDATE statement that for a couple of thousand records, changes the PARTY_ID to a different value; in effect, this performs the de-duplication.

Next, we look for any rows that have been deleted:

SET    deleted = 'Y'
WHERE  deleted IS NULL
  SELECT null
  WHERE  src.par_id = dest.old_par_id);

The idea is that the data in the MIGRATE table is *exactly* what we will insert, unmodified, into the target schema. In a year’s time, we could go back to this MIGRATE schema and see what we actually inserted when the system went live. In addition, we’ll be able to go back to the LEGACY schema and see exactly how the data looked in the old system; and we’ll be able to use tables like MIGRATE.TBMS_PARTY to map back-and-forth between the old and new systems.

The final stage of the process is the “L” of “ETL”. This, again, uses a MERGE statement:

  WHERE  s.party_id = s.old_par_id /*i.e. not a duplicate*/
  AND    (s.deleted IS NOT NULL
          OR NOT EXISTS (
            SELECT null
            FROM   NEW.TBMS_PARTY d
            WHERE  d.party_id = s.party_id
            AND    (d.db_modified_on = s.db_modified_on
                    OR (d.db_modified_on IS NULL
                        AND s.db_modified_on IS NULL))
           ) )  
  ) src
ON (src.party_id = dest.party_id)
  party_type_code = src.party_type_code ,
  first_name      = src.first_name      ,
  surname         = src.surname         ,
  date_of_birth   = src.date_of_birth   ,
  business_name   = src.business_name   ,
  db_created_by   = src.db_created_by   ,
  db_created_on   = src.db_created_on   ,
  db_modified_by  = src.db_modified_by  ,
  db_modified_on  = src.db_modified_on
  party_id        ,
  party_type_code ,
  first_name      ,
  surname         ,
  date_of_birth   ,
  business_name   ,
  db_created_by   ,
  db_created_on   ,
  db_modified_by  ,
  db_modified_on  )
  src.party_type_code ,
  src.first_name      ,
  src.surname         ,
  src.date_of_birth   ,
  src.business_name   ,
  src.db_created_by   ,
  src.db_created_on   ,
  src.db_modified_by  ,
  src.db_modified_on  )

A few things to note here:

  • The SELECT clause brings back each row from the intermediary table that has not been merged to a new record (by the way, those records are needed because they are used when transforming PAR_ID values in child tables) or that has not been modified since it was previously loaded.
  • The MERGE inserts any new rows, updates all columns for modified rows, and deletes rows that have been marked for deletion.
  • NO transformation of data happens here.
  • If any data fails any validation, the MERGE logs the error and continues, using a table created using this:
  err_log_table_name  => 'ERR$_TBMS_PARTY',
  err_log_table_owner => 'MIGRATE');

I can then query this error table to see if there were any problems, e.g.:


A common issue is a failed check constraint, e.g. where the old system failed to validate something correctly. We’d then go back and either change the transformation to work around the problem, or send the data back to the business and ask them to fix it in the source.

Each stage of this ETL solution can be restarted and re-run. In fact, that’s what we will be doing; a few weeks prior to go-live, we’ll get a preliminary extract of the old system into the LEGACY schema, and run all my scripts. Then, at go-live, when the old system is taken down, we’ll wipe the LEGACY schema and refresh it from Prod. We will then re-run the scripts to take changes through.

All the scripts for each table had the same structure: one script to create the intermediary table; one script to do the merge into the intermediary table; and one script to merge into the final destination. With the exception of the SELECT statement in the first merge script, which differed greatly for each table, these scripts were very similar, so I started by generating them all. For this I used queries on the data dictionary to generate all the SELECT lists and x = y lists, and after a bit of work I had a complete set of ETL scripts which just needed me to go in and make up the SELECT statement for the transformation.

For this case, a relatively simple data migration problem, this method seems to have worked well. It, or a variation on it, might very well work for you too.