How to do monthly refresh of large DB tables without interrupting user access to them

Question

Hi all, I have four DB tables in an Oracle database that need to be rewritten/refreshed every week or every month. I am writing this script in PHP using the standard OCI functions, that will read new data in from XML and refresh these four tables. The four tables have the following properties

TABLE A - up to 2mil rows, one primary key (One row might take max 2K data)

TABLE B - up to 10mil rows, one foreign key pointing to TABLE A (One row might take max 1100 bytes of data)

TABLE C - up to 10mil rows, one foreign key pointing to TABLE A (One row might take max 1100 bytes of data)

TABLE D - up to 10mil rows, one foreign key pointing to TABLE A (One row might take max 120 bytes of data)

So I need to repopulate these tables without damaging the user experience. I obviously can't delete the tables and just repopulate them as it is a somewhat lengthy process.

I've considered just a big transaction where I DELETE FROM all of the tables and just regenerate them. I get a little concerned about the length of the transaction (don't know yet but it could take an hour or so).

I wanted to create temp table replicas of all of the tables and populate those instead. Then I could DROP the main tables and rename the temp tables. However you can't do the DROP and ALTER table statements within a transaction as they always do an auto commit. This should be able to be done quickly (four DROP and and four ALTER TABLE statements), but it can't guarantee that a user won't get an error within that short period of time.

Now, a combination of the two ideas, I'm considering doing the temp tables, then doing a DELETE FROM on all four original tables and then and INSERT INTO from the temp tables to repopulate the main tables. Since there are no DDL statements here, this would all work within a transaction. Then, however, I wondering if the memory it takes to process some 60 million records within a transaction is going to get me in trouble (this would be a concern for the first idea as well).

I would think this would be a common scenario. Is there a standard or recommended way of doing this? Any tips would be appreciated. Thanks.

Answer 1

Why not add a version column? That way you can add the new rows with a different version number. Create a view against the table that specifies the current version. After the new rows are added recompile the view with the new version number. When that's done, go back and delete the old rows.

Answer 2

"I'm a little turned off by the idea of having to add a where version=1 clause to all of my select statements."

(a) Worse, you probably want to make that 'where version = select version from CURRENTVERSION' or some such.

(b) But you don't have to add that condition in every select you got. Make your tables "private" with names such as STORED_TABLEA, and views such as (roughly) "CREATE VIEW TABLE_A AS "SELECT * FROM STORED_TABLEA WHERE VERSION = CURRENTVERSION"". Maybe you should have to mind about what happens to someone who is cursoring the view right when the CURRENTVERSION gets updated.

And of course there is always the real right solution to such problems : use the database for its intended purpose (which is to store permanent data), and ditch the XML.

Answer 3

In Oracle your can partition your tables and indexes based on a Date or time column that way to remove a lot of data you can simply drop the partition instead of performing a delete command.

We used to use this to manage monthly archives of 100 Million+ records and not have downtime.

http://www.oracle.com/technology/oramag/oracle/06-sep/o56partition.html is a super handy page for learning about partitioning.

Answer 4

You could have a synonym for each of your big tables. Create new incarnations of your tables, populate them, drop and recreate the synonyms, and finally drop your old tables. This has the advantage of (1) only one actual set of DML (the inserts) avoiding redo generation for your deletes and (2) the synonym drop/recreate is very fast, minimizing the potential for a "bad user experience".

Reminds me of a minor peeve of mine about Oracle's synonyms: why isn't there an ALTER SYNONYM command?

Answer 5

I'm assuming your users don't actually modify the data in these tables since it is deleted from another source every week, so it doesn't really matter if you lock the tables for a full hour. The users can still query the data, you just have to size you rollback segment appropriately. A simple DELETE+INSERT therefore should work fine.

Now if you want to speed things up AND if the new data has little difference with the previous data you could load the new data into temporary tables and updating the tables with the delta with a combination of MERGE+DELETE like this:

Setup:

CREATE TABLE a (ID NUMBER PRIMARY KEY, a_data CHAR(200));
CREATE GLOBAL TEMPORARY TABLE temp_a (
   ID NUMBER PRIMARY KEY, a_data CHAR(200)
) ON COMMIT PRESERVE ROWS;
-- Load A
INSERT INTO a 
   (SELECT ROWNUM, to_char(ROWNUM) FROM dual CONNECT BY LEVEL <= 10000);
-- Load TEMP_A with extra rows
INSERT INTO temp_a 
   (SELECT ROWNUM + 100, to_char(ROWNUM + 100) 
      FROM dual 
   CONNECT BY LEVEL <= 10000);
UPDATE temp_a SET a_data = 'x' WHERE mod(ID, 1000) = 0;

This MERGE statement will insert the new rows and update the old rows only if they are different:

SQL> MERGE INTO a
  2  USING (SELECT temp_a.id, temp_a.a_data
  3           FROM temp_a
  4           LEFT JOIN a ON (temp_a.id = a.id)
  5          WHERE decode(a.a_data, temp_a.a_data, 1) IS NULL) temp_a
  6  ON (a.id = temp_a.id)
  7  WHEN MATCHED THEN
  8     UPDATE SET a.a_data = temp_a.a_data
  9  WHEN NOT MATCHED THEN
 10     INSERT (id, a_data) VALUES (temp_a.id, temp_a.a_data);

Done

You will then need to delete the rows that aren't in the new set of data:

SQL> DELETE FROM a WHERE a.id NOT IN (SELECT temp_a.id FROM temp_a);

100 rows deleted

You would insert into A then into the child tables and deleting in reverse order.

Answer 6

What we do in some cases is have two versions of the tables, say SalesTargets1 and SalesTargets2 (an active and inactive one.) Truncate the records from the inactive one and populate it. Since no one but you uses the inactive one, there should be no locking issues or impact on the users while it is populating. Then have view that selcts all the information from the active table (it should be named what the current table is now, say SalesTargets in my example). Then to switch to the refreshed data, all you have to do is run an alter view statement.

Answer 7

I assume that this refreshing activity is the only way of data changing in these tables, so that you don't need to worry about inconsistencies due to other writing processes during the load.

All that deleting and inserting will be costly in terms of undo usage; you also would exclude the option of using faster data loading techniques. For example, your inserts will go much, much faster if you insert into the tables with no indexes, then apply the indexes after the load is done. There are other strategies as well, but both of them preclude the "do it all in one transaction" technique.

Your second choice would be my choice - build the new tables, then rename the old ones to a dummy name, rename the temps to the new name, then drop the old tables. Since the renames are fast, you'd have a less than one second window when the tables were unavailable, and you'd then be free to drop the old tables at your leisure.

If that one second window is unacceptable, one method I've used in situations like this is to use an additional locking object - specifically, a table with a single row that users would be required to select from before they access the real tables, and that your load process could lock in exclusive mode before it it does the rename operation.

Your PHP script would use two connections to the db - one where you do the lock, the other where you do the loading, renaming and dropping. This way the implicit commits in the work connection won't terminate the lock in the other table.

So, in the script, you'd do something like:

Connection 1: Create temp tables, load them, create new indexes

Connection 2:

LOCK TABLE Load_Locker IN SHARE ROW EXCLUSIVE MODE;

Connection 1: Perform renaming swap of old & new tables

Connection 2: Rollback;

Connection 1: Drop old tables.

Meanwhile, your clients would issue the following command immediately after starting a transaction (or a series of selects):

LOCK TABLE Load_Locker IN SHARE MODE;

You can have as many clients locking the table this way - your process above will block behind them until they have all released the lock, at which point subsequent clients will block until you perform your operations. Since the only thing you're doing inside the context of the SHARE ROW EXCLUSIVE lock is renaming tables, your clients would only ever block for an instant. Additionally, putting this level of granularity allows you to control how long the clients would have a read consistent view of the old table; without it, if you had a client that did a series of reads that took some time, you might end up changing the tables mid-stream and wind up with weird results if the early queries pulled old data & the later queries pulled new data. Using SET TRANSACTION SET ISOLATION LEVEL READ ONLY would be another way of addressing this issue if you weren't using my approach.

The only real downside to this approach is that if your client read transactions take some time, you run the risk of other clients being blocked for longer than an instant, since any locks in SHARE MODE that occur after your load process issues its SHARE ROW EXCLUSIVE lock will block until the load process finishes its task. For example:

10:00 user 1 issues SHARE lock
10:01 user 2 issues SHARE lock
10:03 load process issues SHARE ROW EXCLUSIVE lock (and is blocked)
10:04 user 3 issues SHARE lock (and is blocked by load's lock)
10:10 user 1 releases SHARE
10:11 user 2 releases SHARE (and unblocks loader)
10:11 loader renames tables & releases SHARE ROW EXCLUSIVE (and releases user 3)
10:11 user 3 commences queries, after being blocked for 7 minutes

However, this is really pretty kludgy. Kinlan's solution of partitioning is most likely the way to go. Add an extra column to your source tables that contains a version number, partition your data based on that version, then create views that look like your current tables that only show data that shows the current version (determined by the value of a row in a "CurrentVersion" table). Then just do your load into the table, update your CurrentVersion table, and drop the partition for the old data.

Answer 8

Have you evaluated the size of the delta (of changes).

If the number of rows that get updated (as opposed to inserted) every time you put up a new rowset it not too high, then I think you should consider importing the new set of data into a set of staging tables and do an update-where-exists and insert-where-not-exists (UPSERT) solution and just refresh your indexes (ok ok indices).

Treat it like ETL.

Answer 9

Am I the only one (except Vincent) who would first test the simplest possible solution, i.e. DELETE/INSERT, before trying to build something more advanced?

Then, however, I wondering if the memory it takes to process some 60 million records within a transaction is going to get me in trouble (this would be a concern for the first idea as well).

Oracle manages memory quite well, it hasn't been written by a bunch of Java novices (oops it just came out of my mouth!). So the real question is, do you have to worry about the performance penalties of thrashing REDO and UNDO log files... In other words, build a performance test case and run it on your server and see how long it takes. During the DELETE / INSERT the system will be not as responsive as usual but other sessions can still perform SELECTs without any fears of deadlocks, memory leaks or system crashes. Hint: DB servers are usually disk-bound, so getting a proper RAID array is usually a very good investment.

On the other hand, if the performance is critical, you can select one of the alternative approaches described in this thread:

• partitioning if you have the license
• table renaming if you don't, but be mindful that DDLs on the fly can cause some side effects such as object invalidation, ORA-06508...

Answer 10

I'm going with an upsert method here.

I added an additional "delete" column to each of the tables.

When I begin processing the feed, I set the delete field for every record to '1'.

Then I go through a serious of updates if the record exists, or inserts if it does not. For each of those inserts/updates, the delete field is then set to zero.

At the end of the process I delete all records that still have a delete value of '1'.

Thanks everybody for your answers. I found it very interesting/educational.