Details on TAF support in Oracle AS 10g

Question

Hi experts,
I'm currently writing a Master's Thesis on High Availability of Applications based on Sun’s JEE platform.
With Oracle being one of the big players in this market I'm going to investigate the features of the Oracle product stack and compare it with other solutions including opensource.
At the moment I'm looking for details on TAF-support in Oracle AS 10g. Let’s assume one has installed the latest release of the Oracle 10g database in a RAC configuration on two or more physical nodes to store the applications data. Let’s further assume one has set up the latest release of the Oracle AS 10g on different hosts building an AS farm with clustered OC4J-instances. Now one would like to connect a JEE application running within these OC4J-instances to the RAC database.
Therefore one can configure an OC4J datasource to use a service definition supporting network failover and transparent application failover (TAF) specified in the tnsnames.ora on the AS hosts or directly describe the service in the data-sources.xml. To my knowledge only datasources configured to use the OCI client can be used with TAF. Everything right so far?
Now the interesting part. The best what TAF can do is automatically reconnect to the next known database node if it’s current node is not reachable or fails, and continue processing an open cursor for select so that the application can continue to use it’s current connection.
If the application is running a database transaction it can only be rolled back and rerun after failover.
In a good old plain Java application one had to implement the OracleOCIFailover interface and register a callback with the db connection in order to receive failover events. Therefore you need to process the query in a different thread.
In a JEE server only the container should be responsible for creating threads, or? How does the Oracle AS support TAF? How can a user application get notice of a failover and react appropriately. Is the behaviour always the same no matter one decides to implement the data access via simple DAOs or entity beans (CMP/BMP)? I assume it is safe to use a connection pool and container will take care of the necessary cleanup jobs, right? Will the complete connection pool be reconstructed or only connections to the failed node?
Does the application server internally use these OCI callbacks and wrap them up into special Exceptions? Is there anything an application programmer has to do in order to make his data access code ready for TAF, especially in case of failover while the db is processing transactional code?
Any insights and/or code snippets would be greatly appreciated!
Ah, one last question. From a HA perspective is it generally a bad idea to put application logic into the database (use CDM RuleFrame and so on) as there is no support in preserving package states or do you think this makes no difference as one has always to repeat everything since the last commit point no matter in which tier the code resides?
During the next days I’m going to set up the above environment for testing. A 10.0.2.1 RAC on RHE4 is already up and running fine. AS 10.1.3.0.0 on RHE4 is following today or tomorrow.
Many thanks for all answers!
Regards,
Eric

Answer 1

Hi,
well, I thought I could dig into the OC4J documentation before installing the server. After reading the OC4J Service Guide for 10.1.2 and 10.1.3 I'm a little bit confused.
First I read the 10.1.2 Service Guide and almost everything seemed to be clear. I found a lot of answers to the questions above. I'd like to sum them up. Please correct me if I got something wrong!
- TAF only works on top of OCI driver
- TAF can be used with connection pooling
- OC4J takes care of connection pool cleanup (OC4J and OCI Pool)
These two pools are mentioned. It's not clear to me how they work together. Do one of you know some source where I can get more information about this?
- after a failover all connections in the pool are checked by OC4J and only alive ones will survive
- in case of a failover TAF issues an exception
- the application has to catch this exception and acknowledge with a rollback, after that the transaction can be restarted (this seems to be the responsibility of the applications code, oc4j will not replay something automatically; If I got it right, CMP entity beans will be an exception to this rule - there OC4J will proceed transparently...)
- the application can use OCI callbacks to customize the failover behavior (well I would like to see some source code example here, as I can't imagine how this could look like)
After that I read the 10.1.3 service guide. The section about datasources in conjunction with HA is much shorter. Besides TAF new? mechanism is introduced, namely Fast Connection Failover (FCF). There is not much information about it but references to the JDBC Guide, because FCF is a feature of the implicit JDBC connection cache. Where cache seems to be just another word for pool.
By the way, the example on p. 4-36 looks wrong at first glance. Enabling FCF and setting:
<property name="fastConnectionFailoverEnabled" value="false"/> ???
Seems to be a typo in the docu, or?
I got the impression, that the text is written as if FCF being TAF's successor, without explicitly saying this. Is this right?
So I proceeded reading the JDBC Developer Guide to lean more about FCF. Here is what I learned:
- since 10g available
- FCF is driver independend (you can also use the thin driver)
- prerequisites are implicit connection cache (pool) enabled, ONS must be configured
- in case of a failover the transcation will be rolled back automatically (not by the application); whereas using TAF you have to use a rollback to acknowledge the error
- only one exception ORA-17008 to catch
- application has to obtain a new connection from the cache (pool), where as using TAF the old connection object can be used, and resend the transaction (same as with TAF)
So when to use what? The comparison on p. 27-8 was not very helpful to find an answer to this question.
On p. 28-1 one can read that TAF does not work with the OCI Connection Pool. That's seems to be contradictory to the statement in Service Guide of 10.1.2 that OC4J takes care of both pools (OC4J and OCI). Or has something changed in 10.1.3 in that regard? May be this is because of the changes in Cache or Pool Management?
What is the advantage of using TAF within OC4J in 10g, if there is any?
What about select statements with FCF? You can only read about transactions. I'd assume the application would get the same exception executing a query (ORA-17008) as executing a transaction, right? So TAF should have the advantage of transparently proceeding selects after a failover. Am I right?
I'd love to see more examples and comparisons of TAF and FCF.
Regards,
Eric