1300 to 3100 lines of code for XA support (was: Re: plans and estimates for supporting JDBC 3.0)

On Wednesday 27 October 2004 16:33, Oliver Jowett wrote:
> Yes; there's no point in implementing XA* until we have a 2PC
> implementation on the server side to back it up.

Ok. I hear that Heikki's 2PC patches are unlikely to get merged any
earlier than 8.1.

> As I understand it, you do not need an XA implementation to be JDBC3
> compliant anyway.

Yes and no. Section 6.5 [1] does not list XA as an explicit
requirement for 3.0 compliance. So, yes, you're right.

On the other hand, Section 6.2 [2] seems to suggest that you must
implement XA, if the backend supports 2PC.

> > Is there a plan for doing this and an estimate of the required
> > effort? (As far as I understand, support for the XA stuff needs
> > to be added on the backend side. Ignoring that for a second, are
> > there any estimates for the amount of work required on the driver
> > side?)
>
> It depends on what the server interface to 2PC ends up looking like.

I spoke with Tom Lane about this. If I understood him correctly,
chances are, the interface will end up looking very much like whatever
Heikki currently has.

> With current proposals it doesn't seem like much work on the driver
> side.

To get a better feel for what "not much work" actually looks like, I
googled for existing implementations.

The lower bound estimate is provided by implementations that "emulate"
distributed transactions. The way this works is, you implement
XAResource#prepare(Xid) [3] to always return XA_OK. In the second
phase, when XAResource.commit(Xid, int) [4] is called, the local
transaction is actually committed.

Despite the fact that this approach cannot provide transactional
semantics, it's pretty widely used.

The MySQL JDBC driver used to have this up to and including version
2.0.14 [5]. Since it didn't really work, it was removed from later
versions. MySQL's implementation was about 2100+ lines of code [6].

JBoss provides (or used to provide) similar wrappers in its
org.jboss.pool.jdbc.xa.wrapper.* package, but I can't seem to find the
relevant sources [7].

Jonas, too, provides XA wrappers implemented in 1300+ lines of code
[8].

There's also XAPool. The size of its "standard" implementation is
about 3100 lines of code[9]. (They have a test case for PostgreSQL
[10].)

So, based on this, I'd say it takes about 1300 to 3100 lines of code
to provide fake XA support.

As a side note, it looks like the PostgreSQL JDBC driver used to
provide emulated XA support [11].

As far as real implementations go, MaxDB (previously known as SAPDB)
claims to support 2PC and its JDBC driver implements XA. The size of
MaxDB's implementation is about 1800+ lines of code [12].

One could also look at the size of the relevant .class files in the
Oracle JDBC driver as an upper bound estimate.

Vadim

------------------------------------------------------------------------------

Footnotes

1. JDBC 3.0 API Compliance, p. 40:

| A driver that is compliant with the JDBC 3.0 API must do the
| following:
|
| * Comply with the JDBC 2.0 API requirements
|
| * Include the following required interfaces:
|
| * java.sql.ParameterMetaData
| * java.sql.Savepoint
|
| * Fully implement the DatabaseMetaData interface, including the
| following methods added in the JDBC 3.0 API:
|
| * supportsSavepoints
| * supportsNamedParameters
| * supportsMultipleOpenResults
| * supportsGetGeneratedKeys
| * getSuperTypes
| * getSuperTables
| * getAttributes
| * getResultSetHoldability
| * supportsResultSetHoldability
| * getSQLStateType
| * getDatabaseMajorVersion
| * getDatabaseMinorVersion
| * getJDBCMajorVersion
| * getJDBCMinorVersion

2. Guidelines and Requirements, p. 38

| * Drivers should provide access to every feature implemented by
| the underlying data source, including features that extend the
| JDBC API. When a feature is not supported, the corresponding
| methods throw an SQLException. The intent is for applications
| using the JDBC API to have access to the same feature set as
| native applications.

3. http://java.sun.com/j2se/1.4.2/docs/api/javax/transaction/xa/XAResource.html#prepare%28javax.transaction.xa.Xid%29

4. http://java.sun.com/j2se/1.4.2/docs/api/javax/transaction/xa/XAResource.html#commit%28javax.transaction.xa.Xid,%20boolean%29

5. http://prdownloads.sourceforge.net/mmmysql/mm.mysql-2.0.14-you-must-unjar-me.jar?download

6. This is for mm.mysql-2.0.14-you-must-unjar-me.jar:
| $ find . -name \*.java | grep xa/
| ./org/gjt/mm/mysql/xa/ClientConnection.java
| ./org/gjt/mm/mysql/xa/TwoPhaseConnection.java
| ./org/gjt/mm/mysql/xa/TxConnection.java
| ./org/gjt/mm/mysql/xa/XAConnectionImpl.java
| ./org/gjt/mm/mysql/xa/XADataSourceImpl.java
| $ find . -name \*.java | grep xa/ | xargs cat | wc -l
| 2176

7. http://jboss.sourceforge.net/snapshots/jboss-head-snapshot.tar.gz

This 66MB file does not have anything under
jboss-head/pool/src/main/org/jboss/pool/jdbc/xa/wrapper

The CVS repository has way too much stuff in it. I'm not sure
where to look for org.jboss.pool.jdbc.xa.wrapper.* classes:
http://cvs.sourceforge.net/viewcvs.py/jboss/

8. http://cvs.forge.objectweb.org/cgi-bin/viewcvs.cgi/jonas/jonas/src/org/objectweb/jonas/jdbc_xa/

| $ cat ./jonas/src/org/objectweb/jonas/jdbc_xa/*.java | wc -l
| 1326

9. http://cvs.forge.objectweb.org/cgi-bin/viewcvs.cgi/xapool/xapool/src/org/enhydra/jdbc/standard/

| $ cat xapool/src/org/enhydra/jdbc/standard/*.java | wc -l
| 3107

10. http://cvs.forge.objectweb.org/cgi-bin/viewcvs.cgi/*checkout*/xapool/xapool/test/jotmxapooltest/config/postgres.properties?rev=1.1

11. http://people.redhat.com/~vadimn/scratch/pgsql-jdbc/changelog.html#2002-08-14-16:35

12. Download maxdb-source-7_5_00_18.tgz from
http://www.mysql.com/products/maxdb/
http://dev.mysql.com/downloads/maxdb/7.5.00.html
http://dev.mysql.com/get/Downloads/MaxDB/7.5.00/maxdb-source-7_5_00_18.tgz/from/pick

The source for the JDBC driver is in
V75_00_18/MaxDB_ORG/sys/src/jv/sapdbc.src.jar

After unjarring,

| $ grep -lF XA com/sap/dbtech/jdbcext/*.java | xargs cat | wc -l
| 1806
| $ cat com/sap/dbtech/jdbcext/XA*.java | wc -l
| 1138