Re: Unregistering the driver from DriverManager

Christopher,

(1) Would you mind testing if PR #241
<https://github.com/pgjdbc/pgjdbc/pull/241> on Github fits your needs and
works in the OSGi environment you're using?
Note that it don't check presence of BundleContext via Class.forName()
because this class could be erroneously present in the classpath even
without an OSGi Framework being started.

(2) Driver.deregister() method (named after DriverManager method) has been
exposed so as to be used in other contexts like Java EE.

Alexis

2014-12-29 14:17 GMT+01:00 Christopher BROWN <brown(at)reflexe(dot)fr>:

> There are two main situations where it would be useful to automatically
> unregister the driver:
>
> 1) OSGi - and the suggestion of using BundleActivator.stop() would be a
> good fit here (as long as care is taken to ensure resolution=optional for
> other dependencies)
>
> 2) Java EE web applications, using a ServletContextListener, perhaps using
> annotations as described in
> https://blogs.oracle.com/swchan/entry/servlet_3_0_annotations (but this
> would exclude older application servers)
>
> With regards to (2), I generally place JDBC drivers in the main
> classloader of the application server, as opposed to embedding in
> WEB-INF/lib when working with webapps. Also, not all of the webapps I have
> to deal with (from time to time, it's not my main focus) are up to Servlet
> 3.0, many as still stuck on 2.5. And in any case, embedding JDBC drivers
> in webapps (without matching versions) then accessing them via
> DriverManager is may cause class lookup problems.
>
> A good solution to (1) above to me would be like this then (building on
> the suggestion of Alexis):
>
> - keep the static block in driver
> - check -- via Class.forName("org.osgi.framework.BundleContext") -- if
> OSGi classes are visible, implying that the driver has been loaded as a
> bundle, and skip the DriverManager registration in the static block (unless
> forced via a system property, just in case something breaks for someone
> relying on current behavior)
> - register the driver in DriverManager in the BundleActivator.start()
> method
> - unregister it (same instance, kept as a reference) in the
> BundleActivator.stop() method
>
> The only reason I mention (2) is because it might be useful to share some
> common code. Or not. In any case, (1) is the only requirement at this
> time and (2) isn't as much of a problem.
>
> --
> Christopher
>
>
> On 29 December 2014 at 13:45, Alexis Meneses <alexis(dot)meneses(at)gmail(dot)com>
> wrote:
>
>> If the only concern is OSGi environments, I think that unregistering
>> could be handled in a BundleActivator
>> <http://www.osgi.org/javadoc/r4v43/core/org/osgi/framework/BundleActivator.html>.stop()
>> implementation bundled with the driver.
>>
>> See pending issue #71 <https://github.com/pgjdbc/pgjdbc/issues/71> on
>> Github.
>>
>>
>> 2014-12-29 12:48 GMT+01:00 Dave Cramer <pg(at)fastcrypt(dot)com>:
>>
>>> I have no objection to an unregister static method being added. It's not
>>> in the API so it would not effect anything really
>>>
>>> Dave Cramer
>>>
>>> dave.cramer(at)credativ(dot)ca
>>> http://www.credativ.ca
>>>
>>> On 29 December 2014 at 04:53, Christopher BROWN <brown(at)reflexe(dot)fr>
>>> wrote:
>>>
>>>> Hello,
>>>>
>>>> I'm starting to integrate the Postgresql JDBC driver into an OSGi
>>>> environment, as an OSGi bundle. I'm evaluating the different ways to avoid
>>>> a classloader leak with DriverManager when hot-swapping the driver bundle
>>>> without restarting the host application, and am seeking suggestions on best
>>>> practice regarding the Postgresql JDBC driver.
>>>>
>>>> Another bundle (which I provide, it's not third-party) will directly
>>>> depend upon it (loading classes directly, namely org.postgresql.Driver);
>>>> when the Postgresql JDBC driver classes are loaded, the other bundle will
>>>> create a DataSource using a JDBC connection pool, and register the
>>>> DataSource as an OSGi service. Normally, that's all that will happen
>>>> during the application lifecycle, but in principle, it's possible for the
>>>> administrator to want to replace say the 9.3 driver with the 9.4 driver by
>>>> removing the 9.3 ".jar" at runtime, and replacing it with the 9.4 ".jar",
>>>> all at runtime; when the first ".jar" is deleted, the dependent bundle is
>>>> knocked offline, unregistering the DataSource automatically, and notifying
>>>> all clients; when the second is installed, the application is once again
>>>> fully-functional (and all this normally occurs within a few hundred
>>>> milliseconds).
>>>>
>>>> Looking at the source code, I can see that the org.postgresql.Driver
>>>> class registers itself in a "static" block with DriverManager (which is the
>>>> correct behavior regarding the JDBC spec). However, short of a brute-force
>>>> loop -- like this one: http://stackoverflow.com/a/5315467 (enhanced to
>>>> check the class name of each driver, to avoid clobbering unrelated driver
>>>> registrations) -- is there any other approach possible or that could be
>>>> added, say a NonRegisteringDriver (superclass of Driver, with all logic
>>>> except for the static initializer) or an "unregister()" static method, or a
>>>> field containing the registered Driver instance?
>>>>
>>>> Thanks,
>>>> Christopher
>>>>
>>>>
>>>
>>
>