jerryshao commented on PR #8252:
URL: https://github.com/apache/gravitino/pull/8252#issuecomment-3227455738

   > > > > My biggest concern is the maintainability of the clear resource 
logic. It seems that we are doing some complex work that should belong to the 
JVM level's GC, and this workload will get out of control as the project 
develops.
   > > > > I don't know if we have other alternative solutions to solve the OOM 
problem, but the current solution seems not to be the optimal one in the long 
run. WDYT? @jerryshao @FANNG1
   > > > 
   > > > 
   > > > I agree. My thought is that though we may have lots of catalogs, the 
number of providers is limited, which means the number of classloaders are 
limited, we don't have to create a classloader for each catalog. We can reuse 
the classloader for the same type of catalogs, and with this we can reduce the 
class-related OOM issues.
   > > > Also, many engines like Spark/Trino that support multiple connectors 
will face the same problem, especially for Trino, which can also support 
dynamic connector register/unregister. How do they handle this problem? We'd 
better have a long-term solution.
   > > > For now, I think we can have a short solution, but we should clearly 
separate this from the main logic, so we can remove it easily. Also, we should 
add the comments as possible as we can.
   > > 
   > > 
   > > I see.
   > > I will try to introduce the class loader reuse mechanism in the next 
version, and possibly there will be static variable sharing problems like `UGI` 
exists, so conducting full tests is necessary before this changes.
   > 
   > In Trino, there will be only one classloader for each type of catalog.
   
   We can also have the same solution as I mentioned above. The problem we face 
is the same as Trino (they also have the same UGI issue). Please investigate it 
and have a long-term solution.


-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: [email protected]

For queries about this service, please contact Infrastructure at:
[email protected]

Reply via email to