5:21 p.m.
reference http://opensource.atlassian.com/projects/hibernate/browse/HHH-5697
For multi-tenancy implemented by sepaerate schema we need the ability to tell
the ConnectionProvider about the tenant for the given getConnection() request.
I really see 3 approaches to this:
1) Have 2 hierarchies here. The current ConnectionProvider contract remains
the same. Add a new MultiTenantConnectionProvider with methods accounting for
tenant
2) Just alter the ConnectionProvider contract to pass information in. If we
go this route I prefer the "parameter object" pattern where we pass in
ConnectionOptions interface (see issue).
3) Use contextual lookup. ConnentionProviders interested in (or capable of
understanding) mulit-tenancy would perform some kind of "contextual"
(ThreadLocal, etc) lookup for the needed information.
Thoughts? Discussions?
---
Steve Ebersole <steve(a)hibernate.org>
http://hibernate.org
4:25 p.m.
I think for a user the best option (easiest to understand) would be 1).
3) is too magical. 2) is putting everything in one sack.
Adam
On Mar 22, 2011, at 10:21 PM, Steve Ebersole wrote:
--
Adam Warski
http://www.warski.org
http://www.softwaremill.eu
5:24 p.m.
But 3 is how it's done today since we don't have the other solutions.
So yeah it's magical, but people do manage to make it work. I will
concede that it is not always perceived as straightforward, however.
Something to think about in the grander scheme of things is how someone
would deal with obtaining a tenant-specific connection when resource
injection is in use. #3 solves this but what about #1 and #2? There
are some pretty neat things that I can envision wrt CDI but I haven't
thought through what the implementations would look like.
@Inject @TenantSpecific @UserDatabase private EntityManager em;
How would you drill the tenant information into the CP?
-CB
On 03/23/2011 04:25 PM, Adam Warski wrote:
5:58 p.m.
Well I think I am leaning towards a "parameter object" no matter what we do
otherwise. So call this param object interface ConnectionOptions. It would
have a single method today (err, tomorrow when I write all this :) to get the
tenant id.
The tenant id is really associated with the session. Either session
implements that interface or hands it off to a delegate it manages, but thats
the "scope"... the session.
So in your examnple, you would inject the entitymanager's session with the
proper tenant id. Ideally I want this to be specified up front when the
session is opended to make sure there is no possibility of changing it mid-
stream, but that further relies on
http://opensource.atlassian.com/projects/hibernate/browse/HHH-2860
Anyway, in thinking about this sme more I think the way I really like is this:
1) ConnectionProvider remains as is
2) introduce a new provider contract (names as always are preliminary):
public interface SpecializedConnectionProvider extends ConnectionProvider {
public static interface ConnectionOptions {
public String getTenantIdentifier();
}
public Connection getSpecializedConnection(ConnectionOptions);
public void releaseSpecializedConnection(ConnectionOptions, Connection);
}
Having SpecializedConnectionProvider extend ConnectionProvider addressed
another issue I encountered when implementing this. How do you account for
situations where you need *some* connection? Like when reading
java.sql.DatabaseMetaData on "startup". The other option for that is to not
have SpecializedConnectionProvider extend ConnectionProvider (so they are back
to distinct hierarchies) and adding an extra method to
SpecializedConnectionProvider. This additional method would either give
access to a connection with out providing tenant information or if obtaining
java.sql.DatabaseMetaData is deemed the only acceptable use case there then
just have a method to scrape that info passing back DatabaseMetaData is not an
option if that is the decision because of
java.sql.DatabaseMetaData#getConnection
Also, I was not sure if there is really a need for a specialized "release this
connection" method in the case where SpecializedConnectionProvider extends
ConnectionProvider. There is still the ConnectionProvider#closeConnection
method. Now obviously if the herarchies are distinct we would need release
capability on both...
On Wednesday, March 23, 2011, at 04:24 pm, Chris Bredesen wrote:
---
Steve Ebersole <steve(a)hibernate.org>
http://hibernate.org
1:10 p.m.
Here is the other issue with "separate hierarchies"...
Code that deals with connections now becomes more complicated, having to first
see if there is a SpecializedConnectionProvider and if so grab a connection
from it, otherwise check the ConnectionProvider.
Obviously there are techniques to work around that like making all Connection
access route back through the Session which could expose a singular contract
for getting a connection. Then the "dirtiness" could be contained to just
that "thing".
On Wednesday, March 23, 2011, at 04:58 pm, Steve Ebersole wrote:
---
Steve Ebersole <steve(a)hibernate.org>
http://hibernate.org
11:31 a.m.
For info, I like #2 the best
ConnectionOptions can deal in the future with:
- schema based diff
- user based diff ala Oracle VPD
interface ConnectionOptions {
TenantType getTenantType();
String getDefaultSchema();
String getUser() //is that how VPD filters out?
..some more techniques later
}
enum TenantType {
NONE,
SCHEMA,
USER
}
With getTenantType, a ConnectionProvider can return the right connection or yell if it
does not support it. We could also imagine asking the ConnectionProvider to return the
array of supported tenantTypes so that we can raise the exception at startup time.
On 22 mars 2011, at 22:21, Steve Ebersole wrote:
11:43 a.m.
VPD is really the same notion as a tenant. So the ConnectionProvider having
access to the tenant already solves that
On Tuesday, March 29, 2011, at 10:31 am, Emmanuel Bernard wrote:
---
Steve Ebersole <steve(a)hibernate.org>
http://hibernate.org
12:22 p.m.
Yes but on top of Oracle, I could use the VPD approach or the more portable but less
integrated schema approach, right? Somehow the user will be able to chose and the
connection provider will do different magic tricks. Or am I missing some step?
On 29 mars 2011, at 17:43, Steve Ebersole wrote:
12:30 p.m.
The connection provider is different yes. The information needed is the same.
For VPD you issues an ALTER SESSION command on the connection to tell it the
"tenant"
On Tuesday, March 29, 2011, at 11:22 am, Emmanuel Bernard wrote:
---
Steve Ebersole <steve(a)hibernate.org>
http://hibernate.org
1:36 p.m.
The reason I initially pushed the difference is because one could imagine some kind of map
between the tenantId passed tot he session and the schema name that end up being used. But
that might be a bit over engineered and the tenant id + tenantType (tenantStrategy is
probably better) is enough info.
Emmanuel
On 29 mars 2011, at 18:30, Steve Ebersole wrote:
1:52 p.m.
Maybe i am just missing the point of your specific passed values.
As currently implemented we really expect a custom ConnectionProvider to deal
with the tenant identifier. I really only see JNDI-based
DataSourceConnectionProvider as something for which we might reasonable supply
a implementation.
To be honest the ConnectionProvider is not the difficult part to this. The
isolation within the Session and second level cache are by far the more useful
things.
Of course I am open to suggestions here if anyone thinks we should provide
support for concrete, tenant-aware ConnectionProviders out-of-the-box *and*
has an idea what that looks like and how to actually do it in a clean way.
On Tuesday, March 29, 2011, at 12:36 pm, Emmanuel Bernard wrote:
---
Steve Ebersole <steve(a)hibernate.org>
http://hibernate.org
4:39 p.m.
There is something that is escaping me.
In your view, you would have one ConnectionProvider implementation per underlying
multi-tenant strategy + connection lookup strategy?
ie
JNDIDatasourceConnectionProvider (no multi tenancy support)
SchemaJNDIDatasourceConnectionProvider (multi tenancy with one tenant per schema)
VPDJNDIDatasourceConnectionProvider (multi tenancy delegating isolation to the VPD tech)
And one could write a C3P0 alternative etc for these three if he is motivated.
Or is that something else?
Emmanuel
On 29 mars 2011, at 19:52, Steve Ebersole wrote:
5112
days inactive
5119
days old
11 comments
4 participants
participants (4)
-
Adam Warski -
Chris Bredesen -
Emmanuel Bernard -
Steve Ebersole