[jira] Updated: (DERBY-3192) Cache session data in the client driver

["Dyre Tjeldvoll (JIRA)" <ji...@apache.org>]

     [ https://issues.apache.org/jira/browse/DERBY-3192?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]

Dyre Tjeldvoll updated DERBY-3192:
----------------------------------

    Attachment: derby-3192.prelim1.diff

I've attached a preliminary patch (prelim1) which implements caching of the isolation level in the client driver. It is a fully working solution and all the tests pass, but there is a particular implementation issue which I especially would like to get feedback on:

When executing a query with a scrollable result set a complete row will be fetched each time Statement.flowExecute() is called, even if this requires multiple round-trips. This means that trailing piggy-backed session data may end up being processed either by the code assembling a complete row for a scrollable result set, or by flowExecute (normal case). This will complicate the logic in because flowExecute needs to know if the session data was processed when the row was assembled, or if the session data is still waiting to be read.

In the attached patch I've chosen to work around this problem by detecting split rows destined for a scrollable result set on the server, and then delay sending the session data until the final fragment of the row is sent. Doing it this way has three benefits:
1) No piggy-backing logic needs to be added to the row-assembly logic
2) No complication of flowExecute
3) No unnecessary piggy-backing is done while a row is re-assembled

The primary drawback is that the DRDAConnThread logic gets more complicated and normal DRDA flow is interrupted. Normally each DRDA request is receives a corresponding reply in next message going back to the client. In this patch the answer to a session info request is delayed until all fragments of a row have been sent. This breaks the normal DRDA chaining mechanism. The details can be seen by searching for comments beginning with 'TODO DT' in the patch.

All review comments will be much appreciated.

> Cache session data in the client driver
> ---------------------------------------
>
>                 Key: DERBY-3192
>                 URL: https://issues.apache.org/jira/browse/DERBY-3192
>             Project: Derby
>          Issue Type: Improvement
>          Components: JDBC, Network Client, Network Server, Performance, SQL
>    Affects Versions: 10.3.1.4
>            Reporter: Dyre Tjeldvoll
>            Assignee: Dyre Tjeldvoll
>         Attachments: derby-3192-test.fup1.diff, derby-3192-test.fup2.diff, derby-3192-test.v1.diff, derby-3192-test.v1.stat, derby-3192.prelim1.diff
>
>
> The reason for doing this is to avoid a rather
> substantial performance hit observed when the client driver is used
> together with an appserver that uses connection pooling. There are two
> problems:
> 1) The connection pool will compare the isolation level it has
> stored for the connection with the value returned from
> Connection.getTransactionIsolation() each and every time someone
> requests a new connection from the pool.
> 2) The users of the connection pool (ab)use it to avoid having to keep
> track of their current connection. So each time a query needs to be
> executed a call to the connection pool's getConnection() method is
> made. Getting a connection from the connection pool like this also
> means that a new PreparedStatement must be prepared each time.
> The net result is that each query results in the following sequence:
> getConnection()
> getTransactionIsolation() --> roundtrip + lookup in server's statement cache
> prepareStatment()         --> roundtrip + lookup in server's statement cache
> executeQuery()            --> roundtrip
> Arguably this is a "user error" but when suggesting this I'm kindly
> informed that this works "just fine" with other datbases (such as
> PostgreSQL and ORACLE). 
> The reason why it works is that these databases do statement caching
> in the driver. I've tried to implement a very (too) simple statement
> cache in Derby's client driver and to re-enable caching of the
> isolation level (see
> https://issues.apache.org/jira/browse/DERBY-1148). With these changes
> I observe a marked performance improvement when running with appserver
> load. 
> A proper statment cache cannot be implemented without knowing what the
> current schema is. If the current schema has changed since the
> statement was prepared, it is no longer valid and must be evicted from
> the cache.
> The problem with caching both the isolation level and the current schema in
> the driver is that both can change on the server without the client
> detecting it (through SQL and XA and possibly stored procedures).
> I think this problem can be overcome if we piggy-back the information we would 
> like to cache on messages going back to the client. This can be done by
> utilizing the EXCSQLSET DRDA command. According to the DRDA spec (v4, volume 3, 
> page 359-360) it is possible to add one or more SQLSTT objects after SQLCARD in the reply,
> I think it would be possible to cache additional session information when this becomes relevant.  It
> would also be possible to use EXCSQLSET to batch session state changes
> going from the client to the server.

-- 
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.