[jira] [Commented] (DERBY-5265) LOB replication limit

["Knut Anders Hatlen (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/DERBY-5265?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13046361#comment-13046361 ] 

Knut Anders Hatlen commented on DERBY-5265:
-------------------------------------------

I don't know why the maximum value was set to 1MB. Increasing/removing the limit sounds like a reasonable request. Note that there are 10 of these buffers, so when the property is set to 1 MB, you end up with 10 MB of buffer space.

I think buffering the logs in memory was chosen for convenience (so that replication didn't need to change the logic that decides when the checkpoint is allowed to delete log files, and probably also because hooking on to the code that wrote the log to disk was easier than actually reading the log back from disk) on assumption that a small buffer would be sufficient for most minor network glitches. Adding the ability to read log files from disk when the buffer is full would certainly make replication more robust, and it would be a useful improvement.

> LOB replication limit
> ---------------------
>
>                 Key: DERBY-5265
>                 URL: https://issues.apache.org/jira/browse/DERBY-5265
>             Project: Derby
>          Issue Type: Improvement
>          Components: Replication
>            Reporter: Brett Wooldridge
>
> I have an issue with replication as it pertains to LOBs (BLOBs and CLOBs).
> According to the documentation...
> If the master looses connection with the slave, "transactions are
> allowed to continue processing while the master tries to reconnect with 
> the slave. Log records generated while the connection is down are 
> buffered in main memory. If the log buffer reaches its size limit before 
> the connection can be reestablished, the master replication
> functionality is stopped."
> And the documentation for derby.replication.logBufferSize says the
> maximum size of the buffer is 1048576 (1MB).
> This seems to imply that if I have a database in which I store LOBs
> which are, for example, 256K in size, and the connection between 
> master and slave is severed, I can perform 4 inserts or less before 
> the master gives up.  I would like to file a request that this limit be raised
> considerably or eliminated altogether.
> I have two servers (master and slave) running 64-bit JVMs, 64GB of memory each,
> SSD drives, connected by 10GbE fiber.  I would like to dedicate as much memory 
> as I want to deal with a disconnect/resume scenario (to avoid the onerous failover).  
> At an insertion rate of 16 rows per second (~4MB), currently the setup would
> tolerate a connection interruption of a fraction of a second.  A 1GB buffer would
> afford a connection interruption of ~250 seconds (for example, rebooting the fiber 
> switch).
> Lastly, why does Derby even bother to buffer logs in memory?  Can't it just keep 
> an offset/marker into the transaction log files, or better insert a special replication
> log entry, and replay transactions from there, rather than buffering them in memory?

--
This message is automatically generated by JIRA.
For more information on JIRA, see: http://www.atlassian.com/software/jira