[jboss-jira] [JBoss JIRA] Commented: (JGRP-1251) NAKACK: overwriteDigest() doesn't adjust sequence number

[Bela Ban (JIRA)]

    [ https://jira.jboss.org/browse/JGRP-1251?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12563797#comment-12563797 ] 

Bela Ban commented on JGRP-1251:
--------------------------------

The issue is caused by B installing an outdated digest sent from A (as part of the state transfer) and trashing the NakReceiverWindow and creating a new one. Creating a new NRW means that all existing messages sent by B (but not yet received by A) are lost and cannot be retransmitted (that's why the error message is seen).

Example:
- B has digest B: 0 5 (5) // lowest stable seqno=0, highest delivered=5, highest received=5
- B fetches the state from A
- A returns the state plus the digest B: 0 3 (3)
- This means messages B4 and B5 have not yet been seen by A, and are therefore *not yet* part of the state. In other words, B4-5 have to be delivered *again* by B, even though B already delivered them !!
- WHY ?
- B delivered B4-5, which caused some state changes
- Then B received and set the state, replacing the existing state
- B's new state (fetched from A) does *not* have messages B4-5 in it, so they need to be received again by A and B, to come up with the same state !

SOLUTION:
- We *cannot* set B's seqno to 3, as this would have B send *new* (and *different* !) messages 4 and 5 again !
- We therefore leave B's seqno
- We need to set highest_delivered in B's NRW to 3 though !!
- This causes B to re-deliver messages 4 and 5



> NAKACK: overwriteDigest() doesn't adjust sequence number
> --------------------------------------------------------
>
>                 Key: JGRP-1251
>                 URL: https://jira.jboss.org/browse/JGRP-1251
>             Project: JGroups
>          Issue Type: Bug
>            Reporter: Bela Ban
>            Assignee: Bela Ban
>             Fix For: 2.10.2, 2.11.1, 2.12
>
>         Attachments: GroupCommunication.java, tmp.xml, ViewHandler.java
>
>
> NAKACK.overwriteDigest() is used by the 2 state transfer protocols (STATE_TRANSFER and STREAMING_STATE_TRANSFER).
> When A has a digest A:250,B:10,C:160, and gets the state (as a result of calling JChannel.getState() with a digest of A:230,B:10,C:160, it'll set its own digest to A:230 (from A:250). 
> However, A's sequence number (seqno) will remain at 250 ! This means, when A sends the next message, it'll send A:251, but because we set the digest to A:230, A will ask itself to retransmit messages A:230-250 !
> SOLUTION: when receiving digest A:230, A should set its seqno to 230
> WORKAROUND: use FLUSH, which doesn't ship digests on state transfer

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