[
https://jira.jboss.org/jira/browse/JGRP-948?page=com.atlassian.jira.plugi...
]
Bela Ban commented on JGRP-948:
-------------------------------
Hmm, hearsay info might be OK:
- We have A: {A,B} B: {A,B,C} and C: {C}
- A and B merge first, the digests are {A,B} and {A,B,C}, te resulting digest is {A,B,C}
- The resulting digest has hearsay information about C. The digest for C is probably
outdated and contains lower seqnos than C actually has
- Sending any messages from A or B to C will have C discard them because C's view is
still {C}
- Now A and C merge. A: {A,B,C} C: {C}
- Now we get C's real digest and A and C both update their digest for C and reset
their NakReceiverWindows
--> In this scenario, hearsay information was overwritten with correct information once
C merged
MERGE: merging digests based on hearsay
---------------------------------------
Key: JGRP-948
URL:
https://jira.jboss.org/jira/browse/JGRP-948
Project: JGroups
Issue Type: Bug
Reporter: Bela Ban
Assignee: Bela Ban
Fix For: 2.6.10, 2.8
A merge involves consolidating digests from all members. However, we don't get the
digests directly from involved members, but also include 'hearsay', e.g. in A: {A}
and B: {B,C}, B adds digest information from C. This is what I call 'hearsay', the
digest information about C isn't provided by C directly, but by B.
If B hasn't been able to communicate with C for a while, e.g. because B's link to
C was asymmetrical, then B's information about C's digest might be incorrect.
E.g. if B returns #3 as C's highest seqno, but in effect C's highest seqno is
#20, then members other than C will not be able to get messages C#3 - C#20 from C. They
would therefore queue all subsequent messages from C !
SOLUTION:
- On merging, partition coordinators cannot just return their digests, but have to fetch
the digests for the partition from all members
- The digest consolidation only accepts digests from direct members, e.g. when B returns
a digest of (B:24,C:7,D:15) we only use the B part of it
- Simplification of the above: have members only return their lowest, highest received
and highest delivered seqnos, but *not* seqnos they 'know' (hearsay) from other
members
- When a merge coordinator gets digest information from {A,B,D} in a merge between {A,B}
and {C,D}, then it has to fail the merge because it is missing innformation from C !
--
This message is automatically generated by JIRA.
-
If you think it was sent incorrectly contact one of the administrators:
https://jira.jboss.org/jira/secure/Administrators.jspa
-
For more information on JIRA, see:
http://www.atlassian.com/software/jira