On 9/13/2006 10:15 PM, cbbrowne at ca.afilias.info wrote:
>> On 9/11/2006 5:47 PM, Christopher Browne wrote:
>>> I thought I should drop a note to the list explaining why there's no RC4
>>> just yet.
>>>
>>> Jan Wieck has discovered a case where multiple MOVE SET requests lead to
>>> loss of data; he had anticipated that a relatively minor modification to
>>> the ACCEPT_SET event would address this, but is still finding some "data
>>> loss" cases.
>>
>> I fixed one part of the problem today and will probably finish all of it
>> tomorrow. MOVE SET was entirely broken for subscribers other than the
>> old or new origin. On those, the setsync status was not adjusted
>> correctly. If the old origin had a higher advanced event sequence
>> (likely), this caused the subscribers to ignore all syncs from the new
>> origin until its event sequence exceeded that.
> 
> This seems like reason for a 1.1.6 release, assuming it can be applied to
> 1.1.5 without completely unreasonable effort...
> 
> I *very* much look forward to seeing that soon...

Now this will knock you out of your socks. 1.1.5 does not get out of 
sync in this 3 node setup (you can use test_2_pgbench and the two 
handover scripts from test_1, they are compatible).

The bizarre thing is that the code in moveSet_int(), that adjusts the 
setsync to the last seen SYNC from the new origin (not precise but 
apparently good enough) exists unchanged in HEAD ... but for some reason 
failed to do its job.


Jan

-- 
#======================================================================#
# It's easier to get forgiveness for being wrong than for being right. #
# Let's break this rule - forgive me.                                  #
#================================================== JanWieck at Yahoo.com #