[Pgcluster-general] rlog issue

[Markus Schmidleitner]

hi list, hi mitani!

i'm currently working on backporting some stuff from the cybercluster
project to pgcluster and doing some code-refactoring (make
network-operations centralized and non-blocking, remove dead code and so
on).

while invastigating the rlog code i found out that rlog's are always
sent to lower cascade but never to the upper one. could anyone explain
why?


for example assume following system composition:

replicator1 ---- replicator2
     |      \  /     |
     |       /\      |
     |      /  \     |
     |     /    \    |
  cluster1       cluster2


- cluster1 has replicator1 as replication server
- cluster2 has replicator2 as replication server
- replicator1 has cluster1 and cluster2 as cluster server
- replicator2 has cluster1 and cluster2 as cluster server
- replicator1 is upper and replicator2 is lower

(i know this is quiet a little strange config but it is necessary for us
- we have multiple database nodes (server) and each of them runs a
cluster and a replication server - each cluster is connected to it's
local replicator, each replicator is connected to all clusters and the
replicators are cascaded)


for example performing INSERT on cluster1:

as far as i understand rlog is working the following way:
  1. cluster1 sends query to replicator1
  2. replicator sends query over rlog-method to lower cascade
(replicator2)
  3. replicator1 executes query on cluster2 and cluster1 (eg. sends some
ok-message to source (cluster1) and executing query on cluster2)
  4. cluster2 sends query to replicator2
  5. replicator2 recognizes the executing query comes from rlog, check
wheter the query was executed on cluster2 and send result back to
cluster2
  6. cluster2 and cluster1 send result back to replicator1
  7. replicator1 send some ok-message back to cluster1 (source)

i think that's a short description of how rlog is working.


so what about if source is cluster2?
replicator2 will not perform any rlog becouse rlog-messages are only
sent to lower cascade! 
is this expected behavior or a bug????

is there any other check to prevent replication missbehavior and if so,
why using still rlog? am i missing something?


my backport is working nicely so far and as soon as i finished all of my
tests i will put a patch on the list!


regards,
Markus Schmidleitner


__________________________________________________________________________________
Dieses Mail wurde vom Infotech SecureMail Service ueberprueft und fuer sicher befunden.
Fuer weitere Informationen zu Infotech SecureMail Service waehlen Sie bitte: www.infotech.at

This email has been scanned by Infotech SecureMail Service and it has been classified as secure.
For more information on Infotech SecureMail direct your web browser to: www.infotech.at