Databases Reference
In-Depth Information
Disaster Recovery
ure scenarios, Oracle would likely recommend a combination of both Oracle RAC and
Data Guard, though the user should be cautious of the increased cost of the solution.
IBM HADR allows for fix packs (patch sets) and OS level fixes to be applied in a
rolling fashion. For example, the following steps can be deployed to maintain maximum
availability while patches are applied:
1.
Stop HADR on the standby.
2.
Apply the DB2 fix or OS fix on the standby.
3.
Start HADR on the standby—database will automatically resynchronize.
4.
Perform a switch-roles takeover.
5.
Stop HADR on the new standby (old primary).
6.
Apply the DB2 fix or OS fix to this server.
7.
Start HADR on the new standby (old primary)—database will automatically
resynchronize.
At this point, since HADR is intended to be a peer-to-peer HA solution, you can
leave the roles as they are above. Alternatively, you can perform a takeover to switch
roles again to get back to the original primary/standby configuration. With Oracle Data
Guard, a physical standby database does not support rolling upgrades. Both primary
and standby servers must be using the same patch set.
B.1
Standby Remains “Hot” during Failover
The following sequence of events occurs on the standby during an HADR takeover:
1.
Last log buffer is replayed (if not already done).
2.
Undo of in-flight transactions occurs—note that the buffer pool on the standby
is likely full of all the recent updates so there is likely little to no random data
page I/O during undo recovery.
3.
New transactions are allowed to access the database. Note that the buffer pool
and all other memory structures remain allocated.
With Data Guard, in order to convert a standby into a primary (during either
failover or when switching roles) the standby database must be shutdown and started up
again. This results in buffer caches, catalog caches, and package caches (library caches)
being torn down and recreated. Therefore, a significant “brown out” period would fol-
low a Data Guard failover. According to a presentation given by Angelo Pruscino, Prin-
cipal Architect, Oracle Corporation, there is an issue with warming the buffer cache on
a cold failover that can take “5+ minutes” to resolve.
