Note: "permalinks" may not be as permanent as we would like,
direct links of old sources may well be a few messages off.
On 23/03/17 10:20 AM, Philipp Reisner wrote: > Hi, > > Avoiding data divergence is the motto of this drbd-9 release. One the > one hand the fencing mechanism as known from the drbd-8.x world finally > works in drbd-9 as expected. Including the crm-fence-peer mechanism! > > But those who care about fencing know that it is not trivial to implement. > It is about making sure the cluster does the right thing, if the main > replication network is interrupted. That involves having a second network > for pacemaker communication or to control fencing devices. -- Quite > frequently we see that such a "second network" is just an other VLAN on > the same physical infrastructure -> Which makes the whole fencing effort > point less. > > An other approach to avoid data divergence in partitioned clusters is > to require quorum to modify the data. This approach does not need a > second network or fencing devices. But it only works if you have at > least 3 replicas. > > The most outstanding convenience feature implemented in this cycle: > intentional diskless nodes are no longer red in the output of the > status command. > > Please help testing this release candidate. > > drbd > 9.0.7rc1-1 (api:genl2/proto:86-112/transport:14) > -------- > * various fixes to the 2-phase-commit online resize > * fix fencing and disk state transition and from Consistent; Now it is > as complete as it was under drbd-8.4; Necessary for crm-fence-peer > * fix moving previous current-uuid into bitmap slot for peer with > inconsistent disk state. Old code produced false "split-brain detected" > * fix calculation of authoritative nodes. Old code could lead to to > false "split-bain detected" > * udev workaround: Do not fail down if there is a read-only opener, instead > wait up to a second if the read-only-opener goes away. > * a primary notes if a "far away" nodes outdates itself, and records the > fact in its meta-data. > * report back to userspace if a node (or peer) is diskless by intention > * restore "Outdated" peer disk state from the meta-data upon attach > * Quorum to avoid data divergence, an alternative to fencing when the > resource has 3 or more replicas > * improve error reporting in case a promote/down/disconnect/detach > operation failes. E.g. tells which peer declined a state transition > * compiles with Linux 4.10 > > http://www.drbd.org/download/drbd/9.0/drbd-9.0.7rc1-1.tar.gz > http://git.drbd.org/drbd-9.0.git/tag/refs/tags/drbd-9.0.7rc1 > > drbd-utils > 8.9.11rc1 > -------- > * better integration of drbd-proxy in drbd9 scenarios > * OCF resource agent and crm-fence-peer for drbd9 > * introduce quorum support > * expose information if node is intentionally diskless > * do not set peer-device-options twice > * various JSON output fixes > * udev fixes for newer ubuntu releases > * set bitmap to 0 on metadata creation > > http://www.drbd.org/download/drbd/utils/drbd-utils-8.9.11rc1.tar.gz > http://git.drbd.org/drbd-utils.git/tag/refs/tags/v8.9.11rc1 > > best, > phil YES!!! <3 -- Digimer Papers and Projects: https://alteeve.com/w/ "I am, somehow, less interested in the weight and convolutions of Einstein’s brain than in the near certainty that people of equal talent have lived and died in cotton fields and sweatshops." - Stephen Jay Gould