On 03/16/2017 08:59 AM, Joe Julian wrote:
In the last few releases, we have changed replace-brick command such
that it can be called only with "commit force" option. When invoked,
this is what happens to the volume:
a. distribute only volume: the given brick is replaced with a empty
brick with 100% probability of data loss.
b. distribute-replicate: the given brick is replaced with a empty
brick and self heal is triggered. If admin is wise enough to monitor
self heal status before another replace-brick command, data is safe.
c. distribute-disperse: same as above in distribute-replicate
My proposal is to fully deprecate replace-brick command for
"distribute only" volumes. It should print out a error "The right way
to replace brick for distribute only volume is to add brick, wait for
rebalance to complete and remove brick" and return a "-1".
I am responding late, assuming this is still not done or WIP. Correct me
if I am wrong.
Yes, we need the above. As it really does not make any sense in the way
the current (replace brick) command is structured to lose a pure
distribute brick.
It makes sense.
I just don't see any use of add-brick before remove-brick except the
fact that it will
help to keep the overall storage capacity of volume intact .
What is the guarantee that the files on the brick which we want to
replace
would migrate to added brick?
If a brick, which we want to replace, is healthy and we just want to
replace it then perhaps we should provide
a command to copy those files to new brick and then remove the old
brick.
We used to have a command that did just that. It was replace-brick.
This involves some rebalance trickery IMO (if we leverage that to
replace the brick in a distribute only volume), which we do not have. I
would suggest a github issue be added for such support, and take in the
prevention of current replace-brick on distribute only volumes as the
first priority (for obvious reasons as stated by talur).
Shyam
_______________________________________________
Gluster-devel mailing list
Gluster-devel@xxxxxxxxxxx
http://lists.gluster.org/mailman/listinfo/gluster-devel
