[Gluster-devel] geo-rep regression because of node-uuid change
Xavier Hernandez
xhernandez at datalab.es
Wed Jun 21 08:26:10 UTC 2017
That's ok. I'm currently unable to write a patch for this on ec. If no one can do it, I can try to do it in 6 - 7 hours...
Xavi
On Wednesday, June 21, 2017 09:48 CEST, Pranith Kumar Karampuri <pkarampu at redhat.com> wrote:
On Wed, Jun 21, 2017 at 1:00 PM, Xavier Hernandez <xhernandez at datalab.es> wrote:I'm ok with reverting node-uuid content to the previous format and create a new xattr for the new format. Currently, only rebalance will use it.
Only thing to consider is what can happen if we have a half upgraded cluster where some clients have this change and some not. Can rebalance work in this situation ? if so, could there be any issue ? I think there shouldn't be any problem, because this is in-memory xattr so layers below afr/ec will only see node-uuid xattr.This also gives us a chance to do whatever we want to do in future with this xattr without any problems about backward compatibility.
You can check https://review.gluster.org/#/c/17576/3/xlators/cluster/afr/src/afr-inode-read.c@1507 for how karthik implemented this in AFR (this got merged accidentally yesterday, but looks like this is what we are settling on)
Xavi
On Wednesday, June 21, 2017 06:56 CEST, Pranith Kumar Karampuri <pkarampu at redhat.com> wrote:
On Wed, Jun 21, 2017 at 10:07 AM, Nithya Balachandran <nbalacha at redhat.com> wrote: On 20 June 2017 at 20:38, Aravinda <avishwan at redhat.com> wrote:On 06/20/2017 06:02 PM, Pranith Kumar Karampuri wrote:Xavi, Aravinda and I had a discussion on #gluster-dev and we agreed to go with the format Aravinda suggested for now and in future we wanted some more changes for dht to detect which subvolume went down came back up, at that time we will revisit the solution suggested by Xavi.
Susanth is doing the dht changesAravinda is doing geo-rep changes Done. Geo-rep patch sent for review https://review.gluster.org/17582
The proposed changes to the node-uuid behaviour (while good) are going to break tiering . Tiering changes will take a little more time to be coded and tested. As this is a regression for 3.11 and a blocker for 3.11.1, I suggest we go back to the original node-uuid behaviour for now so as to unblock the release and target the proposed changes for the next 3.11 releases. Let me see if I understand the changes correctly. We are restoring the behavior of node-uuid xattr and adding a new xattr for parallel rebalance for both afr and ec, correct? Otherwise that is one more regression. If yes, we will also wait for Xavi's inputs. Jeff accidentally merged the afr patch yesterday which does these changes. If everyone is in agreement, we will leave it as is and add similar changes in ec as well. If we are not in agreement, then we will let the discussion progress :-) Regards,Nithya--
Aravinda Thanks to all of you guys for the discussions! On Tue, Jun 20, 2017 at 5:05 PM, Xavier Hernandez <xhernandez at datalab.es> wrote:Hi Aravinda,
On 20/06/17 12:42, Aravinda wrote:I think following format can be easily adopted by all components
UUIDs of a subvolume are seperated by space and subvolumes are separated
by comma
For example, node1 and node2 are replica with U1 and U2 UUIDs
respectively and
node3 and node4 are replica with U3 and U4 UUIDs respectively
node-uuid can return "U1 U2,U3 U4"
While this is ok for current implementation, I think this can be insufficient if there are more layers of xlators that require to indicate some sort of grouping. Some representation that can represent hierarchy would be better. For example: "(U1 U2) (U3 U4)" (we can use spaces or comma as a separator).
Geo-rep can split by "," and then split by space and take first UUID
DHT can split the value by space or comma and get unique UUIDs list
This doesn't solve the problem I described in the previous email. Some more logic will need to be added to avoid more than one node from each replica-set to be active. If we have some explicit hierarchy information in the node-uuid value, more decisions can be taken.
An initial proposal I made was this:
DHT[2](AFR[2,0](NODE(U1), NODE(U2)), AFR[2,0](NODE(U1), NODE(U2)))
This is harder to parse, but gives a lot of information: DHT with 2 subvolumes, each subvolume is an AFR with replica 2 and no arbiters. It's also easily extensible with any new xlator that changes the layout.
However maybe this is not the moment to do this, and probably we could implement this in a new xattr with a better name.
Xavi
Another question is about the behavior when a node is down, existing
node-uuid xattr will not return that UUID if a node is down. What is the
behavior with the proposed xattr?
Let me know your thoughts.
regards
Aravinda VK
On 06/20/2017 03:06 PM, Aravinda wrote:Hi Xavi,
On 06/20/2017 02:51 PM, Xavier Hernandez wrote:Hi Aravinda,
On 20/06/17 11:05, Pranith Kumar Karampuri wrote:Adding more people to get a consensus about this.
On Tue, Jun 20, 2017 at 1:49 PM, Aravinda <avishwan at redhat.com
<mailto:avishwan at redhat.com>> wrote:
regards
Aravinda VK
On 06/20/2017 01:26 PM, Xavier Hernandez wrote:
Hi Pranith,
adding gluster-devel, Kotresh and Aravinda,
On 20/06/17 09:45, Pranith Kumar Karampuri wrote:
On Tue, Jun 20, 2017 at 1:12 PM, Xavier Hernandez
<xhernandez at datalab.es <mailto:xhernandez at datalab.es>
<mailto:xhernandez at datalab.es
<mailto:xhernandez at datalab.es>>> wrote:
On 20/06/17 09:31, Pranith Kumar Karampuri wrote:
The way geo-replication works is:
On each machine, it does getxattr of node-uuid and
check if its
own uuid
is present in the list. If it is present then it
will consider
it active
otherwise it will be considered passive. With this
change we are
giving
all uuids instead of first-up subvolume. So all
machines think
they are
ACTIVE which is bad apparently. So that is the
reason. Even I
felt bad
that we are doing this change.
And what about changing the content of node-uuid to
include some
sort of hierarchy ?
for example:
a single brick:
NODE(<guid>)
AFR/EC:
AFR[2](NODE(<guid>), NODE(<guid>))
EC[3,1](NODE(<guid>), NODE(<guid>), NODE(<guid>))
DHT:
DHT[2](AFR[2](NODE(<guid>), NODE(<guid>)),
AFR[2](NODE(<guid>),
NODE(<guid>)))
This gives a lot of information that can be used to
take the
appropriate decisions.
I guess that is not backward compatible. Shall I CC
gluster-devel and
Kotresh/Aravinda?
Is the change we did backward compatible ? if we only require
the first field to be a GUID to support backward compatibility,
we can use something like this:
No. But the necessary change can be made to Geo-rep code as well if
format is changed, Since all these are built/shipped together.
Geo-rep uses node-id as follows,
list = listxattr(node-uuid)
active_node_uuids = list.split(SPACE)
active_node_flag = True if self.node_id exists in active_node_uuids
else False
How was this case solved ?
suppose we have three servers and 2 bricks in each server. A
replicated volume is created using the following command:
gluster volume create test replica 2 server1:/brick1 server2:/brick1
server2:/brick2 server3:/brick1 server3:/brick1 server1:/brick2
In this case we have three replica-sets:
* server1:/brick1 server2:/brick1
* server2:/brick2 server3:/brick1
* server3:/brick2 server2:/brick2
Old AFR implementation for node-uuid always returned the uuid of the
node of the first brick, so in this case we will get the uuid of the
three nodes because all of them are the first brick of a replica-set.
Does this mean that with this configuration all nodes are active ? Is
this a problem ? Is there any other check to avoid this situation if
it's not good ? Yes all Geo-rep workers will become Active and participate in syncing.
Since changelogs will have the same information in replica bricks this
will lead to duplicate syncing and consuming network bandwidth.
Node-uuid based Active worker is the default configuration in Geo-rep
till now, Geo-rep also has Meta Volume based syncronization for Active
worker using lock files.(Can be opted using Geo-rep configuration,
with this config node-uuid will not be used)
Kotresh proposed a solution to configure which worker to become
Active. This will give more control to Admin to choose Active workers,
This will become default configuration from 3.12
https://github.com/gluster/glusterfs/issues/244
--
Aravinda
Xavi
Bricks:
<guid>
AFR/EC:
<guid>(<guid>, <guid>)
DHT:
<guid>(<guid>(<guid>, ...), <guid>(<guid>, ...))
In this case, AFR and EC would return the same <guid> they
returned before the patch, but between '(' and ')' they put the
full list of guid's of all nodes. The first <guid> can be used
by geo-replication. The list after the first <guid> can be used
for rebalance.
Not sure if there's any user of node-uuid above DHT.
Xavi
Xavi
On Tue, Jun 20, 2017 at 12:46 PM, Xavier Hernandez
<xhernandez at datalab.es
<mailto:xhernandez at datalab.es>
<mailto:xhernandez at datalab.es
<mailto:xhernandez at datalab.es>>
<mailto:xhernandez at datalab.es
<mailto:xhernandez at datalab.es>
<mailto:xhernandez at datalab.es
<mailto:xhernandez at datalab.es>>>>
wrote:
Hi Pranith,
On 20/06/17 07:53, Pranith Kumar Karampuri
wrote:
hi Xavi,
We all made the mistake of not
sending about changing
behavior of
node-uuid xattr so that rebalance can use
multiple nodes
for doing
rebalance. Because of this on geo-rep all
the workers
are becoming
active instead of one per EC/AFR subvolume.
So we are
frantically trying
to restore the functionality of node-uuid
and introduce
a new
xattr for
the new behavior. Sunil will be sending out
a patch for
this.
Wouldn't it be better to change geo-rep
behavior
to use the
new data
? I think it's better as it's now, since it
gives more
information
to upper layers so that they can take more
accurate decisions.
Xavi
--
Pranith
--
Pranith
--
Pranith
--
Pranith
--Pranith
--Pranith
--Pranith
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.gluster.org/pipermail/gluster-devel/attachments/20170621/3de10c9e/attachment-0001.html>
More information about the Gluster-devel
mailing list