| Age | Commit message (Collapse) | Author | Files | Lines |
|---|
|
In change Ib62001c03e1e08f58cf0c6e0ba07a8879a584084 we switched the
rabbitmq queues HA mode from ha-all to ha-exactly. While this gives us a
nice performance boost with rabbitmq, it makes rabbit less resilient to
network glitches as we painfully found out via
https://bugzilla.redhat.com/show_bug.cgi?id=1441635.
Will propose another THT change to actually change the default to
-1 so we get this ha-mode:all by default.
Change-Id: I9a90e71094b8d8d58b5be0a45a2979701b0ac21c
Partial-Bug: #1686337
Co-Authored-By: Damien Ciabrini <dciabrin@redhat.com>
Co-Authored-By: John Eckersberg <jeckersb@redhat.com>
|
|
This commit implements composable HA for the pacemaker profiles.
- Everytime a pacemaker resource gets included on a node,
that node will add a node cluster property with the name of the resource
(e.g. galera-role=true)
- Add a location rule constraint to force running the resource only
on the nodes that have that property
- We also make sure that any pacemaker resource/property creation has a
predefined number of tries (20 by default). The reason for this is
that within composable HA, it might be possible to get "older CIB"
errors when another node changed the CIB while we were doing an
operation on it. Simply retrying fixes this.
- Also make sure that we use the newly introduced
pacemaker::constraint::order class instead of the older
pacemaker::constraint::base class. The former uses the push_cib()
function and hence behaves correctly in case multiple nodes try
to modify the CIB at the same time.
Change-Id: I63da4f48da14534fd76265764569e76300534472
Depends-On: Ib931adaff43dbc16220a90fb509845178d696402
Depends-On: I8d78cc1b14f0e18e034b979a826bf3cdb0878bae
Depends-On: Iba1017c33b1cd4d56a3ee8824d851b38cfdbc2d3
|
|
When we create a pacemaker resource it must happen from a single node.
If it happens from multiple nodes an immediate error will be returned by
pcs.
For the pacemaker roles we enforce this by leveraging the recently
introduced <SERVICE_NAME_bootstrap_short_node_name> which gives us
the first hostname per-service, regardless of the role.
(introduced via I03e8685f939e8ae1fcd8b16883b559615042505d)
With this approach if a pacemaker service belongs to two different
roles (say role Controller on node A and role galera on node B), it
will only create the resource from one of the two and not both (which
would return an error).
Only setting Partial-Bug for this one, because it addresses the issue
from the pacemaker resource creation POV (which is always affected). But
the issue itself is a race that we're theoretically affected by since
the composable roles work landed. While I have tried to fix the more
general case in previous attempts, I think it is best if we start a
discussion on how to fix it, because each approach has a bunch of
potential drawbacks and is quite invasive on how we do things. A
discussion slot for this has been proposed for the Atlanta PTG.
Change-Id: I662398cab60d523d204b57a5674ca8f5c0f2e68a
Partial-Bug: #1615983
|
|
When we observe the 'stop timeout' values of pacemaker resources:
rabbitmq and redis, they are set to 90s. But for all other services, it
is set to 200s.
The overcloud deployment sometimes fails due to this with the error:
Error: Could not complete shutdown of rabbitmq-clone, 1 resources
remaining
Error performing operation: Timer expired
This patch updates the timeout for Redis and RabbitMQ to avoid this
error.
Change-Id: I8a3b3951a896ee3e8e5e09778e8ea4717e76a1b4
|
|
It turns out that reducing number of rabbitmq queues in cluster
significantly improves performance of cluster especially in the case of
failover recovery time. Right now the cluster uses ha-all mode for rabbitmq
queues.
It is best to change this to "ha-exactly" mode and reduce the number
of queue copies to ceil(N/2) where N is number of controllers in the
cluster - so in typical scenario of 3 controller It would be 2 by
default.
It does not make much sense to keep the copies of queues over whole
cluster since if the quorum of nodes is lost then the rest of cluster
nodes will be stopped anyway. We let the user override this with a
parameter.
I.e. for a 3 node controlplane cluster we will go from this:
pcs resource show rabbitmq
Resource: rabbitmq (class=ocf provider=heartbeat type=rabbitmq-cluster)
Attributes: set_policy="ha-all ^(?!amq\.).* {"ha-mode":"all"}"
To this:
pcs resource show rabbitmq
Resource: rabbitmq (class=ocf provider=heartbeat type=rabbitmq-cluster)
Attributes: set_policy="ha-all ^(?!amq\.).* {"ha-mode":"exactly","ha-params":2}"
According to Marin Krcmarik's testing recovery time from failure was
reduced significantly.
Co-Authored-By: Marian Krcmarik <mkrcmari@redhat.com>
Change-Id: Ib62001c03e1e08f58cf0c6e0ba07a8879a584084
Partial-Bug: #1628998
|
|
Write restart flag file for services managed by Pacemaker into
/var/lib/tripleo/pacemaker-restarts directory. The name of the file must
match the name of the clone resource defined in pacemaker. The
post-puppet restart script will restart each service having a restart
flag file and remove those files.
This approach focuses on $pacemaker_master only (we don't want to
restart the pacemaker services 3 times when we have 3 controllers), so
it relies on the assumption that we're making the matching config
changes across the pacemaker nodes.
Change-Id: I6369ab0c82dbf3c8f21043f8aa9ab810744ddc12
|
|
As we are staring to manually check overcloud services
the first step is to check that the puppet profiles
are all aligned.
Changes applied:
No logic added or removed in this submission.
Removed unused parameters.
Align header comments structure.
All profiles parameters sorted following:
"Mandatory params first sorted alphabetically
then optional params sorted alphabetically."
Note: Following submissions will check pacemaker,
cinder, mistral and redis services in the base profiles
as some of them has the $pacemaker_master parameter
defaulted to true.
Change-Id: I2f91c3f6baa33f74b5625789eec83233179a9655
|
|
Add RabbitMQ composable role, and keep the same logic that we had in
THT.
Implements: blueprint refactor-puppet-manifests
Change-Id: I961bdbe1cc6dd1d4a315de616439f9fc77d793ae
|
Michele Baldessari
Emilien Macchi
Jiri Stransky
Carlos Camacho