This patch updates the Octavia documentation in support of the OpenStack documentation migration[1]. [1] https://specs.openstack.org/openstack/docs-specs/specs \ /pike/os-manuals-migration.html Change-Id: I97fd038b8050bfe776c3fca8336d9090f8236362 Depends-On: Ia750cb049c0f53a234ea70ce1f2bbbb7a2aa9454
16 KiB
Operator Maintenance Guide
This document is intended for operators. For a developer guide see
the ../../contributor/guides/dev-quick-start
in this
documentation repository. For an end-user guide, please see the ../../user/guides/basic-cookbook
in this documentation
repository.
Monitoring
Monitoring Load Balancer Amphora
Octavia will monitor the load balancing amphorae itself and initiate failovers and/or replacements if they malfunction. Therefore, most installations won't need to monitor the amphorae running the load balancer.
Octavia will log each failover to the corresponding health manager logs. It is advisable to use log analytics to monitor failover trends to notice problems in the OpenStack installation early. We have seen neutron (network) connectivity issues, Denial of Service attacks, and nova (compute) malfunctions lead to a higher than normal failover rate. Alternatively, the monitoring of the other services showed problems as well, so depending on your overall monitoring strategy this might be optional.
If additional monitoring is necessary, review the corresponding calls
on the amphora agent REST interface (see ../../contributor/api/haproxy-amphora-api
)
Monitoring Pool Members
Octavia will use the health information from the underlying load balancing application to determine the health of members. This information will be streamed to the Octavia database and made available via the status tree or other API methods. For critical applications we recommend to poll this information in regular intervals.
Monitoring load balancer functionality
For production sites we recommend to use outside monitoring services. They will use servers distributed around the globe to not only monitor if the site is up but also parts of the system outside the visibility of Octavia like routers, network connectivity, etc.
Monitoring Octavia Control Plane
To monitor the Octavia control plane we recommend process monitoring of the main Octavia processes:
- octavia-api
- octavia-worker
- octavia-health-manager
- octavia-housekeeping
The Monasca project has a plugin for such monitoring (see Monasca Octavia plugin). Please refer to this project for further information.
Octavia's control plane components are shared nothing and can be scaled lineary. For high availability of the control plane we recommend to run at least one set of components in each availability zone. Furthermore, the octavia-api endpoint could be behind a load balancer or other HA technology. That said, if one or more components fail the system will still be available (though potentially degraded). For instance if you have installed one set of components in three availability zones even if you lose a whole zone Octavia will still be responsive and available - only if you lose the Octavia control plane in all three zones will the service be unavailable. Please note this only addresses control plane availability; the availability of the load balancing function depends highly on the chosen topology and the anti-affinity settings. See our forthcoming HA guide for more details.
Additionally, we recommend to monitor the Octavia API endpoint(s). There currently is no special url to use so just polling the root URL in regular intervals is sufficient.
There is a host of information in the log files which can be used for log analytics. A few examples of what could be monitored are:
- Amphora Build Rate - to determine load of the system
- Amphora Build Time - to determine how long it takes to build an amphora
- Failures/Errors - to be notified of system problems early
Rotating the Amphora Images
Octavia will start load balancers with a pre-built image which contain the amphora agent, a load balancing application, and are seeded with cryptographic certificates through the config drive at start up.
Rotating the image means making a load balancer amphora running with an old image failover to an amphora with a new image. This should be without any measurable interruption in the load balancing functionality when using ACTIVE/STANDBY topology. Standalone load balancers might experience a short outage.
Here are some reasons you might need to rotate the amphora image:
- There has been a (security) update to the underlying operating system
- You want to deploy a new version of the amphora agent or haproxy
- The cryptographic certificates and/or keys on the amphora have been compromised.
- Though not related to rotating images, this procedure might be invoked if you are switching to a different flavor for the underlying virtual machine.
Preparing a New Amphora Image
To prepare a new amphora image you will need to use diskimage-create.sh as described in the README in the diskimage-create directory.
For instance, in the octavia/diskimage-create
directory,
run:
./diskimage-create.sh
Once you have created a new image you will need to upload it into glance. The following shows how to do this if you have set the image tag in the Octavia configuration file. Make sure to use a user with the same tenant as the Octavia service account:
openstack image create --file amphora-x64-haproxy.qcow2 \ --tag <amphora-image-tag> --private \ --disk-format qcow2 --container-format bare /var/lib/octavia/amphora-x64-haproxy.qcow2
If you didn't configure image tags and instead configured an image id, you will need to update the Octavia configuration file with the new id and restart the Octavia services (except octavia-api).
Generating a List of Amphorae to Rotate
The easiest way to generate a list is to use nova to list all the amphorae:
openstack server list --name amphora* --all -c ID -c Status -c Networks
Take note of the amphorae IDs and IPs on the management network.
If you are using an ACTIVE-STANDBY topology it might be beneficial to rotate first the BACKUP amphora before the ACTIVE one. In this case you will need to use the Octavia database to query for that:
mysql octavia -e 'select id, compute_id, lb_network_ip, role from amphora where status="ALLOCATED" or status="READY";'
Take note of the compute ids, the role (ACTIVE, BACKUP), and the ip
on the management network. You can also find this IP either via
openstack server show <id>
or
openstack server list --all -c ID -c Status -c Networks | grep <id>
.
Rotate an Amphora
The idea is to force Octavia to start up a new amphora because Octavia thinks this one has failed. The most graceful way to do this is to shut down the port on the management network.
Use the ip on the management newtork to find the port-id:
openstack port list | grep <ip on mgmt net>
Take note of the port-id and initiate a failover by shutting down this port:
openstack port update --admin-state-up False <port-id>
You can observe the failover by querying nova
openstack server list --all | grep <id>
until the
server isn't found any longer.
Best Practices/Optimizations
To speed up the failovers, the spare pool can be temporarily
increased to accommodate the rapid failover of the amphora. In this case
after the new image has been loaded into glance, shut down or initiate a
failover of the amphora in the spare pool. They can be found, for
instance, by looking for the servers in
openstack server list --all
who only have an ip on the
management network assigned but not any tenant network. Alternatively,
use this database query:
mysql octavia -e 'select id, compute_id, lb_network_ip from amphora where status="READY";'
After you have increased the spare pool size and restarted all Octavia services, failovers will be greatly accelerated. To preserve resources, restore the old settings and restart the Octavia services. Since Octavia won't terminate superfluous spare amphora on its own, they can be left in the system and will automatically be used up as new load balancers are created and/or load balancers in error state are failed over.
Warning
If you are using the anti-affinity feature please be aware that it is not compatible with spare pools and you are risking both the ACTIVE and BACKUP amphora being scheduled on the same host. It is recommended to not increase the spare pool during fail overs in this case (and not to use the spare pool at all).
Since a failover puts significant load on the OpenStack installation by creating new virtual machines and ports, it should either be done at a very slow pace, during a time with little load, or with the right thottling enabled in Octavia. The throttling will make sure to prioritize failovers higher than other operations and depending on how many failovers are initiated this might crowd out other operations.
Note
In Pike a failover command is being added to the API which allows to failover a load balancer's amphora while taking care of the intricacies of different topologies and prioritizes administrative failovers behind other operations. This function should be used instead of the ones described above once it becomes available.
Rotating Cryptographic Certificates
Octavia secures the communication between the amphora agent and the control plane with two-way SSL encryption. To accomplish that, several certificates are distributed in the system:
- Control plane:
- Amphora certificate authority (CA) certificate: Used to validate amphora certificates if Octavia acts as a Certificate Authority to issue new amphora certificates
- Client certificate: Used to authenticate with the amphora
- Amphora:
- Client CA certificate: Used to validate control plane client certificate
- Amphora certificate: Presented to control plane processes to prove amphora identity.
The heartbeat UDP packets emitted from the amphora are secured with a symmetric encryption key. This is set by the configuration option heartbeat_key in the health_manager section. We recommend setting it to a random string of a sufficient length.
Rotating Amphora Certificates
For the server part Octavia will either act as a certificate
authority itself, or use ../Anchor
to issue amphora certificates to be used by
each amphora. Octavia will also monitor those certificates and refresh
them before they expire.
There are three ways to initiate a rotation manually:
- Change the expiration date of the certificate in the database.
Octavia will then rotate the amphora certificates with newly issued
ones. This requires the following:
- Client CA certificate hasn't expired or the corresponding client certificate on the control plane hasn't been issued by a different client CA (in case the authority was compromised)
- The Amphora CA certificate on the control plane didn't change in any way which jeopardizes validation of the amphora certificate (e.g. the certificate was reissued with a new private/public key)
- If the amphora CA changed in a way which jeopardizes validation of
the amphora certificate an operator can manually upload newly issued
amphora certificates by switching off validation of the old amphora
certificate. This requires a client certificate which can be validated
by the client CA file on the amphora. Refer to
../../contributor/api/haproxy-amphora-api
for more details. - If the client certificate on the control plane changed in a way that
it can't be validated by the client certificate authority certificate on
the amphora, a failover (see
rotate-amphora-certs
) of all amphorae needs to be initiated. Until the failover is completed the amphorae can't be controlled by the control plane.
Rotating the Certificate Authority Certificates
If there is a compromise of the certificate authorities' certificates, or they expired, new ones need to be installed into the system. If Octavia is not acting as the certificate authority only the certificate authority's cert needs to be changed in the system so amphora can be authenticated again.
# Issue new certificates (see the script in the bin folder of Octavia if Octavia is acting as the certificate authority) or follow the instructions of the third-party certificate authority. Copy the certificate and the private key (if Octavia acts as a certificate authority) where Octavia can find them.
# If the previous certificate files haven't been overridden, adjust the paths to the new certs in the configuration file and restart all Octavia services (except octavia-api).
# Review rotate-amphora-certs
above to determine if and how the
amphora certificates needs to be rotated.
Rotating Client Certificates
If the client certificates expired new ones need to be issued and installed on the system:
# Issue a new client certificate (see the script in the bin folder of Octavia if self signed certificates are used) or use the ones provided to you by your certificate authority.
# Copy the new cert where Octavia can find it.
# If the previous certificate files haven't been overridden, adjust the paths to the new certs in the configuration file. In all cases restart all Octavia services except octavia-api.
If the client CA certificate has been replaced in addition to rotating the client certificate the new client CA certificate needs to be installed in the system. After that initiate a failover of all amphorae to distribute the new client CA cert. Until the failover is completed the amphorae can't be controlled by the control plane.
Changing The Heartbeat Encryption Key
Special caution needs to be taken to replace the heartbeat encryption key. Once this is changed Octavia can't read any heartbeats and will assume all amphora are in an error state and initiate an immediate failover.
In preparation, read the chapter on best_practice
in the Failover section. In particular,
it is advisable if the throttling enhancement (available in Pike)
doesn't exist to create a sufficient number of spare amphorae to
mitigate the stress on the OpenStack installation when Octavia starts to
replace all amphora immediately.
Given the risks involved with changing this key it should not be changed during routine maintenance but only when a compromise is strongly suspected.
Note
For future versions of Octavia an "update amphora" API is planned which will allow this key to be changed without failover. At that time there would be a procedure to halt health monitoring while the keys are rotated and then resume health monitoring.