Skip to content

BOSH Disaster Recovery Process for AZ Outages

In it's current state, BOSH, not being a highly available service, does not support a multi-AZ deployment. This means that in the event of an AZ outage, it is not directly possible to fallback to another replica instance for example. This has not always been necessary since BOSH is a deploytime-relevant service, and has less relevance during runtime. This, however, does not mean that large productive environments depending on BOSH for their deployments could do without BOSH in scenarios where AZ outages extend over longer time periods, since, for instance, relevant fixes and security updates should still be delivered in time.

However, it is possible that we externalize the state to the best possible extent to highly available infrastructure offerings, and fallback to another AZ in case of an AZ outage. This can be achieved if certain prerequisites are fulfilled.


1. Using domain names

When an AZ outage happens, and we switch to a new director, it is important that BOSH's deployments are still able to reach the new director in the other AZ. The first step to make this posible is to use domain names for the director instead of IPs. This, however, does come with a challenge of updating the the certificates used by BOSH with alternative names (IPs being used to connect to the director and new hostname configured) for hostname validation.

2. Externalizing the database

BOSH should still be aware of its deployments, their instances, blob references etc. in the event of its redeployment to another AZ. This could be made possible by using an external database instead of the default colocated configuration. This can be done by consuming highly available RDS services by IaaS providers, and then configuring the director database properties and the UAA database properties to consume these external database instances. See director db configuration and uaa configuration for further help in this regard.

3. Externalizing the blobstore

It would also make sense to use an external blobstore like AWS S3 which is highly available than consuming blobs from the colocated DAV blobstore. This would help to fail over to a new director with less worries of re-fetching the blobs (keeping in mind this is a time consuming process) when the need arises. As of now, it is already possible to use highly available blobstores provided by AWS, Azure, AliCloud and GCP. Switching over to a new blobstore can be smooth if there's a blobstore migration mechanism implemented using the existing BOSH Blobstore CLIs (by simply 'putting' the blobs using the s3cli from the director VM to the S3 bucket for example in case of AWS). Another way of doing this would be to recreate the blobstore externally by uploading the blobs with --fix or running the deploy commands with --fix-releases. This however is costly in terms of time required to re-fetch the blobs. See this to start configuring the director to use an external blobstore.

Steps for recovering a bosh director in the event of an AZ outage:

We can assume that we are working with multiple deployments across two availability zones: zone_1 and zone_2, and that our director is currently deployed in zone_1. Then, let's also assume that there's a sudden outage in zone_1.

1. Isolating the director

To isolate the director in zone_1 from the database to prevent it from modifying the state when the zone is up, it is important that:\ a. The database passwords are rotated\ b. All existing database connections are killed via psql (by executing pg_terminate_backend for example)

2. Deploying another director

There might be differences in deploying BOSH directors in another AZ based on whether we use the create-env approach or the bosh deploy approach.

After isolating the director deployed by create-env in zone_1 the process to deploy a second one in zone_2 can be treated similar to a scratch installation, implying that the zone used in the director's deployment manifest is changed to zone_2 and the bosh-state.json is dropped before calling create-env and setting up the second director in zone_2.

For directors deployed by bosh deploy, the deployment manifest should be updated to usezone_2. Next, the number of instances in the instance_group has to be increased to 2 and a second static ip (available in zone_2) should be configured (in order to prevent data mismatch in the records of the deployer of this director and therefore to support both directors in the two zones).

However, while redeploying the director with the updated manifest, the director which this drector is the deployment of would first try to delete the old bosh in zone_1. This will most probably fail in case of a zone outage. This can be prevented by invoking ignore on the deployed bosh in zone_1.

3. Disabling resurrection

After bosh is successfully deployed in zone_2 it will now take care for repairing its deployments. Most of them would be in error situation since they would have instances in the failing zone. The director would continuously trigger scan_and_fix tasks which would then also fail. Therefore, an option would be to deploy the new directors with disabled resurrection. After the zone outage is over, the directors in zone_1 can be dropped and the one in zone_2 can be used further. Of course with enabled resurrection. Resurrection can be enabled/disabled via bosh update-resurrection on (or off).

>Here< is a video presentation on the topic from the 2023 CF Day at Heidelberg.