When/how does Zero-Downtime-Failover actually fail-over?

[c.anselm]

Hi there,

I’ve been reading up on the exact functionality of the zero-downtime-failover (ZDF) feature.

According to https://support.cloudflare.com/hc/en-us/articles/360037465932-Preventing-site-downtime#6epLHd3oJb13iY9R06WPjN there is a limitation that under certain circumstances the failover only retries once:

Cloudflare currently retries only once for HTTP 521, 522, and 523 response codes.

Though the meaining of this sentence is not really clear to me.

Does this mean that CF will only perform ZDF and only do it for one other origin if one of these three response codes is received from the origin server? If so this would seem strange as those response codes should only be used by proxies and not by applications.

Or does this mean that CF will perform ZDF on any Status 5xx, but only do it once if 521-523 is received from origin?

What’s the exact mechanic here? When will ZDF be performed and when will CF give up and serve an error page?

Also does ZDF update the “healthy”-state of an origin?

[simon]

Hi there, I think this topic here should address most of your questions - have a read and let us know:

Zero-downtime-failover question

I had a question about the zero-downtime-failover feature included with business plans. We currently have an A record registered with a single IP address. In the DNS UI, this record appears as ‘Proxied’. I would like to add a second IP to the same domain, that points to a duplicate k8s cluster hosting identical content as the original cluster. The intention is that both clusters will serve content simultaneously, and if one goes down, Cloudflare will stop sending requests to that cluster. It sou…

[c.anselm]

So what you’re saying is that ZDF will not attempt to failover at all when an origin responds with HTTP status 503 Service Unavailable?

[simon]

That is correct - you could use Load Balancing to achieve something like this as an alternative - since there will be proactive health-checks defined.

[c.anselm]

… which would still cause some requests to receive the 503 until the monitor picks up the changed status on the health-check endpoint

I guess we’ll have to workaround this limitation from our side then, thanks for the clarification.

[simon]

It’s good feedback - would you create a topic in https://community.cloudflare.com/c/feedback/prodreq/30 detailing what you’d like? Others can comment/upvote and I’ll loop in our product team to take a look.