A ctlplne studio product
22 October 2026 · 3 min read

Renewal is not deployment

Every certificate outage post-mortem has the same sentence in it: the renewal ran. It did. The file was on disk. The load balancer never reloaded, the keystore was never rebuilt, and the CDN was still holding a copy from last year.

Renewal automationDiscovery and inventoryThe 47-day schedule

Two different jobs

Renewal is a conversation with an authority: prove the name, get a new certificate. Deployment is a conversation with everything that serves that name: give it the new certificate, make it use it, confirm it did. Automating the first without the second produces the most confusing outage there is, because every dashboard says renewed.

A renewed certificate fanning out to five kinds of target, each of which takes it differently renewed certificate on disk, in the control plane web server or proxycloud certificate storeappliance APIkeystore inside an applicationKubernetes secret write, validate, reloadimport via the provider APIbind to the SSL profilerebuild keystore, restartupdate secret, roll pods
Five shapes of target. The verb on the right is the part the renewal never did.

The five places

  1. Web servers and proxies. nginx, Apache, Caddy, HAProxy, IIS, Traefik. The pattern is write the certificate and key (or a PKCS#12 bundle), validate the configuration, then reload or update the dynamic file provider. Skip the reload and the process keeps the old certificate in memory until someone restarts it, weeks later.1
  2. Cloud certificate stores. AWS Certificate Manager, Azure Key Vault, Google Certificate Manager. There is no file; the certificate is imported through the provider's API and the load balancer or CDN in front picks it up from there. Each provider has its own idea of when.1
  3. Appliances. F5 BIG-IP, Citrix NetScaler, Cisco ASA and ISE, Fortinet, Palo Alto. They speak HTTPS management APIs, and the job is not "upload" but "bind this certificate to that SSL profile or listener".1 The old certificate stays bound until you change the binding.
  4. Keystores inside applications. Java KeyStore and PKCS#12 files that a JVM reads once at startup. The keystore has to be rebuilt deterministically and the application told to reload it or restarted.1 Mail servers (Postfix, Dovecot) and Envoy's secret discovery follow their own variants.
  5. Kubernetes. A TLS secret updated in place, then whatever reads it (an ingress controller, a sidecar, a pod that mounted it) has to notice. In-cluster automation handles this well for in-cluster certificates. It knows nothing about the four places above.

The check

Deployment is done when something that is not the deployer connects to the endpoint and sees the new fingerprint. Not "the API returned 200", not "the file changed": a TLS handshake from outside, compared against the certificate you meant to install. Do it after every deployment, and keep doing it on a schedule, because the old certificate has a way of coming back after a failover, a config restore, or an image rebuild from last quarter's snapshot.

What this looks like at twelve renewals a year

At one renewal a year, a person could do the five conversations by hand and remember which appliance needed the binding step. At twelve per certificate, deployment has to be a connector: code that knows the shape of each target, runs with the least access that shape needs, and records what it did. trstctl ships connectors for the first four kinds of target, running in a sandbox that declares exactly which host it may dial, and records each deployment as an event beside the issuance it belongs to. For Kubernetes, its in-cluster agent writes the TLS secret. The appliance and cloud connectors are tested against stand-ins for the vendor APIs so far, not against real devices.1,2

Sources

  1. trstctl documentation, Deployment connectors: the initial connector set (web servers, cloud certificate stores, keystores, Envoy, mail servers, F5) and the appliance connectors (Citrix, Cisco, Fortinet, Palo Alto), each declaring only the network access it needs.
  2. trstctl documentation, Compliance: every served mutation is an event in a hash-linked, tenant-scoped audit log.