Is the vision in your head of your pipelines that of lean, graceful atheletes? Do branch builds simply test your changes swiftly and anything that hits master builds artifacts finished with the flourish of a ephemeral "deployment"?
Your pipelines are overweight slobs, unwilling to truly do real work.
Conventional wisdom dictates that deployments occur at the ends of pipelines by
running a simple task, say
kubectl apply or similar, with the produced
artifact mentioned. This act is transient and for many pipelines means rolling
back is an act of rerunning the whole pipeline, an individual step in the
pipeline, or even reverse-engineering the action in the deployment step and
performing it manually, given the level of desperation.
Build artifacts aren't deployments. By turning deployments from "transient action" into their own artifact you can scrobble across deployments with little fuss. A deployment artifact can be anything that describes the act of deploying. This might be a script, a set of versions packaged together, or even a specification like a kubernetes manifest. Once you have release artifacts and deployment artifacts start the exercise regime for your pipelines by building and publishing all the things.
Won't all this extra work cost more money and time? The reality is that amortizing the cost of storing your artifacts and building whenever you get a chance helps provide options so you don't have to do extra work when it is the most untimely to do so. What costs more? Having a terrible mean-time-to-resolution (MTRR) and frequent outages or paying for more build bots and storage space? If you haven't learned the cost of burning the trust of your end users, then you have an important lesson to learn.
Scrobbling deployments not only helps reduce the blast radius of botched code hitting the pool of production by increasing your MTRR but it also gives you the opportunity for functionality such as preview deployments. Some approaches may provide previews in different deployment environments entirely whereas others allow service or resource "naming" (e.g. unique URLs or distinct IP addresses) to route traffic accordingly. Some blend the two together. This last approach is often how services such as zeit and linc.sh do their previews for branch builds. It depends on how much reproducibility you care about to get a sanity check before deploying to production.
The one wrench in all of this is the matter of shared state; sometimes the complication of going backwards or forwards from a certain deployment involves running or reversing migrations, reinstating or removing coupled infrastructural changes, or even having third party services paid and available. There are islands of deployments that may become totally inaccessible due to the above and the best advice I can provide in the briefest period of time is that all of these can be (somewhat) circumvented by ensuring nothing exists that isn't code. To address the noted issues (which is incomplete, mind you):
Fear not the Friday deploy when you have options at hand; fear the duct-tape and popsicle-stick infrastructure that makes Friday deploys a nightmare.