A web product can be built for two years and still never touch a real server. One team did exactly that: the only deployment target was an office test machine used by two internal testers, each of whom had to edit their hosts file to reach it. The gap between writing code and shipping it produces surprises that no amount of desk testing prepares you for, and the first real deployment to a web farm is a memorable experience.
A 2 a.m. Lesson
The junior member of that team was spared the first rollout. Near the end of the first three-month fix cycle, the team leader asked whether he could come to the office at 2 a.m. on Tuesday to do a deployment. The answer was a cheerful yes — taken as a joke. It was not. By Tuesday morning the team leader and the architect were sitting together at one machine, hours into a deployment gone badly wrong.
A well-rested brain found the fault quickly: a code change the dev manager had made to cookie handling that behaved fine on the internal test server but broke the world on the real web servers.
Sleep deprivation and the confusion over who would be present fed the panic, which the outage then amplified. A deeper problem mattered more: the two seniors had no way of knowing whether a failure came from the new code or from human error in a manual process of copying zipped binaries and site files to multiple servers, editing config files by hand, and comparing and updating database schemas in the early hours.
None of that work needed to be manual. A proven automated process would have eliminated the haystack entirely. The team could have woken at 6 a.m., verified an automated release before shifting traffic to it, and fixed any issues without disturbing real users. The site would have been more stable, and expensive developers would have spent more time developing.
The Hidden Costs of the Checklist
Later years brought many more 2 a.m. deployments, breakfast paid for by the company, and a day off if the release could be proven working — assuming the drive home went safely.
The manual checklist was perpetually incomplete and outdated, yet it was never blamed for deployment-day failures. Sometimes an inaccurate checklist was the direct cause; other times it merely made diagnosis or the rollback decision far harder. Rolling back meant doing it all again the next day, and the sleep debt kept compounding.
Web development can tempt teams further than other work. An internship writing Windows installers meant shipping code to machines the developer could not access, so the process had to be right. On the web, updating production straight from a dev machine looks like the easy fix, especially when someone wants it immediately — which is why automating deployment should come with locking things down so automation is the only path to production.
Frank and the Rewards of Heroics
As responsibilities grew, research turned up material that took direct aim at manual deployers, including Damian Brady's talk “Octopus Deploy and how to stop deploying like an idiot”, and Octopus founder Paul Stovell's line on databases: keep yours out of source control and you don't deserve one — go use Excel. The tone recalls Australia's road-safety campaign “If you drink then drive, you're a bloody idiot,” which scared people straight by insulting a destructive choice.
Brady's talk targets a hypothetical deployment manager named Frank who keeps being a hero by adding risk and wasted time to a process that could be automated — and an automated process would not overwrite the config file.
Some Franks are worse than idiots. The “deployment manager” is often a cunning operator dressed as a harmless one: engineering hires tend to have above-average IQs, and the person in that role will have searched for easier options and chosen not to take them. The tradeoff is that Frank seems important, that other developers look foolish trying to follow his process while he is on leave, and that he alone gets the trips to clients because he can get the product working on a new server. Companies should be careful which behaviors they reward — and Conway's law applies to deployment processes too.
Lockdown, Approvals, and Other Motivations
Conway's observation is why Octopus Deploy has built-in manual intervention and approval steps; some hierarchy and division of responsibility in a process is normal and necessary. Other reasons to keep deployments manual are nonsense. Complex manual deployments remain more widespread than they need to be, and for developers still living that life in similar conditions, the past is a place to visit, not to remain.
There is an argument for the team-building value of a 2 a.m. deployment, much as a military deployment may teach troops valuable lessons even when the stated cause isn't the real one and the costs exceed expectations. A more pleasant reference: Adobe's San Jose “Photoshop floor,” with time-capsule conference rooms including a 90s room housing a working Macintosh Classic running Photoshop 1.0.
Choosing Your Point on the Automation Dial
Automation is a spectrum. On one side is heavy control for a complicated automatic process; on the other is the person who runs Can I Use, for whom adopting git-ftp was a life upgrade over manually copying modified files to a web server — process left simple, little energy spent on sophistication. Bitbucket Pipelines and GitHub Actions sit in the middle: more automated than git-ftp from a dev machine, less complicated than Octopus together with TeamCity, which is overkill on a simple project.
The point on that spectrum is a function of the tool and the reward it brings, measured as reduced complexity versus added complexity. The 2 a.m. adventure above fits whenever a team manually deploys a complicated site with database changes, background processes, web farms, and SLAs.
Read the reward in the negative: a tool costs more than a project can afford. Therefore, the matching rule applies to automators too. A solo developer with a side project who wires Azure DevOps to push into Octopus Deploy pays for complexity that never returns its value in money or cognitive load. Octopus is a deceptively complex tool, and those go-to-market tradeoffs can be invisible from outside.
So the whole comparison is contingent on the reward calculation. Choose the simpler route unless genuine complexity, a lack of complexity, and the middle ground call for it.
The Costs of Manual Work
A solo developer with a side project who hooks Azure DevOps up to push into Octopus Deploy pays too much in money and cognitive load.
Certainty matters more when the processes are shared. Whether an all-in-one platform such as Jmix can win out over other technologies may end up unknown.
The Octopus case also shows that using the tool doesn't end complexity, it only constrains and possibly improves it. On-demand instant application images come with their own unpredictability: after testing Firebase timing, for instance, a teammate noticed a change that would cause a repeatable lag.
A Car Driven by a Computer Is Still a Car
“Serverless” is a misnomer, and how much the inaccuracy matters is debatable. The analogy: early cars earned the name “horseless carriages” because the carriage could move without a horse — a true paradigm shift. Cars without human drivers should not be called driverless, since the driver is now an AI; self-driving is the better term. Self-driving vehicles often work well, but ignoring the limits of how they work can be fatal.
One hobby project — a rap battle chat on Firebase — needed a cloud function to score each post with the same rhyme detection algorithm that powered the front end. It worked in testing under the Cloud Functions emulator. After the first deployment it fell short in performance, probably from a cold start, since cost was missing and adding instances was a solution. Much as years before, code behaved dramatically differently on a dev machine than on real Firebase — as though a server still existed while offering little ability to tweak it. Fixing one thing only raised the price of another.
Remove certainty and options remain ones that are variable priced.
Rainbows End, a sci-fi novel by Vernor Vinge, makes the parallel explicit. The curmudgeonly main character cuts open a car that wasn't designed to be serviced, finding every component labeled a no user-serviceable part, and he is told the car is “Russian dolls all the way down.” The choice pointed at him cuts down the cost and effort. Even by those numbers and hard engineering, you still find those moving parts wanting.
Which brings the problem home in a smaller way: an article is a deployment artifact too, and the tools available suit different costs. A blog at one shop runs on a poll on the company's internal social network, in which the new post waits in a markdown file and no preview step exists — the input may come from a newsletter with the write-up following, and the choices fork. A prompt two months before tackling the author-side sequence suggests that the answer to "How can Jon create such a deep, lyrical song from the same formula he relies on every time?" is known. The other arrangement has the editor access work she points on but nothing needs a static generator by itself.



