Lesson #1: Don't push code on Friday afternoon.
Lesson #2: Beer, Code and Commit is totally fine. Just don't push! Wait until next day to review and push/deploy it..
As a general rule, the best time to launch is first thing on a Tuesday. Why Tuesday morning? We are also well aware of "Monday morning madness" too - pre-planning stress for Monday morning is akin to masochism.
The technical staff all talk about the dangers of Friday evening deployments.
The company has ignored/overridden these warnings in the past, and sometimes experienced catastrophic failures because of it.
STILL, sometimes the marketing staff will have weekend promotions that are considered more important than these concerns.
In particular, a weekend promotion will be implemented, tested on Wednesday, and ready to go. And only after it's active, is a bug found in it on Friday at 5pm. And the promo is more important than the risk of failure.
I'm not saying this as a "bitter techie"... I'm just explaining that this is how it happens.
I think that the "it's more important" argument is often correct. Not running the promo costs the company $X of opportunity cost, and X is often very well known, from past promotions. Balance "losing $X" against "well, there's only a small chance of something going wrong", and that's why the risk is often taken.
Take the risk enough times, and eventually something goes wrong.
We exclusively push code on Fridays.
Funnily enough, change control actually runs a lot more smoothly in terms of getting past senior non-technical managers when you include elements like that
continuous deployment to a dev/test environment is the easiest for most organisations to move to. Due to the live environment being mission-critical, they can't afford the risk of any degradation of service. So you push regularly (after passing test suites) to a test environment, get a small subset of users working, and at some point then push out to live from that. But I suspect that isn't what you are referring to, as this is too similar to typical change management.
the alternative definition of continuous deployment is that of constantly pushing to live, initially for a subset of users then rolling it out from there gradually, but always on the live environment. In many large, often 'cloud-y' solutions, that subset might be all users. Except you can't have a public transport ticketing system fail at peak times. A hospital patient record system must stay available for staff, and give sufficient notice before any possible impact to service to allow manual processes to be used. Payroll, accounts, HR systems... all of these have failure modes that have financial penalties at best.
Hence continuous deployment only makes sense when you can afford to risk service, possibly with significant impact.
> Don't code when you're distracted.
Same story here, I can't remember the exact scenario but I was concurrently acting under all three of my titles (Developer, Architect, Escalation Engineer/Critical Debugger). The customer (who was 7 hours different to us) had been battling for 3 months and we were getting nowhere (all thanks to a, since fixed, bug in WinDBG which essentially came down to broken stack traces in certain scenarios), for those 3 months I had been working 20 hour shifts (development by day, support by night).
Under that strain I eventually made a screw up with the dev and it cost QA time. The MD of my region had a sit down with me and I
> cried literal tears
Needless to say they were impressed that I someone cared so much and sent me home to sleep. The next day I came in and decided to go through the 800kloc codebase line-by-line and see what could be causing the issue - I found it in a few hours.
Lesson #4: Get sleep.
Who ships a product release immediately after a commit?
Turns out? Nope.
I also find that the arguments I had with my bosses come around again, from my employees.
I still don't really understand how business works - but I've learned enough to understand that I don't understand how business works.
My question is: Who never makes mistakes? I certainly do not belong in that set.
And I like Patrick's candid business anecdotes. Snarky comments like this might disencourage people from writing useful advice here. Which would be very unfortunate.
this is the primary difference I see between programmers and sysadmins, development and operations. I know programmers who don't own cellphones at all, while I know some sysadmins who take tertiary backup communication devices on vacation.
It's a difference in focus.
Of course, most programming jobs have /some/ operational responsibilities, and most sysadmin jobs have /some/ development responsibilities, but most people see themselves as primarily one or the other, and act accordingly.
Excuse me for caviling at your post, but "fortuitously" is a synonym for "accidentally", not "fortunately".
Thanks for the comment though - I personally always appreciate a gracious correction.
Lesson #1: Don't code when you're distracted.
Some hours later, the problem manifested. The queue workers came down, and AR (which is totally dependent on them for its core functionality) immediately stopped doing the thing customers pay me money to do. My monitoring system picked up on this and attempted to call me -- which would have worked great, except my cell phone was in a box that wasn't unpacked yet.
Lesson #2a: If you're running something mission critical, and your only way to recover from failure means you have to wake up when the phone rings, make sure that phone stays on and by you.
Later that evening I felt a feeling of vague unease about my change earlier and checked my email from my iPad. My inbox was full of furious customers who were observing, correctly, that I was 8 hours into an outage. Oh dear. I ssh'ed in from the iPad, reverted my last commit, and restarted the queue workers. Queues quickly went down to zero. Problem solved right?
Lesson #3: If at all possible, avoid having to resolve problems when exhausted/distracted. If you absolutely must do it, spend ten extra minutes to make sure you actually understand what went wrong, what your recovery plan is, and how that recovery plan will interact with what went wrong first.
AR didn't use idempotent queues (Lesson #4: Always use idempotent queues), so during the outage, every 5 minutes on a cron job every person who was supposed to be contacted that day got one reminder added to the queue. Fortuitously, AR didn't have all that many customers at the time, so only 15 or so people were affected. Less than fortuitously, those 15 folks had 10 to 100 messages queued, each. As soon as I pressed queues.restart() AR delivered all of those phone calls, text messages, and emails. At once.
Very few residential phone systems or cell phones respond in a customer-pleasing manner to 40 simultaneous telephone calls. It was a total DDOS on my customers' customers.
I got that news at 3 AM in the morning Japan time, at my new apartment, which didn't have Internet sufficient to run my laptop and development environment to see e.g. whose phones I had just blown up. Ogaki has neither Internet cafes nor taxis available at 3 AM in the morning. As a result, I had to put my laptop in a bag and walk across town, in the freezing rain, to get back to my old apartment, which still had a working Internet connection.
By the time I had completed the walk of shame I was drenched, miserable, and had magnified the likely impact that this had on customers' customers in my own mind. Then I got to my old apartment and checked email. The first one was, as you might expect, rather irate. And I just lost it. Broke down in tears. Cried for a good ten minutes. Called my father to explain what had happened, because I knew that I had to start making apology calls and wasn't sure prior to talking to him that I'd be able to do it without my voice breaking.
The end result? Lost two customers, regained one because he was impressed by my apology. The end users were mostly satisfied with my apologies. (It took me about two hours on the phone, as many of them had turned off their phones when they blew up.)
You'd need a magnifying glass to detect it ever happened, looking on any chart of interest to me. The software got modestly better after I spent a solid two weeks on improved fault tolerance and monitoring.
Lesson the last: It's just a job/business. The bad days are usually a lot less important in hindsight than they seem in the moment.