I work on services too. Typically you don’t want to revert the entire PR, just the bit that turns it ‘on’. If you make this bit as small as possible, it’s often simpler and easier to inspect the correctness of reverting a flag flip or a config option change than it is to revert actual code/config. Reverting a ton of code is what you do when you don’t need it anymore. It seems like a bad idea to do it in a hurry (unless you have no choice, but that’s a separate point about designing features that have this property and making it a strict requirement during code review where possible).

Then again, reverting commits in your source tree seems like a poor fix to the problem in any form. What you really want is rolling back to the last known working version (which you hopefully can do without needing source tree changes), rather than revert one of n changes, that may depend on each other and rolling forward.

Especially if you are building a new feature that is riskier or bigger, you will put the main bit of code behind a flag/config option (or perhaps even multiple to incrementally test larger and larger bits).

Sure. I'm just talking about binary deploy. Of course you're gonna use flags to flight things in an industry context, depending on the company and its CI/CD practices.

Let's say the binary deploy (before flighting) is fucked somehow. Then what do you do?

Well, if rollbacks (which shouldn’t require affecting the source tree) are a thing then that. Otherwise you’re right and code needs to be reverted and hope it works.

But in the 2nd case, I’d make sure to increase the priority of having known good rollback versions available (and rollbacks performable) and also carefully consider what CI/CD could be added to catch more broken binaries (e.g canary or staging if it’s important enough) and what code review practices could have prevented it.

Ok I agree, you roll back to the known working version. The easiest way to do that is revert the whole PR (or data deploy in case of flags, ofc). My point is not "flags vs. no flags". My point is "each PR should generate one commit because that's easy to revert".

The commit dag of git is a cool feature but shouldn't be in `main`. It's so much easier to work with a linear history and one where each commit contains all the required context to figure out "could this have broken something".

I don’t think reverting the broken code should be on the critical mitigation path at all.

Imagine a scenario where a bug didn’t immediately cause an issue or where your release contains more than 1 new PR. If you suspect the latest version of the binary is broken, your first instinct should be to use a version that isn’t. Figuring out the change and rolling it back should come after the rollback, when you have more time to think.

Deciding whether to revert a change is tactical question. Often the issue will be because you tickled an unknown bug in a different part of the code. In that case, it’s a lot easier to fix forward than revert the code that tickled the big and go through the multiple steps of fixing the bug and redoing.

Linear history is good, and having multiple commits in a PR doesn’t prevent it. The only change is adding n (ideally well crafted) consecutive commits rather than 1.

The problem is who is doing the crafting. (And the approving.) At least squash based approach limits to 1 the number of commits an untrustworthy "crafter" can occupy.

I think most code review already requires good faith on behalf of the reviewer already.

I do see what you mean about untrustworthy crafter. If we want to preserve master history, then the damage of a bad commit chain is worse than of bad code (which can be fixed/undone).

However, I think that the truly adversarial case is rare (and an exception could be made and master history be rewritten in that case). In most cases though, hopefully your coworkers and not deliberately trying to sabotage the codebase :). And I don’t think the commit chain needs to be a work of art or anything, just mainly avoiding typo commits and similar, so it shouldn’t be difficult to do when approached in good faith.

> the truly adversarial case is rare

The problem is not adversarial, or due to malice. The problem is ignorance and expediency driven by a desire to push code and little incentive to cleanup your git history.

The easy fix is to squash PRs.

The hard fix is to enforce that devs become "crafters" and to define what is and isn't "good faith".

Ratorx, do we work on the same team at google? I feel like I've heard you before?

Possibly? I’m an SRE. I don’t think my position is too different from the SREs I know :)

Replying to the second bit of your comment:

Commits before they are in master are fungible. If you have a PR with bad commits, you should treat it the same as bad code and force a refactor of the history. That’s enforceable during code review. I’d even do the same if their commits were too big and could reasonably be broken down into smaller changes.