On large teams I think the "cherry pick" workflow (Gerrit style) beats the "pull request" workflow (GitHub/gitlab style). On smaller teams it's the other way around. I think it's somewhere around 10-20 people actively committing that the cherry pick workflow comes out ahead.
This is exactly it. When I worked in a ~10 person team I just didn't get it, PRs worked quite well (with some basic discipline, they're not perfect). When I moved to a... much larger company... I don't know how PRs would work here, it would be way too unwieldy. The Gerrit style works fantastically here.
It makes every commit small, so they can be reviewed quickly and easily.
Small commits can often be tested faster, since irrelevant tests don't need to run.
Small commits are less risky. The smaller the delta of change, the lower the probability that something breaks.
Small commits get merged sooner; big commits take time to build up. Merging early front-loads your integration risk; merging later puts integration risk just before delivery.
Breaking a big feature into small commits means using feature flags to control whether a feature is enabled or not (since control paths will generally be incomplete). This means you separate the delivery of the code from the delivery of the feature, and has the added benefit that you can turn off a feature that has a problematic rollout without needing to redeploy code.
1 commit == 1 reviewable unit == 1 PR == 1 CL == 1 feature == 1 fix is a perfectly reasonable way of working.
I used to work at companies where no one squashed their commits and the entire git logs were filled with 80% non-sense like "temp" or "bad" or "working" with the other 20% being coherent changes. What's the point of doing this I ask?
I don’t quite get the situation. After merging a PR/MR (with the squash feature of Gitlab or GitHub) the equality you describe holds and there’s a nice sequential git log.
I quite like that I can tell my colleagues they can commit whenever they feel like it (and should probably commit more often than they feel like) because it becomes one clean commit in the end anyways.
In my experience the situation with GitHub is either that
- you apply the commits as-is on the target branch and end up with a ton of pointless "fix this" and "fix that" commits that were created during the review process,
- you squash everything and lose what might otherwise have been meaningful, useful commit boundaries for the purpose of bisecting, reverting or just investigating the history of changes or
- you rework your commits and force push changes to the branch you want to pull and create confusion as to what changed because that's not usually how review is conducted with GitHub.
I think the pull request is a bad abstraction. Not terrible, and certainly easy to grasp, but bad enough that everything you do with it is a compromise. At the very best, if you apply with the squash+rebase strategy and limit the scope of your PRs to what makes perfect sense as just one commit in the target branch, it's just a convoluted way of working with individual commits.
I wish git worked like some of the other dvcs (bazaar/breezy). The default merge type from a branch should be a merge commit and git log should only show the first parent (commits directly on the current branch).
This gives you a very clean log on main (it only shows commits directly on this branch). There is no need to squash, rebase, or anything else. And if you want to dig down into individual commits that happened on a branch, you can!
This can be a bit replicated by forcing a merge type of merge and setting an alias of log to "log --first-parent", but since that isn't the default, that isn't what you see when you look at the commit logs on platforms like github.
I never understood why git decided to show a flat list of every commit that happened, even if the commit originally happened on a separate branch.
there is an inbetween .... i insist people interactively rebase those commits out. In some contexts it is actually important to have traceability of iterative proof of work towards the final result.
Well are we talking about commits pre- or post-merge? I don’t care how many commits you put into the PR / MR as long as they squash down to a single commit upon merge.
The reason people like to work this way is that it lets you very effectively respond to review. You address them by amending the commit to incorporate the feedback.
This enables good interdiff review, so you can re-review just the new stuff in the new version of the commit and not the entire thing all over again.
I haven’t used this project but I have used Gerrit. It has its drawbacks (like terrible UX) but its style of code reviews were the most sensible and commit of every PR might not be as bad as it sounds. GitHub’s PR reviews are atrocious and it’s unfortunate they have become the gold standard.
In Gerrit, you commit every review, and the author has to edit individual commits to address them (using git rebase). This may sound PITA but it makes the history absolutely clean, and makes it easier for both reviewers and authors to review and address suggestions.
On Github, on the other hand, reviewing a large PR is just insanely hard. Making sure comment was addressed properly is hard as well, they can get lost in a sea of suggestions. They also become separate commits instead of being part of the commit itself. The commit should always been treated as unit of work rather than the branch.
IMHO equating commits and PRs puts undue pressure on the scope and quality of a given commit, adding potential for unnecessary stress and eliminating the benefits of an additional buffer / layer for aggregation of changes. A PR representing a sizable feature or refactor might naturally contain a dozen commits, each dedicated to a logical
area or a requisite subset of the whole. Assuming on principle a goal of keeping main in a known-good state, such intermediate and incomplete changes (fine in an unstable feature branch) would wreak havoc.
It's equivalent to asking, "Why would you have more than one story in an epic (or task in a story)?".
In Gerrit, I would say that the equivalent of the epic is the topic, stacks are the stories and commits are the tasks.
It certainly puts pressure on the scope and quality of commits, but I think that's only undue for a short-lived project where you don't foresee a long period of maintenance. Consistently high quality commits with a clear scope is a godsend when you are investigating the history of a project. Some small pressure now is IMO better than greater pressure later when you are debugging an issue at three in the morning and scrolling through either a bunch of "fix stuff" commits or 1000+ line PR squashes.
IMHO equating commits and PRs puts undue pressure on the scope and quality of a given commit, adding potential for unnecessary stress and eliminating the benefits of an additional buffer / layer for aggregation of changes
You can do as many commits as you want locally. Then go back and squash them before pushing. I think that addresses all your points, if I am not mistaken.
If your PR has more than one commit, each one should be deployable in isolation. Which means you can split your giant PR into smaller ones that can be reviewed independently.
I’ve worked under both systems, but isn’t the purity you’re describing a bit of a dodge in that you wind up force pushing amended commits when you find you forgot something?
People say they care about the "story" behind the PR. But no one cares about that story if it's about forgetting to fix a test and a typo in a comment.
The extra commits are just noise that make you think the original commit is a source of truth in a blame when it has been amended 3 times more in the same PR, but the link isn't apparent anymore.
Force pushing is bad to a published branch, not a feature branch (not that you really have force pushes in Gerrit anyway). There are versions of Gerrit tooling where you can have a branch as dirty as you want locally, but only the final aggregated change is visible for review of that's what you prefer too.
Sorry, “dodge” was a loaded word. I suppose what I mean is, what’s the difference between a PR with multiple commits and a commit with multiple commits?
I’m not sure what a “commit with multiple commits” is. I direct we’re talking past each other slightly :)
I think this is maybe what you’re asking about: with PRs, you often respond to feedback by adding new commits on to the PR. This is because code review is based on reviewing an entire branch. In a stacked diff based system, you respond to feedback by amending the commit. This is because reviews are tied to individual commits, not branches.
The reason people prefer this is that it keeps changes small and focused, and makes sure you have a high quality history. It also lets you assign different reviewers to different parts of your stack, which is helpful for all sorts of reasons. It also means that you can land earlier commits while waiting on review for the later commits, instead of holding it all up at once. Lots of stuff like this.
Any time. This stuff took me a while to wrap my head around, since I never worked at Google or Meta. I don't think I'm the best at explaining it yet either...
Ha, yeah. I worked at Box.com for a couple years and it took me months to figure out why my giant PRs full of genius were getting so much pushback. Just two different workflows depending on scale.
Comments
Who is creating a separate PR for each commit on their feature/fix branch?
sounds like crazy town.
I just dont understand why someone would operate like this.
Lets assume you're squash merging your feature branchs to your local main, then you're raising the them as prs.
why would you do this?
On large teams I think the "cherry pick" workflow (Gerrit style) beats the "pull request" workflow (GitHub/gitlab style). On smaller teams it's the other way around. I think it's somewhere around 10-20 people actively committing that the cherry pick workflow comes out ahead.
This is exactly it. When I worked in a ~10 person team I just didn't get it, PRs worked quite well (with some basic discipline, they're not perfect). When I moved to a... much larger company... I don't know how PRs would work here, it would be way too unwieldy. The Gerrit style works fantastically here.
It makes every commit small, so they can be reviewed quickly and easily.
Small commits can often be tested faster, since irrelevant tests don't need to run.
Small commits are less risky. The smaller the delta of change, the lower the probability that something breaks.
Small commits get merged sooner; big commits take time to build up. Merging early front-loads your integration risk; merging later puts integration risk just before delivery.
Breaking a big feature into small commits means using feature flags to control whether a feature is enabled or not (since control paths will generally be incomplete). This means you separate the delivery of the code from the delivery of the feature, and has the added benefit that you can turn off a feature that has a problematic rollout without needing to redeploy code.
This is standard practice in the "stacked diffs" world: one review, one commit.
1 commit == 1 reviewable unit == 1 PR == 1 CL == 1 feature == 1 fix is a perfectly reasonable way of working.
I used to work at companies where no one squashed their commits and the entire git logs were filled with 80% non-sense like "temp" or "bad" or "working" with the other 20% being coherent changes. What's the point of doing this I ask?
I don’t quite get the situation. After merging a PR/MR (with the squash feature of Gitlab or GitHub) the equality you describe holds and there’s a nice sequential git log.
I quite like that I can tell my colleagues they can commit whenever they feel like it (and should probably commit more often than they feel like) because it becomes one clean commit in the end anyways.
In my experience the situation with GitHub is either that
- you apply the commits as-is on the target branch and end up with a ton of pointless "fix this" and "fix that" commits that were created during the review process,
- you squash everything and lose what might otherwise have been meaningful, useful commit boundaries for the purpose of bisecting, reverting or just investigating the history of changes or
- you rework your commits and force push changes to the branch you want to pull and create confusion as to what changed because that's not usually how review is conducted with GitHub.
I think the pull request is a bad abstraction. Not terrible, and certainly easy to grasp, but bad enough that everything you do with it is a compromise. At the very best, if you apply with the squash+rebase strategy and limit the scope of your PRs to what makes perfect sense as just one commit in the target branch, it's just a convoluted way of working with individual commits.
I wish git worked like some of the other dvcs (bazaar/breezy). The default merge type from a branch should be a merge commit and git log should only show the first parent (commits directly on the current branch).
This gives you a very clean log on main (it only shows commits directly on this branch). There is no need to squash, rebase, or anything else. And if you want to dig down into individual commits that happened on a branch, you can!
This can be a bit replicated by forcing a merge type of merge and setting an alias of log to "log --first-parent", but since that isn't the default, that isn't what you see when you look at the commit logs on platforms like github.
I never understood why git decided to show a flat list of every commit that happened, even if the commit originally happened on a separate branch.
there is an inbetween .... i insist people interactively rebase those commits out. In some contexts it is actually important to have traceability of iterative proof of work towards the final result.
Well are we talking about commits pre- or post-merge? I don’t care how many commits you put into the PR / MR as long as they squash down to a single commit upon merge.
When you work this way, each commit is expected to be able to land independently.
Doesn’t sound like it leaves much room for error. How do you address PR / MR comments? Force push?
The reason people like to work this way is that it lets you very effectively respond to review. You address them by amending the commit to incorporate the feedback.
This enables good interdiff review, so you can re-review just the new stuff in the new version of the commit and not the entire thing all over again.
it lets you maintain version history when working, then most workflows auto squash on merge
I haven’t used this project but I have used Gerrit. It has its drawbacks (like terrible UX) but its style of code reviews were the most sensible and commit of every PR might not be as bad as it sounds. GitHub’s PR reviews are atrocious and it’s unfortunate they have become the gold standard.
In Gerrit, you commit every review, and the author has to edit individual commits to address them (using git rebase). This may sound PITA but it makes the history absolutely clean, and makes it easier for both reviewers and authors to review and address suggestions.
On Github, on the other hand, reviewing a large PR is just insanely hard. Making sure comment was addressed properly is hard as well, they can get lost in a sea of suggestions. They also become separate commits instead of being part of the commit itself. The commit should always been treated as unit of work rather than the branch.
Why would you have more than one commit for a PR? That sounds like crazy town.
IMHO equating commits and PRs puts undue pressure on the scope and quality of a given commit, adding potential for unnecessary stress and eliminating the benefits of an additional buffer / layer for aggregation of changes. A PR representing a sizable feature or refactor might naturally contain a dozen commits, each dedicated to a logical area or a requisite subset of the whole. Assuming on principle a goal of keeping main in a known-good state, such intermediate and incomplete changes (fine in an unstable feature branch) would wreak havoc.
It's equivalent to asking, "Why would you have more than one story in an epic (or task in a story)?".
In Gerrit, I would say that the equivalent of the epic is the topic, stacks are the stories and commits are the tasks.
It certainly puts pressure on the scope and quality of commits, but I think that's only undue for a short-lived project where you don't foresee a long period of maintenance. Consistently high quality commits with a clear scope is a godsend when you are investigating the history of a project. Some small pressure now is IMO better than greater pressure later when you are debugging an issue at three in the morning and scrolling through either a bunch of "fix stuff" commits or 1000+ line PR squashes.
You can do as many commits as you want locally. Then go back and squash them before pushing. I think that addresses all your points, if I am not mistaken.
If your PR has more than one commit, each one should be deployable in isolation. Which means you can split your giant PR into smaller ones that can be reviewed independently.
I’ve worked under both systems, but isn’t the purity you’re describing a bit of a dodge in that you wind up force pushing amended commits when you find you forgot something?
People say they care about the "story" behind the PR. But no one cares about that story if it's about forgetting to fix a test and a typo in a comment.
The extra commits are just noise that make you think the original commit is a source of truth in a blame when it has been amended 3 times more in the same PR, but the link isn't apparent anymore.
Force pushing is bad to a published branch, not a feature branch (not that you really have force pushes in Gerrit anyway). There are versions of Gerrit tooling where you can have a branch as dirty as you want locally, but only the final aggregated change is visible for review of that's what you prefer too.
Why is that a dodge? that's the expected way to work in this system, and it should be able to show you the interdiff between those amends.
Sorry, “dodge” was a loaded word. I suppose what I mean is, what’s the difference between a PR with multiple commits and a commit with multiple commits?
I’m not sure what a “commit with multiple commits” is. I direct we’re talking past each other slightly :)
I think this is maybe what you’re asking about: with PRs, you often respond to feedback by adding new commits on to the PR. This is because code review is based on reviewing an entire branch. In a stacked diff based system, you respond to feedback by amending the commit. This is because reviews are tied to individual commits, not branches.
The reason people prefer this is that it keeps changes small and focused, and makes sure you have a high quality history. It also lets you assign different reviewers to different parts of your stack, which is helpful for all sorts of reasons. It also means that you can land earlier commits while waiting on review for the later commits, instead of holding it all up at once. Lots of stuff like this.
Cool, thanks.
Any time. This stuff took me a while to wrap my head around, since I never worked at Google or Meta. I don't think I'm the best at explaining it yet either...
Ha, yeah. I worked at Box.com for a couple years and it took me months to figure out why my giant PRs full of genius were getting so much pushback. Just two different workflows depending on scale.
Not once they hit master, no. You push bug fix commits.
A PR is a collection of commits?
Integration into main ideally is squashed down to a single merge commit though.
It shouldn't matter how many commits a pr takes to from 0 to finished
You wouldn't. Imagine you have more than one of what you are calling a "feature/fix branch" and they depend on each other.