Pre-commit, pre-push or pre-merge ?
November 14, 2020
On my codebases, I may want to run:
- a linter,
- static analysis,
While I can run them manually, I find it easier for everyone involved if it happens on its own.
It’s about getting feedback as soon as possible without breaking the flow of work.
When and where should they run?
- on save: runs every time I save a file on disk,
- pre-commit: runs before I commit my changes,
- pre-push: runs before I push mu changes to a remote,
- pre-merge: runs on CI when I do a merge request (or “pull-request”).
|on save||very early||very often||locally|
My advice is that the correct answer for a given task will depend on the “lost” time perceived by the developer. Perceived lost time ~= runtime + time to correct the code, if the task is blocking.
The slower the task, the later and less often it should run.
If something feels slow, it will get in the way and people (me included) will work around it.
If I have to wait a long time before knowing if my tests pass or not, I probably will switch to another task and waste time.
If the task itself can correct the code instead of rejecting the commit (e.g. linters), make it so.
Here is my rule of thumb, to be adapted to your needs:
|Runtime||Earliest it can run|
|less than 100 milliseconds||on save|
|less than 3 seconds||pre-commit|
|less than 10 seconds||pre-push|
|above 10 seconds||pre-merge|
|above 5 minutes||split into sub-tasks|
To state the obvious : every team, every codebase, every workflow is different. What works for one won’t necessarily work for others.
The tools we use give us even more options to choose when to run what, according to our preferences:
- IDE Integrations,
- manual or conditional tasks on CI,
- third party tools integrating with our version control system
For instance, I choose to run linters on pre-commit, not on save, even if it’s fast, because I find it distracting to have code jumping around on the screen.
I run unit tests pre-merge but trigger end-to-end tests manually because they are very slow, and would delay feedback too much.
Experiment and gather feedback.