This document is a long-winded guide to preparing and submitting your first contribution to CockroachDB. It's primarily intended as required reading for new Cockroach Labs engineers, but may prove useful to external contributors too.
At the time of this writing, CockroachDB is several hundred thousand lines of Go code, plus a smattering of C++ and TypeScript. This section is a whirlwind tour of what lives where.
Since CockroachDB is open source, most of the instructions for hacking on CockroachDB live in this repo, cockroachdb/cockroach. You should start by reading the top level of this section of the wiki.Â
Then, look at our Go (Golang) coding guidelines and the Google Go Code Review guide it links to. These are linked to from the other pages in this wiki, but they're easy to miss. If you haven't written any Go before CockroachDB, you may want to hold off on reviewing the style guide until you've written your first few functions in go.
Here's what's in each top-level directory in this repository:
build/ Build support scripts. You'll likely only need to interact with [build/builder.sh]. See "Building on Linux" below.
c-deps/ Glue to convince our build system to build non-Go dependencies. At the time of writing, "non-Go dependencies" means C or C++ dependencies.
cloud/kubernetes/ Kubernetes configuration to auto-launch CockroachDB clusters.
docs/ Documentation for CockroachDB developers. See "Internal documentation" below.
monitoring/ Configuration to integrate monitoring frameworks, namely Prometheus and Grafana, with CockroachDB. This configuration powers our internal monitoring dashboard as well.
pkg/ First-party Go code. See "Internal documentation" below for details.
scripts/ Handy shell scripts that aren't part of the build process. You'll likely interact with scripts/gceworker.sh most, which spins up a personal Linux VM for you to develop on in the GCE cloud.
vendor/ A Git submodule that contains the right version of our dependent libraries. For many years, the Go answer to dependency management was "write backwards-compatible code." That strategy is as dangerous as it sounds, so since v1.6, Go will automatically load packages in a subdirectory named
vendor, if it exists. See build/README.md for details on how we maintain this folder.
Besides cockroachdb/cockroach, the cockroachdb GitHub organization is home to several other important open-source components of Cockroach:
cockroachdb/examples-go, which contains small, self-contained Go programs that exercise CockroachDB via the PGWire protocol. You're likely to hear most about block_writer, which writes uniformly random values into a table, and photos, which simulates a more-realistic workload of a photo-sharing site, where some photos and users are orders of magnitude more popular. The other example programs are of a similar scope and purpose, but block_writer and photos are deemed important enough to run constantly against our production clusters.
cockroachdb/examples-orms, which showcases ORMs a toy API that uses an ORM to prepare its responses in several different languages.
Most of the remaining repositories under the cockroachdb organization are forks of existing Go libraries with some small, Cockroach-specific patches.
Documentation on the first-party Go packages that make up CockroachDB is, as of this writing, essentially nonexistent. This is par for the course with code that's evolving as quickly as Cockroach, but it's something we're hoping to improve over time, especially as internal packages stabilize.
The internal documentation that we do have lives in cockroachdb/cockroach/docs. At the time of writing, most of this documentation covers the high-level architecture of the system. Only a few documents hone in on specifics, and even those only cover the features that were found to cause significant developer frustration. For most first-party packages, you'll need to read the source for usage instructions.
Protip: If you prefer Go-generated HTML documentation to reading the source directly, take a look at godoc.org/github.com/cockroachdb/cockroach.
For our internal docs, I recommend the following reading order.
First, browse through the design document, which describes the architecture of the entire system at the highest possible level. You'll likely find there's too much information here to digest in one sitting: you should instead strive to remember what topics are covered, so you can refer to it later with more specific questions in mind.
Then, look through the docs/tech-notes folder and determine if any of the tech notes are relevant to your starter project. Again, you'll likely find that the tech notes contain too much information to process, so instead try to identify the sections that are likely to be useful as you make progress on your starter project.
The one exception to this rule is docs/tech-notes/contexts.md. It's worth learning why so many of our function signatures take a context as their first parameter, like so:
func doOperation(ctx context.Context, ...)
Otherwise, plumbing contexts everywhere will feel like a chore with no upsides.
Finally, I feel obligated to reproduce this disclaimer from the tech notes README:
Standard disclaimer: each document contains parts from one or more author. Each part was authored to reflect its author's perspective on the project at the time it was written. This understanding is necessarily subjective: its context is both the state of the project and the authors', and their reviewers', experience around that particular date. In case of doubt, consult your local historian and your repository's timeline.
In short, our documentation is not authoritative. Trust your reading of the code.
It's time to fire up your editor and get to work! The other pages on this wiki doÂ a good job of describing the basic development commands you'll need. In short, you'll use
make to drive builds.
To produce a
$ make build
To run all the tests in
$ make test PKG=./pkg/rest/of/path
Yes! It's calledÂ
You can useÂ
delve on its own, but the easiest way is to use it via Goland. Goland is a Jetbrains IDE for Go. It has built in support for delve, and will "just work" if you click the standard debug buttons, set breakpoints, etc.
Do note that the debugger can be somewhat slow on live clusters. It may be easiest to use printfs, log.Infof, and so on, depending on what you're trying to do. Use your judgement!
If you're debugging an issue that spans several nodes, you could consider using our distributed tracing. Ask your Roachmate to either walk you through OpenTracing/LightStep or point you at someone who can.
You'll notice that all source files in this repository have a license notification at the top. Be sure to copy this license into any files that you create.
Code at Cockroach Labs is not about perfection. It's too easy to misinterpret "build it right" as "build it perfectly," but striving for perfection can lead to over-engineered code that took longer than necessary to write. Especially as a startup that moves quickly, building it right often means building the simplest workable solution, then iterating as necessary. What we try to avoid at Cockroach Labs is the quick, dirty, gross hackâbut even that can be acceptable to plug an urgent leak.
So, you should get feedback early and often, while being cognizant of your reviewer's time. You don't want to ask for a detailed review on a rough draft, but you don't want to spend a week heading down the wrong path, either.
You have a few options for choosing when to submit:
You can open a PR with an initial prototype using the âDraftâ feature. (âCreate Draft PRâ in the Github interface). This won't send an email notification to anyone watching this repository.
For small changes where the approach seems obvious, you can open a PR with what you believe to be production-ready or near-production-ready code. As you get more experience with how we develop code, you'll find that larger and larger changesets will begin falling into this category.
PRs are assumed to be production-ready (option 3) unless you create it as âDraftâ or say otherwise in the PR description. Be sure to note if your PR is a WIP to save your reviewer time.
If you find yourself with a PR (
git diff --stat) that exceeds roughly 500 line of code, it's time to consider splitting the PR in two, or at least introducing multiple commits. This is impossible for some PRsâespecially refactorsâbut "the feature is only partially complete" should never be a reason to balloon a PR.
One common approach used here is to build features up incrementally, even if that means exposing a partially-complete feature in a release. For example, suppose you were tasked with implementing support for a new SQL feature, like supporting subqueries in UPDATE statements. You might reasonably submit four separate PRs. First, you could land a change that adjusts the SQL grammar and links it to a dummy implementation that simply outputs an error like "subqueries in UPDATE not supported" instead of "syntax error." Then, you might land a change to implement a naive version of the feature, with tests to verify its correctness. A third PR might introduce some performance optimizations, and a fourth PR some refactors that you didn't think of until a few weeks later. This approach is totally encouraged! It's no problem if the first PR (the one that simply prints "not implemented") lands in an unstable release, as long as a change isn't backing you into a corner or causing server panics. Code produced from an incremental approach is much easier to review, which means better, more robust code.
In general, production-ready code:
Is free of syntax errors and typos
Omits accidental code movement, stray whitespace, and stray debugging statements
Documents all exported structs and functions
Documents internal functions that are not immediately obvious
Has a well-considered design
That said, don't be embarrassed when your review points out syntax errors, stray whitespace, typos, and missing docstrings! That's why we have reviews. These properties are meant to guide you in your final scan.
So, you've written your code and written your tests. It's time to send your code for
First, read the Go (Golang) coding guidelinesÂ again, looking for any style violations. It's easier to remember a style rule once you've violated it.
Then, run our suite of linters:
$ make lint
This is not a fast command. On my machine, at the time of writing, it takes about a full minute to run. You can instead run
$ make lintshort
which clocks in at about 30s by omitting the slowest linters.
In the interest of branch tidiness, we ask that even contributors with the commit bit avoid pushing directly to this repository. That deserves to be properly called out:
Protip: don't push to cockroachdb/cockroach directly!
Instead, create yourself a fork on GitHub. This will give you a semi-private personal workspace at github.com/YOUR-HANDLE/cockroach. Code you push to your personal fork is assumed to be a work-in-progress (WIP), dangerously broken, and otherwise unfit for consumption. You can view @bdarnell's WIP code, for example, by browsing the branches on his fork, bdarnell/cockroach.
Then, you'll need to settle on a name for your branch, if you haven't already. This isn't a hard-and-fast rule, but Cockroach Labs convention is to hyphenate-your-feature. These branch identifiers serve primarily to identify to you.
To give you a sense, here's a few feature branches that had been merged in recently when this document was written, and their associated commit summaries:
sql: ensure that DEFAULT exprs are re-evaluated during backfill.
sql: show only constrained cols in fk
It's far more important to give the commit message a descriptive title and message than getting the branch name right.
Be kind to other developers, including your future self, by splitting large commits. Each commit should represent one logical change to the codebase with a descriptive message that explains the rationale for the change.This is explained in our PR organization philosophy.
On your first PR, it's worth re-reading theÂ Git Commit Messages pageÂ to ensure your commit messages follow the guidelines.
Most importantly, all commits which touch this repository should have the format
pkg: message, where
pkg is the name of the package that was most affected by the change. A few engineers still "watch" this repository, which means they receive emails about every issue and PR. Having the affected package at the front of every commit message makes it easier for these brave souls to filter out irrelevant emails.
Protip: often, you'll realize after a bout of hacking that you've actually made n separate changes, but you've got just one big diff since you didn't commit any intermediate work. Explaining how to fix this is out of scope for this guide, but either Git "patch mode" (Google it) or a graphical Git client will take care of it.
See also the "When to submit" section above for advice on when multiple commits should be further split into multiple PRs.
Deciding who should review a PR requires context that you just won't have in your first month. For your starter project, your Roachmate is the obvious choice. For little bugs that you fix along the way, ask your Roachmate or your neighbor if there's someone who's obviously maintaining a particular area of the codebase. (E.g., if you had a question about the code that interfaces with Pebble, I would immediately direct you to @petermattis.)
If you're unable to get a reviewer recommendation, the "Author" listed at the top of the file may be your best bet. You should check that the hardcoded author agrees with the Git history of the file, as the hardcoded author often gets stale. Use
git log FILE to see all commits that have touched a particular file or directory sorted by most recent first, or use
git blame FILE to see the last commit that touched a particular line. The GitHub web UI or a Git plugin for your text editor of choice can also get the job done. You might also try
scripts/authors.sh FILE, which will show you a list of contributors who have touched the file, sorted by the number of commits by that author that touched the file. If there's a clear winner, ask them for a review; if they're not the right reviewer, they'll suggest someone else who is. In cases that are a bit less clear, you may want to note in a comment on GitHub that you're not confident they're the right reviewer and ask if they can suggest someone better.
If you're still unsure, just ask in chat! You'll usually get a response in under a minute.
It's important to have a tough reviewer. In the two months I've been here, I've seen a few PRs land that absolutely shouldn't have landed, each time because the person who understood why the PR was a bad idea wasn't looped into the review. There's no shame in thisâit happens! I say this merely as a reminder that a tough review is a net positive, as it can save a lot of pain down the road.
Also, note that GitHub allows for both "assignees" and "reviewers" on a pull request. It's not entirely clear what the distinction between these fields is, and we're currently split at Cockroach Labs on what we prefer. Choose the field you like most and move on. Ever since GitHub began auto-suggesting reviewers, there seems to be a slight preference for using the reviewer field instead of the assignee field.
Nearly everything you're tempted to put in a PR description belongs in a Git commit message.Â (As an extension, nearly everything you put in a commit message belongs in code comments too.)
You should absolutely write a PR description manually if you're looking for a non-complete reviewâi.e., if your code is a WIP.
Otherwise, something simple like "see commit messages for details" is good enough. On PRs with just one commit, GitHub will automatically include that one commit's message in the description; it's totally fine to leave that in the PR description.
We run our own continuous integration server called TeamCity, which runs unit tests and acceptance tests against all open pull requests.
GitHub displays the status of the latest TeamCity run at the bottom of every pull request. You can click the "Details" link to get insight into a failed build or view real-time status of an in-progress build. Occasionally, a build will fail because of flaky tests. You can verify by running a new build and seeing if the problem disappears; just hit the "Run" button on the top right of the page GitHub links to queue a new build. Less frequently, TeamCity will entirely fail to notice a PR, and GitHub will display "waiting for status to be reported" forever. The "TeamCity Continuous Integration" wiki page describes how to fix this.
Unfortunately, a full tutorial on TeamCity is outside the scope of this document. Ask your Roachmate for an overview if you're confused, or post specific questions in #teamcity in slack.
That's it! You're done. Sit back and get ready for the impending code review.
Code reviews are one of the main ways for Cockroach Labs to guarantee quality in the product. We use them to ensure that all code written meets our strict standards for production-ready code.
A code review is like a discussion. The reviewer can have questions, and can even sometimes request the author to write a deeper justification for changes. The goal of that discussion is to end up with the highest quality product, and create a track record of why and how a change occurred, so we can later look back and understand the choices.
No one is expected to write perfect code on the first try. That's why we have code reviews in the first place!
For more information on what you might see as part of a code review, see What to expect when youâre expecting (someone to review your code at CRL).
As you advance in your work on CockroachDB, youâll inevitably be asked to review someone elseâs code. For more information on whatâs expected of reviewers, see Working as a reviewer.
Except for the smallest of PRs, many teams eschew GitHub reviews in favor of a third-party app called Reviewable (demo video). Reviewable has an incredibly dense UI: you could probably master a new musical instrument faster than you could master Reviewable.
Still, some teams have decided that Reviewable offers a better UI than GitHub for large changes that undergo multiple rounds of feedback. The basic model is the same: a diff of your change where your reviewers can leave inline notes. In Reviewable, unlike GitHub, every comment spawns a new thread of discussion, and the UI will highlight any threads that you haven't acknowledged/responded to.
Reviewable also takes great pains to handle force pushes. When you push a new revision, Reviewable will do its best to match up comment threads to the logically-equivalent location in the new diff. Even if this algorithm fails, Reviewable records each force push as a "revision," and will allow you and your reviewers to track how your PR has evolved as you address review feedback.
You won't need to do anything special to enable Reviewable for your PR; one of our bots will be along shortly after the PR is opened to post a link to the review interface.
Check with someone on your team if to see what your team prefers, but even if your team does not use reviewable for sending each other reviews in your day to day work, you will still want to be familiar with it if for reviews on changes that cross team boundaries.
If someone leaves line comments on your PR without leaving a top-level "looks good to me" (LGTM), it means they feel you should address their line comments before merging. That is, without an LGTM, all comments are roughly in the discussing disposition, even if the Reviewable disposition marker in the bottom right corner hasn't been set as such. If you agree with their feedback, you should make the requested change; if you disagree, you must leave a comment with your counterargument and wait to hear back.
If someone leaves a top-level LGTM/accepts your PR using Github but also leaves line comments, you can assume all line comments are in the satisfied disposition, even if the Reviewable toggle hasn't been set as such. If you disagree with any of their feedback, you should still let them know (e.g., "actually, I think my way is better because reasons"), but you don't need to wait to hear back from them to merge.
You may see comments labeled as "optional" to indicate a one-off satisfied disposition when the comment default is discussing (i.e., you haven't yet received an LGTM).
Similarly, you may see an "LGTM modulo the comments" to indicate that you're free to merge provided that you make the requested changes. If you do disagree with the feedback, then you need to voice your counterargument and wait for the reviewer to respond.
Rarely, someone will express a sentiment like "I feel very strongly that we shouldn't merge this." Disagreements like these are often easier to resolve outside of Reviewable, via an in-person discussion or via chat.
Reviewers may also have comments about subjective matters, for example what code presentation should be considered idiomatic. When in doubt, if the approach suggested by the reviewer doesn't change functionality, it may be a learning opportunity to follow the suggestionâif the reviewer cared to give the suggestion, probably someone else would trip up in the same way too. However, remember that a review is also a two-way discussion and you should feel free to ask your own questions to the reviewer, ask them to motivate their suggestions, and you can also weigh their considerations against your own. Sometimes, an appropriate resolution is to copy a summary of a conversation during the review into a comment inside the source code.
Once you've accumulated a set of changes you need to make to address review feedback, it's time to force-push a new revision. Be sure to amend your prior commitsâyou don't want to tack on a bunch of fixup commits to address review feedback. If you've split your PR into multiple commits, it can be tricky to ensure your review feedback changes make it into the right commit. A tutorial on how to do so is outside the scope of this document, but the magic keywords to Google for are "git interactive rebase."
You should respond to all unresolved comments whenever you push a new revision or before you merge, even if it's just to say "Done."
Protip: Wait to publish any "Done" comments until you've verified that your changes have been force-pushed and show up properly in Reviewable. It's very easy to preemptively draft a "Done" comment as you scroll through feedback, but then forget to actually make the change.
These are the acronyms you'll frequently encounter during code reviews.
CR, "code review"
PR, "pull request"âyou probably knew that one.
PTAL, "please take another look"
RFAL, "ready for another look"âas in, I made some changes, PTAL.
LGTM, "looks good to me"âi.e., ship it!
, "science dog"âi.e., I have no idea what this code does.
TF[YT]R, "thanks for your/the review"
A list of even more Cockroach jargon lives on this repository's wiki.
External contributors: you don't need to worry about this section. We'll merge your PR as soon as you've addressed all review feedback!
To merge code into any CockroachDB repository, you need at least one LGTM (or Github PR approval). Some LGTMs matter more than others; you want to make sure you solicit an LGTM from whoever "owns" the code you're touching.
Occasionally, someone will jump into a review with some minor style nits and leave an LGTM once they see you've addressed the nits; you should still wait for the primary reviewer you selected to LGTM the merits of the design. To confuse matters, sometimes you'll get a drive-by review from someone who was as qualified or more qualified than the primary reviewer. In that case, their LGTM is merge-worthy. Usually it's clear which situation you're dealing with, but if in doubt, ask!
Once you've gotten an LGTM from who you think is the right person, don't be afraid to merge. The git revert command exists for a reason. You can expect to revert at least one PR that you land in your first three months.
When your PR is ready to go, request a merge from our build bot Craig. Craig is a Bors merge bot, whose only job is to be gatekeeper to the repository. Add a comment on the PR of the form
bors r=<reviewer>, replacing
<reviewer> with the GitHub username of the person who gave you the LGTM. Once approved, your PR will be batched up with other PRs approved around the same time. Craig will try to build them as though they were merged to the target branch, and if successful, will merge them. This limits your exposure by ensuring you don't accidentally merge a commit with an obviously broken build or merge skew. (IfÂ Craig says "Permission denied", see the FAQ.)
Congratulations on landing your first PR! It only gets easier from here.
Here's a checklist of action items to keep you sane:
bors r=reviewerto ask Craig to merge