Successfully Merging the Work of 1000 Developers
engineering.shopify.com
engineering.shopify.com
In addition to Bors, there are a number of apps and services that automate this kind of workflow. Here is an incomplete list: https://forum.bors.tech/t/other-apps-that-implement-the-nrsr...
Edited to add: Graydon Hoare (creator of Rust) called this the Not Rocket Science Rule Of Software Engineering (NRSROSE): "automatically maintain a repository of code that always passes all the tests" - https://graydon2.dreamwidth.org/1597.html
Disclaimer: I have contributed code to Bors.
The Servo and Rust projects have used a bors* for many years and have had few problems.
* Servo uses the https://github.com/servo/homu fork of the Homu rewrite of Graydon's original code, and Rust uses a slightly different fork at https://github.com/rust-lang/homu.
Seems a bit extreme to discard something after finding two bugs. Could you have fixed them instead?
With our CI times and PR volume, by the time any CI run completes, master would have drastically changed.
In a bors workflow, bors is the only thing allowed to push to master, so master cannot get out of date.
The Rust projects solves throughput with rollups, which are semi-automated. It would be nice if someone would write fully automated rollup support into a bors, but alas, no one has tried that I know of.
None of this is new at all, btw, I'm just regurgitating MVCC from postgresql.
Also the exact workflow described with staging branch and batch merge is probably another 20-30 lines.
Travis runs your test suite on the merge commit, not the head of the branch.
https://docs.travis-ci.com/user/pull-requests/#double-builds...
The linked page explains the speculative execution aspect, used to ensure every change is tested before merge, with the true state of master at the time of merge despite that state being different than it was when the CI run started ;)
https://cacm.acm.org/magazines/2016/7/204032-why-google-stor...
this was written 3 years ago and scale has only grown. But at the time the numbers quoted in the article are 25,000 authors / day w/ 16,000 commits and another 24,000 commits / day from automated sources
That way you can also add tags for flaky tests etc, to make your builds more reliable.
Edit: I didn't understand before what people had against monorepos, but I guess if you tie an continuous integration builds to each repo then having a monorepo becomes a huge pain point. Are there any open source tools to fix this?
Not to sound like a downer but this is really an article about fixing a broken process because not running CI on branches before merging to master goes against best practices. Would have loved to actually hear about their work process as this whole article could be summed up as "not running CI on branches before merging their commits to master is a great way to ruin master".
Perhaps this problem is what microservices are meant to solve. When you can't coherently integrate code fast enough, attack the bottleneck (master) by splitting it (multiple services).
Imagine an e-commerce site broken into a bunch of services including search and checkout. The search team is making updates daily, trying to improve ranking and drive conversion. The checkout team (assuming that the site is mature and has hit some design equilibrium) may only be releasing changes every couple of months, and if a bug is introduced, the financial impact is a lot higher.
By not bundling the outputs of very different teams together, you can help those that want to "move fast and break things" with their "moving fast" goal, and de-risk breaking everything by reducing the surface area of changes. Microservices-based architectures are a way to reduce friction caused by the structure of your organization and is one outcome of an Inverse Conway Maneuver.
In the SCS literature, you would integrate via async mechanisms across SCSes, provide versioned interfaces, and enforce via consumer-driven contract testing like Pact: https://pact.io
Kinda. Microservices have always been an organizational solution; they're a way to shard your company's work output. Usually that's API contracts, but whatever mechanisms are bottlenecked on the work output is affected, including how many concurrent builds are running due to how many people are touching the code at the same time.
https://eng.uber.com/research/keeping-master-green-at-scale/
I think the success of this system breaks down into several parts:
1) Yup, microservices. You could submit your proto change, which would affect all clients, before actually implementing the code that used the new feature. (Or after, in the case of renaming some field from foo to deprecated_foo and refactoring the clients to stop using that field.) That means you could wrangle that change without having to worry about it affecting your actual feature. (Typically proto changes did not cause any breakages since people were very conservative about what changes they would make. Nobody renames all the fields, invalidating dependent code, or renumbers the fields, invalidating all existing messages. You COULD do those things, but nobody ever did.)
2) Clear dependencies in the build system. The CI system only had to run a small set of tests for most changes, because it knew exactly what tests the change would affect. You had to go way out of your way to depend on code without informing the build system. This is very different from every CI system that I've seen outside of Google, which seem to default to running everything and hoping your programming language or build system magically tracks dependencies. It doesn't; Docker for example will happily use random images that it thinks haven't changed, without actually checking if it has changed. (Consider building your app on top of golang:latest. Go is updated, and docker may or may not pull that new base image. Meanwhile, docker will happily clear its build cache if you edit README.md and no code. The result is that 50% of the time you waste 10 minutes rebuilding stuff that didn't change, and 50% of the time you get an outdated build. And nobody seems to care at all!)
3) Being careful about keeping changes small. I don't know what the average CL size is, but I would aim for 100 lines changed rather than 1000 lines changed. This is something that surprised me post-Google, people go away and work for a week and you have a 2000 line PR to review. These are tough to merge and were relatively rare in my experience at Google. It is not always possible to make every change small, but that should be the norm. Figure out how much work you can do in a day, and try to make a CL/PR that is that size. A lot can churn in a week. A lot less churns in a day. If you respected steps 1 and 2, that means your tests will run fast and it's unlikely that your merge will fail between CI and actually merging. If you have 2000 lines of code across 8 services... you'll probably never get it merged. But I am sure that I have successfully merged ginormous changes before, it's just more work.
All in all, my takeaway from this article is that Shopify is huge but I'm surprised that specialized merge tooling was necessary. I wonder what the underlying problem is; do they really have a 1000 developer monolith? Do they not use a proper build system like Bazel?
Jenkins with a Github plugin behaves almost exactly like this system. Every PR basically has tests run 3 times; once for the branch that the PR is on, once for your branch merged to master, and then once after you do the merge and submit it. TAP presubmit did the "once for your branch" and TAP did the post-merge CI.
TAP presubmit didn't really check that the resulting merge was sound, so you would see TAP presubmit pass, your change get merged, and then have the build break anyway because of the race condition. A merge queue would not have this race condition... so I'm not sure Shopify has one either. The more I think about it the more it sounds like they just rewrote Jenkins. (And for that, I don't blame them.)
My philosophy was to always have integration tests run in the normal CI system. This basically meant creating a test binary that happened to link in the systems you were integrating with, and run tests against that. This is easier when everything is written in the same programming language, and for the cases where it wasn't, I was usually happy with "fakes". (https://testing.googleblog.com/2013/06/testing-on-toilet-fak...)
Other teams really loved the sandbox environment with live instances of everything. They would have some machinery outside the standard CI system to inject their code into this sandbox and run some tests, as well as machinery for keeping their sandbox up to date with production. (And adding test data, etc., etc., which all becomes very complex very quickly.)
Both methodologies have their downsides and upsides.
I generally prefer simplicity and speed; people should be able to run the tests on their workstation 100% of the time without having to set up any external resources. If you have an integration test binary that is built from the build system, this is possible. The downside is that config changes in production can break your system; since you are starting up your own instance of some other team's server, they could theoretically make some config change that breaks your integration. Even if you include their configuration in your in-memory version of their service, there was no guarantee that what is running in production is actually checked in yet. (Debugging in production, emergency rollback to an older prebuilt binary, etc.) These were rare and never caused me problems, however, and not having machinery to maintain a shadow environment meant it was easier to work on the code.
Having a sandbox environment was good because you could "check" (not test) big changes before putting them into production. You could try out your flag flip, database migration, mapreduce, or just load up the website in your browser and send your coworkers a link without affecting production data. And you could test your actual production binary in production-like conditions; as long as you sync'd production changes to your sandbox, your automated test probably ran against something that was very much like production. This let you check for more subtle things like performance regressions before deploying. (I worked on a system to do just that.)
The main problem I had with this method was that it was maintenance-intensive (big teams that used this had entire teams just to maintain the sandbox, and that begat sub teams that maintained the sandbox maintenance) and slow. Building and running another test during CI was relatively fast, but starting up a job in production and scaling it up was significantly slower. This meant that you needed a parallel set of tools to run some subset of this environment locally, and it was always painful. Not having your tests in the standard system meant that downstream dependencies wouldn't see test failures in your system when you made a change, so the "buildcop" would have to detect and fix that.
I found this to be too much overhead, but it is probably necessary when you are developing, say, a mobile application. You will have to write some sort of software to make it possible to try your in-progress code on your personal phone. You will probably want to be able to share links with coworkers. I generally like to push changes to production multiple times a day, and make sure that clients can handle a newer server and still work correctly. This way, as soon as a build passes tests, you can start giving it, say 0.1% of production traffic and keep an eye on the error rates, and promote that to production as quickly as possible. The biggest problem I've run into with this strategy is that 0.1% of Google's traffic is way more than enough for a good canary, but at other places I've worked... 0.1% of traffic might be one request over several days. In that case, you have to have staging and manually bug people to try it out. Sometimes I wonder if that kind of software is worth writing at all, to be perfectly honest. If you get one request a day, maybe just make it open a support ticket, and hire 2 support engineers instead of one software engineer. But I digress ;)
I've seen several blog posts from Google about using fakes and 'hermetic servers' for testing. We use GCP for our product, and unfortunately, Google doesn't seem to care much about making this easy. For example, I think I saw only one or two languages for which the Google Storage client libraries provided "fakes" of a Google Storage server. For PubSub (and maybe one or two other services?) there is the PubSub Emulator, which is unfortunately in Java and isn't supported by any of the CLI tools.
For all their love of fakes and hermetic servers, it would be awesome if they provided them for all the GCP services.
The implication here being that a queue must be processed in-order, so you will ultimately have a perfect sequence of future commits to speculate against, and can incrementally build up each hypothetical future master state for a test build on one of any number of parallel build agents. As the queue depth grows, you would see higher and higher throughput.
Google does it with 50 times the developer count.
> At 1000 developers, the rate of PRs coming in makes it impossible to determine what current master will be when the PR is ready to merge (i.e. when the branch has a green build).
True, it is impossible to catch all errors like this, but you can catch almost every error by building and testing it against current master and then merge it with the master 20 minutes later when the build is done. I have seen maybe one build breakage a year being introduced due to this in projects I've worked on, so it isn't a big deal.
And I'm pretty sure that is the way Google does it too. Test a commit against current master, if tests are green commit. Then run tests against master again (and I think this stage might not run for every single commit) to see if anything broke on the rare times there was an actual conflict. If that run was red, which should be rare, then you can have the system do a bisect to find the offending commit, or just run all the ones that haven't been individually tested.
You'll notice that those listed have had to customize or creat new vcs's to meet their needs.
https://news.ycombinator.com/item?id=17605371 https://news.ycombinator.com/item?id=11789182 https://medium.com/@maoberlehner/monorepos-in-the-wild-33c6e...
At scale, managing complexity can require either a lot of coordination, or a lot of careful planning. Large companies (especially tech companies) don't do either well, so they pick architectures that remove choices, and iterate on them until they are workable. And they have the money and workforce to do it.
Microservices are the same kind of solution, with the same gains and costs for this specific problem.
At this scale you need to be deploying constantly, otherwise deploys are hundreds of commits large and its impossible to triage - what PR in the deploy broke something, is it even safe to rollback, etc. That is the primary reason to automate deploys and manage the deploy queue.
What's the minimum residency time to reliably detect problems with my PR? Add deployment time, double to account for jitter caused by humans being humans (forgetful, lunch, meetings, etc), and there probably are not enough hours in the day for 1000 people to be deploying the same monolith.
To increase residency time you can deploy separate units (You can have multiple deployment units even in a monorepo), and those also reduce the surface area of merges.
Honestly what are they doing with 1000 developers? Duplicated effort goes up considerably with a team and codebase of that size. If you forced me to hire that many people, I'd have a lot of them working on open source, trying to steward feature enhancements that help our process. Because otherwise they'd be running around writing proprietary versions of a bunch of shit that already exists and in a better more documented form.
https://engineering.shopify.com/blogs/engineering/introducin....
Folks, when you hire enough devs, they feel empowered to rewrite the world. I have lived all sides of this phenomenon and rarely is it pretty.
Scaling is a concern that goes in both directions. Shopify has 1000 developers today. How screwed would they be if they suddenly had to drop to 600? Or even if there's a hiring freeze? What happens when the people who wrote these tools go work somewhere else?
When I do tool smithing work these days, it's always with an effort to provide the thinnest of shims around open source or commercial tools with healthy user communities, so that at the end of the day they have a larger pool of resources than what is in house. People move on. Money dries up. Mandates change.
"Being important" in a company is about how much you support new work, not how locked in people are to your old work. If you can't give your old work away then you're shackling yourself, both to your current responsibilities and to the company. I can't believe that I'm the only one who has ever stayed at a company out of guilt for how screwed they'd be if I left. But that quickly turns into resentment which is worse.
If you are important for new work, then you always get new challenges. You stay sharp and your resume looks good. If the company stops doing new work altogether, do you really want to stay there anyway? Plus you could always go back to one of your old projects.
Totally agree that CD is required with this many commits. It’s commonplace on teams with many fewer developers. Was surprised to see you folks roll your own workflows rather than using other systems.
Would also be interesting to see if you tag commits that go to master in instrumentation systems so you have visibility into production metrics and can correlate them with what code was running at the time.
I think the main thing that was missing for me is the rationale behind building this system rather than building a workflow in one of the existing CI/CD tools. Was there a throughout bottleneck in existing tools? Was there something custom about your workflow that wasn’t supported elsewhere? I may be wrong but the workflow you landed upon seems pretty common so I’m curious as to why the need to build and maintain a tool in house for this?
Pull requests are our unit of work, and the queue was created to support all pull requests. We do have feature flags as a tool, but we let our developers make the judgment call on how their changes should be rolled out.
How do you handle the scenario that some developer pushes a send_me_all_the_credit_card_details() function to the code base which does something 'evil'? Do you rely on the reviewer "doing their works properly" to handle that?
I'm not saying formal "signing off"-steps in processes handle it, but some companies does them for that reason.
I’m curious as to why you had a queue instead of a develop branch before moving to CD? Was this to allow arbitrary commits to be launched to production rather than getting them batched by time?
You will want to make your `develop` branch the default branch in git and on GitHub, to make sure pull requests automatically are targeted properly (not doing this would be a major UX pain). However, that means that when you `git clone` a repository you are not guaranteed to get a working version.
The `develop` branch can still be broken, which is a problem that needs to be addressed. While you can revert breaking changes (or force-push it to a previous known good sha), and you can automate this process, the pull request is already marked as merged at this point. This means that developers have to open a new PR whenever that happens.
With the queue approach, pull requests remain open until we are sure they integrate properly. Also, we have the opportunity to use multiple branches to test different permutations of PRs, so we can still progress and merge some PRs even if the "happy path" that includes all PRs does not integrate properly.
The major pain point I agree with on develop is changing the defaults to merge to that rather than master. It’s a shame this is not easier to do in git/github.
I’m not sure I agree with “develop can still be broken” as an issue that supports a queue. Whether it’s a queue or develop, one should run CI on each change to validate that merging it to master will not cause issues. It’s possible for both to be broken via the same scenarios just as it’s possible for master to be broken. Since CI runs before the branch is merged to develop and upon merge, a failure would “stop the world” and prevent more code from being merged unless that code fixes the failure.
I guess I’m not fully understanding how a queue prevents this. Since you don’t have a full picture of the state of master until something is merged from the queue, how do the CI checks in the queue prevent things that branch-based CI checks wouldn’t prevent in a “develop” branch? With branches and develop, pull requests remain open until they can be assured they merge properly with develop as well.
For clarity, I’m not arguing that a develop branch is the way to go, I think CD is much better.
Maybe I’m missing something big here but using multiple branches is permissible in other setups also. You can cherry pick a bunch of commits to a branch and test permutations but only certain branches get deployed to staging and production based on rules.
I’m glad that Shopify has found tools and a process that works. Honestly, I’m just having trouble comparing and constraining this to the other tools that are out there. The article never speaks about other approaches and whether or not they were considered and why you decided to go with a queue. It’s not clear to me if this was a case of improving the existing queue system because it was already in place or whether or not the queue was specifically chosen again because it was better than other alternatives (and why).
The trick of the merge queue is that it splits the "merging a branch / pull request" in two steps:
1. Create a merge commit with master and your PR branch as ancestors.
2. Update the `master` ref to point to the merge commit.
Normally when you press the "Merge Pull Request" button, it will do those two things in one go. By splitting it up in two distinct steps, we can run CI between step 1 and 2, and only fast-forward master if CI is green.
This means that master only ever gets forwarded to green commits. And because the sha doesn't change during a fast-forward, all the CI statuses are retained. Only when we fast-forward will GitHub consider to pull request merged, so we don't have to "undo" pull request merges when they fail to integrate. If the merge commit fails to build successfully, we leave a comment on the PR that merging failed, and the PR is still open.
When we have multiple PRs in the queue, we can create merge commit on top of merge commit, and run CI on those merge commits. When once of these CI runs comes back, we can fast forward master to it, and potentially merge multiple pull requests at once with this approach.
I’d imagine at 1,000 developers and with a monolithic codebase, you’re looking to minimize test runs both from a time and cost of runners perspective.
You may also want to look into Zuul or Bazel if cost of test suite runs is a factor in coming to this solution.
That wouldn't work for us due to the amount of changes we need to ship. If you rebase your branch and wait for CI to come back green, chances are another PR will have merged in the mean time, which means your rebased branch is no longer up to date with master. You end up stuck in a rebase cycle.
For this reason, we have no choice but to batch PRs, which is what the merge queue tool does. Faster CI will reduce this problem and we're working on that as well, but won't fully solve this.
Now the bot merges into a temporary that is fast-forwarded as the new master if CI validates it.
Would you say this is more of a decision based around the constraints of using GitHub or more of the ideal process for Shopify’s needs?
I’m curious because the article doesn’t mention the core reasons that you chose to write your own CD tool versus the other options that exist. The workflow you describe seems readily available in most tools. Perhaps the throughput was causing other options to break?
> We check if Branch CI has passed and if the pull request has been approved by a reviewer before adding the pull request to the queue
I'm starting to think most CI problems are just people not looking at the problem the right way. Here is the problem re-worded:
- When a PR has a green light and someone hits 'merge', it locks anything else from being to merge to master, and you merge your PR. When it finishes merging and deploying, now all the other PRs waiting have to rebuild themselves to see if they will merge with this new state of master. So 100s of PRs are rebuilding every time you merge one PR, and there's constant CI churn.
Here is why that problem exists:
- The system was designed for 1000 developers to all be writing to the same code base.
Here is how you solve that:
- Don't let 1000 developers all write to the same code base. Break the code down into discrete components that different small teams manage. The only bottleneck for that code base is that small team.
This small team is often called the two-pizza team, and their discrete components are often called microservices.
What's your point?
Although hard to get stats without control group - maybe control group could be acquisitions?
"continuous integration" is often just "npm ci && npm run build", and sometimes "npm run test" (or similar for your language). For products that don't make any remote API calls (or when they use a faker service), most of this is done on the same machine and costs very little to do on every commit, making it easier to precisely define which commit broke something.
Our CI takes ~8 hours of machine/VM time, which is about 35 minutes of wall time with our current testing cluster (including non-distributed parts like building). We skip certain long tests during the day, so that brings wall time down to ~13 minutes. But we also test 2-3 branches with decent churn. So even if we're only doing post-merge CI based on the current state of master, we're still getting 5+ commits fairly often.
I want to get to a world where CI is run before and after each merge with master, on every commit (or push/pull), but it seems like it would take so much more resources and infrastructure than we currently have.
I do like the emergency exception. The problem then becomes having a solid policy on when to use it.
Not really sure about the flakiness test (the 25% figure is totally random), as you could have a test that is failing only at a certain hour of the day and you would be furiously rerunning the test suite for nothing because it would always fail, wasting resources. It would be much better to at least single out the failing tests and rerun only those, dropping the ones that pass successfully until max tries are reached or the queue is empty.
https://docs.gitlab.com/ee/ci/merge_request_pipelines/pipeli...
says 25 minutes.
Who would ever think it's a good idea ?
The simple way would be to rebase your branch (or merge in master). However, with the amount of changes that are being merged, by the time the CI result comes in for your rebased branch, the tip of master is already changed making the CI result obsolete. So we went for the queue approach instead.
So overall you came to conclusion that what you explained even if can fail is pretty much the best way you can do it.. (rebase just custom..)
Overall DevOps principles are proven to simply work, you just have to follow them..
What’s odd is that they seem to characterize this as a tools issue rather than a process issue. There are plenty of CI/CD tools that allow for a similar or the same workflow as what they created. It’s also kinda scary that there’s not an emphasis on the overall SDLC and how the specific attributes of a branch or commit should/shouldn’t affect the process. You’d think at 1000 developers it’d be very important to define what is “launchable” as well. Haven’t worked with 1,000 on the same project but even with 20+, the standards and practices around development were always more important than the tooling. Tooling was just meant to represent workflows that were already defined.
Master should be always in a state of release at any moment.
I just cannot imagine it was an unknown practice for some.
Every new merge on master would require to rebase several hundred branches being worked on or awaiting reviewed. Multiply this with the hundreds of commits merged on master every day and you end up with way too much CI jobs to run.