Strangers on the Internet, or Why PRs Are Conversations, Not Notifications
Here’s something I say a lot that usually gets a reaction: we are not strangers on the internet.
The open-source PR model — push your work, notify the team, hope someone picks it up, iterate based on comments — makes total sense when contributors don’t know each other and are working across time zones and organizations with no shared context. It was built for that. It works great for that.
It does not work great for a team of people who sit (or video call) together every week, have shared context on the codebase, and are accountable to each other for delivery outcomes.
When I watch internal teams use the open-source model, I see a few things consistently:
PRs get big before anyone looks at them. The engineer works in isolation until they think the thing is done, then opens a PR with thirty changes. The reviewer now has to reconstruct the entire intent of the work from the diff. This is expensive for the reviewer, and the engineer has already invested too much to pivot meaningfully on structural feedback.
Reviews become rubber stamps or rabbit holes. Either the reviewer approves because the thing looks roughly right and they don’t want to slow it down, or they find a fundamental problem and now you have a PR comment thread that should have been a twenty-minute conversation three days ago.
Coverage and quality go sideways late. By the time someone notices a testing gap or an architectural issue, the engineer is context-switched out and getting back up to speed costs time neither of you wanted to spend.
The model I’ve had the most success with is straightforward: have the conversation early, stage the review, and use the PR to confirm shared understanding rather than initiate it.
Specifically — if you’re working on something that someone else will need to sign off on, loop them in before you start. “I’m going to do this thing. Here’s my plan. Any concerns?” Then, at the midpoint, show them the architecture — not the implementation details, just the structure. Names, boundaries, where things live. Get that directional feedback when it’s cheap to act on. Then finish the implementation and let the PR be the final confirmation: here’s what we talked about, here’s what I built, here’s the test coverage that says it works.
If the PR fails automated checks (coverage below threshold, build broken, linting issues) it doesn’t even need to get to a human. It goes back to the engineer with a clear signal. Fix the basics first. Don’t ask me to spend time on work that doesn’t meet the baseline.
PRs as the ends of conversations. It’s a small mental shift with a surprisingly large impact on how smoothly a team moves.
More Code, More Risk ... More Automation
One of the things I keep saying to engineering teams right now is: this tool should improve the process on all fronts. Not just the building. All of it.
AI-assisted development is genuinely remarkable. Engineers are shipping features faster, clearing backlogs that felt immovable, and getting real leverage from tooling that would’ve taken months to build before. I’m a fan. I’m actively encouraging it.
But faster output without better validation is just faster bugs. I’ve seen it — and heard it from peers managing teams at scale. Engineers trust the output, skip the second look, and something breaks in production. The specific failure mode varies. The root cause almost never does: the code moved faster than the confidence in it.
The issue isn’t that AI makes mistakes. It does, and so do humans. The issue is that when the volume of work goes up, the surface area for mistakes goes up with it. If your validation process doesn’t scale at the same rate, you end up cleaning up behind yourself constantly.
The answer isn’t to slow down. It’s to automate your safety net.
Fast builds. Fast test runs. Automated coverage thresholds that fail a PR before it ever hits a human reviewer. If the test coverage isn’t there, the PR doesn’t get reviewed. Period. That sounds harsh, but it’s actually generous — it tells the engineer clearly and immediately what they need to do before asking anyone else to spend time on their work.
I also think there’s something important here about how we use AI in the validation loop, not just the construction loop. Use it to generate test cases. Use it to review test coverage for gaps. Have it challenge your assumptions about how a feature might fail. A lot of what a good QA person does — thinking adversarially about software — can be augmented with the same tools you’re using to write the code.
The teams that win with AI aren’t the ones moving fastest. They’re the ones who figured out how to move fast and land cleanly.
Testing Is a Team Sport
I’ve had some version of this conversation with multiple engineering leaders lately, and I keep landing in the same place: quality is a team sport, and most teams haven’t figured that out yet.
Here’s the dynamic I see a lot. You’ve got developers who write their code, write their tests (maybe), and hand the whole thing off. Somewhere downstream, a tester or QA person picks it up and starts poking holes. When they find something, it goes back. When they don’t, it ships. Everyone’s doing their job. And yet somehow, bugs still make it to production, morale is still low, and the whole thing feels slower than it should.
The problem isn’t the people. The problem is the model.
Separating “building” from “quality” creates distance — between the person who knows the most about how something was constructed and the process of verifying that it works correctly. That distance is where bugs hide.
The fix isn’t to hire more testers. It’s to collapse that distance. Make the people accountable for delivery also accountable for correctness. Pair them up. If two engineers worked on a feature together, they’re both on the hook if it breaks. Not “well, they were supposed to test it.” No. Both of them own it.
What I’ve seen work is building that expectation into the culture from the top. Not as a punitive thing, but as a shared ownership thing. *We* built it. *We* stand behind it. If it’s wrong, we fix it, we learn from it, and we make sure it doesn’t happen that way again.
The corollary to this is that no single person (or department) in an org should carry “quality” as their sole responsibility. One QA lead against a team of fifteen engineers is not a quality strategy. It’s a bottleneck with a job title. The engineerings and their leadership (the people accountable for delivery) also have to be accountable for what they deliver working correctly. That’s not a radical idea. It’s just engineering maturity.
Quality isn’t a department. It’s an expectation.
Feature Flags for Fun & Profit
By Gerald Singleton & Clint Parker
Intro
Continuous integration & deployment. Automated testing. Refactoring & taking control of the monolith. Reducing cycle time. Increased uptime. Optimizing the data layer. Putting stakeholders in control. Making customers happy.
We’ve done all that. And it was all easier with our aggressive adoption of feature flags. Using flags by default has unblocked all of the more visible initiatives we wanted to achieve. In this document, we’ll showcase what flags are (and aren’t), why we use them to the extent we do, and how you can quickly take advantage of these patterns to improve your codebases.
Guidelines
Not (just) for features
A common misconception is that feature flags are only for toggling application features. Development teams often tend to think that an application will have a finite number of features at any given time, and those features can be managed via configuration settings. This thinking limits the overall flexibility of your application. In our SDLC, we don’t use feature flags just to toggle application features. In fact, we primarily use feature flags as a deployment tool. They enable us to deploy changes to our product frequently, purposefully, and most importantly…safely.
Columns vs. Layers
A foundational concept of feature flags is context vs config. Many applications use a variety of pre-production environments with their own config settings and slightly different expectations. An example would be connection strings for the different versions of the same DB. Or maybe a sandbox vs production integration dependencies. These are usually diagramed as vertical, separated by horizontal lines as layers, hence the phrase “lower environments.” Layers are expensive. There are tricks to reduce costs, but nevertheless, they are expensive.
On the other hand, horizontal segmentations, or columns, tend to be much less expensive and dynamic. These are thought of as usage characteristics, like the number of users, geography, and account settings, which are cheap. For this purpose, we can group all the varieties of horizontal segmentation into the term “context.” Contexts don’t have to have names, nor do they need special dependencies. Contexts are unmanaged. They simply exist within the runtime space of the code.
Context
Context is not a reserved word in this case; it’s a concept. LaunchDarkly does have a specific meaning, but that’s not what this is. As mentioned, context already exists in your application. You already have settings for the application itself, organizations, users, geographies, and time/date. Contexts can be split or combined. The point is that your application can accommodate this perspective and already does. Accepting this is important because the appropriate usage of feature flags will maximize it.
Feature flags are just additional context. You should pick one to three known contexts to start with. In a B2B application, the first context you should accommodate in your flagging system is “Company ID.” The new context is flag state plus company. The myth of tech debt I’m sure you’re asking yourself, “But aren’t you just increasing the technical debt in the system by adding temporary code?” The short answer is…not really. In our experience, when managed correctly, this is not a problem. When feature flags are implemented in a way in which they can be easily removed (i.e. simple conditional statements), they can (and should) be the shortest-lived code sections in the repo. With this in mind, feature flags can actually be a tool that can be used by the development team to reduce overall technical debt. Armed with this new tool, contributors are empowered to improve the code base aggressively.
Risk mitigation
Feature flagging’s top value is as a risk mitigator. They can be used for all sorts of other things, but this should be the top priority. If you have such an amazing codebase that you prefer to fly without this safety net, congratulations, you’ve found the perfect engineering shop and should never leave!!! But in most of the engineering teams I’ve been on this hasn’t been the case. Engineers are usually working in code that has been through several development teams with varying levels of skill and backgrounds. This results in a codebase that is often brittle. This is where proper usage of feature flags can shine. Think of how inexpensive it is to add a flag to mitigate any potential risk to your application. Any side effects can be quickly reverted back to the original behavior with the simple flip of a feature flag.
With the safety net of feature flags in place, you can do the unthinkable…test in production. Feature flags enable you to pick a context (user, company, etc.) and test that hypothesis in a real production environment. The impact would be limited to that context only, and if not, the impact can be quickly disabled without impacting the rest of the team and their deliveries.
Refactors unleashed
With the safety of feature flags, you can take bigger swings especially as it relates to refactoring code. This proves especially beneficial in older code bases where the system may be more brittle. You can refactor a whole vertical slice, deploying frequently and confirming along the way that the change is working in production with little to no impact to your users. If you find that a piece of refactored code that was deployed causes unintended side effects, you don’t have to go through the process of rushing to make a code fix and redeploying your code to production. It’s as simple as turning off the feature flag. You can take the time to mindfully fix the issue and continue refactoring.
Quality mindset / User experience first
Since nothing is now stopping the team from improving the system, quality, and intolerance for degraded user experiences can become the norm. Imagine a world where rapid delivery of value to your users is possible without the historical fear of unintended downtime. Part of the SDLC now involves verifying assumptions in production, constantly maintaining the system, and never disappointing your users. Feature flags can literally let you “swap the engine while driving down the freeway.”
Aggressive/Liberal usage
With all of these benefits, why not aggressively apply feature flags across your code base? The reality is that if the implementation of feature flag usage described here is going to be successful, it must become an enforced standard. It must be a requirement on pull requests that the changes be behind a feature flag. Why does it require that level of enforcement to be successful? Because change is hard, and this change in particular, requires a mindset shift that may not be easy for some engineers. In the same way, writing tests, creating documentation, or following coding standards doesn’t come easy initially.
Implementation examples and guidelines
Identify your flags not just by the change but also by the team and implementer. Start with a common context for your application. In B2B, the context should be the organization identifier. In B2C, geography is a great place to start.
To make sure that this point gets across, we want to repeat: “Feature flags are meant to have a short lifespan.” Ensuring that feature flags are temporary is one of the foundations of implementing the strategies outlined here. Ignoring this foundational topic could lead to situations where you have nested feature flag implementations. This can severely reduce the maintainability of the code base.
public class SampleClass : ISampleClass
{
private readonly IFeatureFlagProvider _featureFlagProvider;
public SampleClass(IFeatureFlagProvider featureFlagProvider)
{
_featureFlagProvider = featureFlagProvider;
}
public async Task DoSomething(string inputValue)
{
if (_featureFlagProvider.IsEnabled(FeatureFlagEnums.FeatureFlag1)
{
if (_featureFlagProvider.IsEnabled(FeatureFlagEnums.FeatureFlag2)
{
/// Some Code Here
}
else
{
/// Some Code Here
}
}
else
{
}
}
}
Figure 1.1 Nested Feature Flags
One of the issues that we encountered pretty quickly when implementing feature flags was merge conflicts. In our initial implementation the FeatureFlags were defined in a single enum class. To solve this issue, the enum class was split into several partial classes with each developer having their own Enum file (FeatureFlagEnums.Dev1.cs,FeatureFlagEnums.Dev2.cs, etc). Within that class file is a partial declaration to the FeatureFlagEnums class where each developer can list the feature flags that they are working on. This gives the developer compile time notifications of potential conflicts.
/// Sample service class
public class CompanyService : ICompanyService
{
public async Task<string> DoSomethingCool(string inputValue1)
{
//Imagine some code here
}
public async Task<string> DoSomethingCooler(string inputValue1)
{
//Imagine some cooler code here
}
}
/// This is a context object
/// It would be used to pass into the feature flag client
public class FeatureFlagContext : IFeatureFlagContext
{
public string CompanyId { get; set; }
}
/// This class would represent your feature flag
/// management. This could be a wrapper around an external feature flag
/// management client such as LaunchDarkly
public class FeatureFlagClient : IFeatureFlagClient
{
public bool IsFeatureFlagEnabledForContext(string contextId, FeatureFlagEnum featureFlag)
{
// Call to your external flag manager here to
// retrieve the flag state for the given context
}
}
/// Pseudo code implementation of the feature flag provider class
/// This IFeatureFlagContext would contain
public class FeatureFlagProvider : IFeatureFlagProvider
{
private readonly IFeatureFlagContext _featureFlagContext;
private readonly IFeatureFlagClient _featureFlagClient;
public FeatureFlagProvider(IFeatureFlagContext featureFlagContext, IFeatureFlagClient featureFlagClient)
{
//The context could be the HttpContext of the session
// (HttpContext.Current) or some other context object.
// The provider will need to account for if your
// context object is null and return the appropriate value
// from the IsEnabled property
_featureFlagContext = featureFlagContext;
_featureFlagClient = featureFlagClient;
}
public bool IsEnabled(FeatureFlagEnum featureFlag)
{
/// Use the context to determine whether the feature is turned on for the specified context
if (_featureFlagContext.CompanyId == null)
{
return false;
}
else
{
return _featureFlagClient.IsFeatureFlagEnabledForContext(_featureFlagContext.CompanyId, featureFlag);
}
}
}
//Once compiled both of these feature flags will be part of the FeatureFlagsEnum object
/// <summary>
/// This specific file belongs to: FeatureFlagsEnum.Dev1.cs
/// </summary>
public static partial class FeatureFlagEnums
{
public const string InternalBugFixIssueFlag = "internal-31119-sample-feature-flag";
}
/// <summary>
/// This specific file belongs to: FeatureFlagEnums.Dev2.cs
/// </summary>
public static partial class FeatureFlagEnums
{
public const string Company321BugFixIssue = "internal-40101-sample-feature-flag";
}
//Use of the feature flag in code would look like this
public class FooService : IFooService
{
private readonly IFeatureFlagProvider _featureFlagProvider;
private readonly ICompanyService _companyService;
public FooService(IFeatureFlagProvider featureFlagProvider, ICompanyService companyService)
{
_featureFlagProvider = featureFlagProvider;
_companyService = companyService;
}
public async Task<string> DoFoo(string inputValue)
{
if (_featureFlagProvider.IsEnabled(FeatureFlagEnums.InternalBugFixIssueFlag))
{
return await _companyService.DoSomethingCool(inputValue);
}
else
{
return await _companyService.DoSomethingCooler(inputValue);
}
}
}
Figure 1.2 Initial Introduction of the feature flag into code.
//Foo service once the feature flag has been removed
public class FooService : IFooService
{
private readonly IFeatureFlagProvider _featureFlagProvider;
private readonly ICompanyService _companyService;
public Foo(IFeatureFlagProvider featureFlagProvider, ICompanyService companyService)
{
_featureFlagProvider = featureFlagProvider;
_companyService = companyService;
}
public async Task<string> DoFoo(string inputValue)
{
return await _companyService.DoSomethingCool(inputValue);
}
}
/// Unused code removed from the company service
public class CompanyService : ICompanyService
{
public async Task<string> DoSomethingCooler(string inputValue1)
{
//Imagine some cooler code here
}
public async Task<string> DoTheCoolestThing(string inputValue1)
{
//Imagine some of the coolest code here
}
}
Figure 1.3 Same code block from Figure 1.1 after the feature flag has been removed. Real World Scenarios
The Big Refactor
If your development team is not made up of AI agents yet, then you’ve heard the phrase “I want to rewrite that whole feature from scratch”. In one of the development teams I worked with, we wanted to remove the use of stored procedures and replace it with an object relational mapping tool (ORM). On the surface it doesn’t sound crazy, until you factor in that the application had over 900+ stored procedures. Where in most development organizations this would be a non-starter, we were able to start work on this immediately. How? Look at Figure 2.1 to see where we started.
/// Parameter Values Class
public class ParameterValue
{
public string ParameterName { get; set; }
public object ParamterValue { get; set; }
public ParameterValue(string parameterName, object parameterValue)
{
ParameterName = parameterName;
ParameterValue = parameterValue;
}
}
/// Generic Data Service
public class DataService : IDataService
{
public async Task GenericQuery1(string inputValue)
{
ExecuteStoredProcedure("GenericStoredProcedure",
new List<ParameterValue>(){
new ParameterValue("Value1", inputValue)
});
}
public async Task GenericQuery2(string inputValue)
{
ExecuteStoredProcedure("AnotherStoredProcedure",
new List<ParameterValue>(){
new ParameterValue("Value1", inputValue)
});
}
private async Task ExecuteStoredProcedure(string procedureName, List<ParameterValue> paramters)
{
/// Code to Execute Stored Procedure against a data store here
}
}
Figure 2.1 - Before any Changes
Nothing super interesting in that code snippet. Your normal boiler plate stored procedure execution. But look at how easy it was for us to start implementing changes to how we are accessing our data with a few lines of code. Look at figure 2.2
/// Generic Data Service
public class DataService : IDataService
{
private readonly IFeatureFlagProvider _featureFlagProvider;
public DataService(IFeatureFlagProvider featureFlagProvider)
{
_featureFlagProvider = featureFlagProvider;
}
public async Task GenericQuery1(string inputValue)
{
// Add a feature flag if statement here
if (_featureFlagProvider.IsEnabled(FeatureFlagsEnum.UseNewQuery1))
{
await GetData(
new List<ParameterValue>(){
new ParameterValue("Value1", inputValue)
});
}
else
{
await ExecuteStoredProcedure("GenericStoredProcedure",
new List<ParameterValue>(){
new ParameterValue("Value1", inputValue)
});
}
}
public async Task GenericQuery2(string inputValue)
{
ExecuteStoredProcedure("AnotherStoredProcedure",
new List<ParameterValue>(){
new ParameterValue("Value1", inputValue)
});
}
/// New method that doesn't use stored procedures
private async Task GetData(List<ParameterValue> parameters)
{
}
private async Task ExecuteStoredProcedure(string procedureName, List<ParameterValue> paramters)
{
/// Code to Execute Stored Procedure against a data store here
}
}
Figure 2.2
Using feature flags we were able to start chipping away at a major refactor while still maintaining the legacy code. By using a context, we could control how many users were executing our new source code. We could deploy several changes with little to no user impact.
2023 Year in Review
My team has seen a lot of changes in the last year. These are things that we didn’t really have in 2022 but are became a part of our day-to-day in 2023.
Feature flags
We started to introduce the concept of flags in late 2022 but didn’t adopt them until 2023. We’ve rewritten the framework a few times. The team has created guidelines for flag creation, management, and removal. We’ve introduced over 200 flags in 2023. The adoption of our feature flag process has led to…
Deploying multiple times per day
In May of 2023, we moved to hourly deploys. We had previously been on a structured 2-week deployment cadence. There are some specific challenges with a 2-week cadence: maintaining the “release branch,” being beholden to the release schedule and work done or not done in time, the fact that we were deploying a bundle of 2 weeks of work, and hotfixes bypassing all the process. We’ve since moved to hourly deploys. We currently deploy on the hour and will be moving to full continuous deployment in January. Production incident remediation times are now tracked in minutes and not hours.
DDOS protection
In 2023, we moved our WAF to Cloudflare. This has given us DDOS protection and a CDN. The DDOS mitigation has proved extremely valuable, as our system has been able to withstand attacks over 10M requests per minute.
WASM
We’ve introduced Blazor to our stack to add frontend code quickly and reliably. We’re using Blazor WASM, which is C# and HTML compiled to WebAssemly. This allows us to use our C# knowledge and best practices (including automated testing) for browser code.
Running on Linux in prod
In the first half of 2023, we migrated our production servers to Linux. In the second half of the year, we migrated our remaining dev and staging servers to Linux. We’ve also migrated our build servers to Linux. These migrations saved costs on the computing side, allowing us to scale up our data side without any overall cost increase.
Latest .NET
Staying on the latest version of the framework is uncommon in most .NET shops. In 2022, we migrated to dotnet 6. In 2023, we’ve done it again and migrated to dotnet 7. In early 2024, we’ll move to the newly released dotnet 8.
Increased automated testing
In August, we increased our expectations around automated testing. We’re now near 40% for total line coverage for all codebases. We’ve adopted behavioral testing across all of the backend code. We’ve introduced Playwright, which allows us to test our frontend code in a more automated fashion.
Codified SDLC
In 2022, our SDLC was very loose and ad-hoc. In 2023, we’ve codified our SDLC. Our SDLC is meant to be flexible while maintaining consistency across the department. Our SDLC guidelines represent sensible defaults, and we hope they will continue to evolve to best serve the teams leveraging them.
Structured teams
At the end of 2023, we had one team of 12, one team of 5, and one team of 2 with QA floating across teams. We’ve since restructured into 3 teams of even size and even staffing.
Job descriptions
I know the engineering team had been working on some job descriptions/matrices, but they never quite made it to fruition. This year, Engineering leadership created measurable job expectations for software engineering levels 1-4. We’ve published these to our team and are using them in our 1:1s and reviews. This gives clarity to both our team members and managers. We’ll be creating similar documents for our managers and QA and DevOps teams in 2024.
Consistent meeting schedule
In addition to the meeting guidelines of our SDLC, we’ve also established a monthly department-wide meeting. This meeting is an opportunity to showcase the great work done each month, share department-level information, and keep each other accountable for our organizational goals.
Company-wide bug reporting
Open bug reporting is a sign of engineering team maturity, and in May of 2023, we opened up our bug reporting process to the whole company. We previously had two competing processes. Not only did this reduce transparency and create confusion, but issues reported in the support team’s system had to be verified and triaged before being added to the engineering backlog. This dual process limited visibility into the bug backlog and also skewed reporting.
This has been one of the most remarkable years of my career. Teams rarely see this much evolution in such a short time. I can’t wait to see what interesting enhancements 2024 delivers.