Hey folks.

After 23 years of fun, I'm leaving Microsoft to try some new things, and therefore this blog will be going into readonly mode.

If you want to continue to read about agile development or design, you might be interested in my new blog:

EricsCodeShack.com

Like the other #no<x> assertions – NoEstimates, NoBugs – I'm really not saying that you shouldn't do TDD. Well, maybe I am…

I was an early TDD advocate, as I really liked the way it helped me organize my thoughts, and I respected some of the other people who were early advocates – people like Ron Jeffries.

But looking back on the 15 years since I started with TDD, I have to say that it really did not live up to my expectations.

What I've seen is a consistent pattern of TDD working in a laboratory setting – developers are quick to pick up the workflow and can create working code and tests during classes/exercises/katas – and then failing in the real world.

My hypothesis for this is very simple. When you look at the TDD evangelists, all of them share something: they are all very good – probably even great – at design and refactoring. They see issues in existing code and they know how to transform the code so it doesn't have those issues, and specifically, they know how to separate concerns and reduce coupling.

If you have those skills, TDD works great; you see what the issues are as the arise and fix them incrementally, and your simple tests prove that your design has low coupling. And, you have the tests to lean on in the future.

A hypothetical

Let's take the TDD workflow and remove the third step – the refactoring step. What would we expect to happen?

Well, we would expect to end up with classes that have multiple concerns in them – because we didn't split them apart – and they would be inconvenient to test. We would need to write a lot of test code, and would need a lot of help to write the code, either creating many hand-written mocks or using mock libraries.

Which I submit is precisely the result that most developers get using TDD; the resulting code looks exactly like what we would expect if developers are skipping the third step. Which isn't really very surprising, given that the tests in most non-TDD code look the same way, and we know that most developers do not have great design/refactoring skills.

At this point we should ask ourselves, "are these developers getting a net positive benefit from using TDD?" Let's list the pros/cons:

Pros

Cons

I think that's a good summary of why many people have said, "TDD doesn't work"; if you don't have a design with low-coupling, you will end up with a bunch of tests that are expensive to maintain and may not guard against regression very well.

And to be fair, the level of design and refactoring skills required is relative to the state of the codebase you're working in. There are a lot of legacy codebases with truly monumental amounts of coupling in them and figuring out how to do anything safely is very difficult and complex.

TDD—, Refactoring++

Instead of spending time teaching people TDD, we should instead be spending time teaching them more about design and especially more about refactoring, because that is the important core skill. The ability to write and refactor code to a state with low coupling, well-separated concerns,  and great names will be applicable in all the codebases that we work in.

That is not to say that I don't dust off Ncrunch now and then to either create a new set of functionality with great tests or to create tests around a concern as I do a refactoring, because I still do that often. I just don't think it's the right place to start from a educational or evangelical perspective.

Code/Test/Code/Test/Code/Test/Code/Test

Before I dig into things, I have a bit of pre-work for you. I would like you to spend 5 minutes listing all the things that you (or your team) do/does as part of the inner loop, and then create a value stream map from that information. Put numbers on all of the steps, but don't worry if they aren't particularly accurate. You can also use a range like 5-10 minutes.

Do you have a diagram that describes your current process? Great.

Here's the first diagram I came up with:

The times are going to vary considerably for a given team and a given developer. Let's expand a little bit on the diagram to add our first handoff. This time, because we're only talking about a single developer, it's a different kind of handoff.

When we start a build, we are handing off the code to another process –the compiler – and we will need to wait until that process is done. And no, that 5-60 minutes is not a misprint; I've worked on a number of teams where a full build was at least an hour.

If you are thinking, "self, that looks like a queue/handoff and a context switch, and that is bad". If so, give yourself a gold star and a hearty "Well done!". Let's flesh this out a bit more….

Before we can test or debug our code, we need to deploy it to a place where we can test or debug it. If we are running unit tests, we don't have a deploy step, which is why it sometimes takes 0 minutes.

Finally, sometimes we are successful, and sometimes we are not. That needs to be represented in the diagram as well…

If our build failed or our test/debug failed, we need to go back and tweak the code. If it worked, then we need to figure out the next thing to do.

Time to search for bottlenecks. Where do we look?

The usual places; we look for handoffs and we look for queues. Waiting for the build and waiting for the deployment are excellent candidates. Before we dig in there, I want to go on a bit of a tangent…

In earlier posts, I talked in depth about developers switching to do something else and the costs of that. I'd like to quantify that using Eric's Theory of Developer Wait Times.

It sounds more grandiose than it really is…

This is basically just a repurposing/reexpression of the research around how users react to delays in programs.

Once we get above 2 minutes, it's time to read email, take a nature break, get some water, read the Facebooks, etc. Because it's a hard interruption, it's really easy for 3 minutes to turn into 15 minutes.

Since developers spend a lot of time waiting for builds, the majority of the teams out there put a high priority on having fast builds; they give developers capable machines on which to do builds, they architect their projects so that components are small and builds can therefore be fast, and the focus on fast-executing unit tests rather than slower integration tests.

Ha Ha. I make joke.

Remember the discussion about how local optimization can de-optimize other parts of the system from the last post? Here's another example.

The amount of time I waste waiting for a build is determined by how long the build takes and how many times I need to wait for it. If my build takes 10 minutes, I will look to batch my work so that I minimize how much time I spend waiting for builds. That means bigger checkins, with all the downsides that come from them; mistakes are more common, design feedback is less likely to acted upon, etc.

Is this obvious, or even noticeable? If you ask most dev leads about the impact of build speed on their team's productivity, they will say something like, "Our build is slower than we would like, but it's not a large drain on our productivity". They say this whether their average build takes 30 seconds or their average build takes 15 minutes. That is based on an estimate of the "time spent waiting for builds" / "total time" metric. Developers will rearrange how they do things (ie "optimize") so that the metric doesn't get too far out of whack in most cases.

What they are missing is the opportunity cost of having slow builds.

• Slow builds means no TDD. I am fairly stubborn and I've tried to modify TDD to make it work when I have slow builds, but it isn't really effective, and there is no way I can convince anybody else to try it.
• Slow builds mean I (probably) can't leverage any of the  continuous testing tools such as NCrunch or dotCover in the C# world or Infinitest in the Java world. Ncrunch is why there is a "none" entry in the wait time table; you just write your product code and your test code and everything happens in the background, with no wait for compile or deploy.
• Slow builds mean my team will write fewer tests.
• Slow builds mean my team will do fewer refactorings.

In lean terms, slow builds are a constraint, and the developer optimization is an attempt to subordinate the constraint. Which is fine as long as you exploit and elevate the constraint to try to remove it as a constraint.

If you can only do one thing, work on getting your developer builds faster.

Having a short deployment cycle is also important, but if your builds are quick and you write unit tests, you will be spending a lot more time in that cycle and less time in the deployment cycle. For deployment I think that simple and automated is more important than pure speed, because mistakes you make in deployment kill your overall cycle time; just forgetting to deploy one changed binary can cost you anywhere from 15 minutes to a couple of hours. On the other hand, if you can make deployment very fast as well, nobody will be tempted to build a faster version of it, which has a lot of other benefits.

Figure out what code to write

We will now move the fill in the left side of the diagram, the part where we figure out what code to write.

There are three ways we figure out what code to write:

1. Sometimes we just think about it, or we already know pretty much what we want to do.
2. Sometimes we have to do some research to figure out what we need to do and/or how to do it.
3. Sometimes we ask somebody.

Where are the bottlenecks here?

One of them is obvious – it's the time that we spend finding somebody to ask and/or the time we spend waiting for an email response to the a question. Unless it's a person-to-person interaction, that part generally takes long enough that we "go do something else".

So, we should just try to limit that path, right?

Well, nobody likes to interrupt other people to find out something you could have found out yourself, and in many (most?) teams there are social pressures that keep you from choosing that branch first, so they choose the upper research branch first.

The flow in that branch typically looks something like:

1. Need to do <x>
2. Do a web search on <x>
3. Read a few articles
4. Go write some code
5. Can't figure out on of the samples.
6. Go back and read some more
7. Modify the code.
8. Build/deploy/test the code
9. Mostly works, but not sure
10. Go back and read some more
11. Try something else
12. Code works, go do something else.

Is that the most efficient way to do things? Or, to put it another way, if we wanted to make this faster, what would we do?

Well, the obvious observation is that our build/test/deploy time has a big effect on the cycle time here; that was one of the reasons I talked about that first. If you can experiment quickly, you can be a bit less sure of what you are doing, spending less time researching and more time doing.

But there's a larger thing at play here.

Our goal is obviously to get to the exit arrow as quickly as possible (assuming good testing, good quality code, etc. ). To do that, we have to write the code that is correct.

So, to state it another way, our bottleneck is figuring out what the right code is. If we can do that quickly, then the coding and testing part is generally pretty straightforward.

How do we become more efficient at that? How do we get rid of the waste?

The answer is simple. The biggest waste is when we go down the wrong path; we don't understand the code correctly, we don't understand what we are trying to do, or we use a more cumbersome way of solving a problem. How do we solve that? Well, we need to apply more minds to the problem.

Some groups have an "ask for help" rule, which says if you have been stuck on a problem for more than 15/30/60 minutes, you are required to ask for help. Which is a good idea, but you've already wasted a lot of time, nobody likes to ask for help, and you have to interrupt somebody else to ask, slowing them down.

What we really need is a way to dedicate extra brainpower to a task from the start. By doing so, we will reduce false starts, research flailing, waiting too long to ask, etc. – and we will write the right code faster.

And, once again, we've invented pairing.

Pairing is more efficient because it directly targets the waste in the system:

• Time spent researching how the system works
• Time spent researching options
• Time spent modifying code when it doesn't work
• Time spent doing more research when we could have just asked somebody.

Why don't teams try to become more efficient at writing good code?

Well, that ties back to one of the weird parts of the profession.

Writing software is a very complex task, especially if you are doing it in a codebase that has a lot of technical debt. Because of that, it is very hard to estimate how long a given task should take, which – transitively – means that it is often hard to determine whether a task is taking longer than it should. We don't want to ask developers why a task is taking so long because that encourages the wrong behavior and we know that sometimes things just take a long time to do.

Pairing gives us a way to hop inside that complex task and get rid of the waste without having to ask the wrong question or mess up the group dynamic. That is one of the reasons it is so powerful.

That is all for this post, and that takes me to the end of where I envisioned the series. I do think there might be a short summary post to wrap things up.

Checkin Queue

1. Developers have to figure out something else to do while they wait, which involves a context switch and/or "looking busy".
2. Because of the time lag, there is possibility that somebody on your team will have their checkin finish before yours, and if their changes are incompatible with yours, you will have to wait.
3. Because the wait feels like wasted time to the developers (it is…), they will try to optimize by making their submissions bigger, which makes their code reviews bigger and harder to understand. None of this is good.
4. If the infrastructure breaks, we are dead in the water until it is fixed.
5. The "run tests" step is a magnet for more tests; if bugs get through the box, it's very tempting to add tests. That is *great* if they are unit tests, and horrible if they are integration tests, as the time to run the tests will get much slower.
6. The attitude is "the system is supposed to find the errors", which makes it easy to be sloppy and just submit things.
7. Failures are generally private, and not considered to be that important.
8. If there is a big failure that makes it though even though it shouldn't have, you need to stop the whole system and block everybody until you get it fixed.

1. We have gotten rid of the queue, which saves time
2. Developers can work in smaller chunks. There is great synergy with pairing and working in small increments.
3. Failures are obvious, which puts social pressure on developers to be more careful.
4. The system rewards the behavior you want to incentivize; it encourages developers to be a little more careful, it encourages them to run tests on their computers (and therefore encourages them to keep those tests simple and fast).
5. It treats developers like adults who can make rational choices around what needs to be done before a checkin is made.

Searching for the bottleneck

1. They are always busy.
2. Work piles up in front of them.
3. Downstream resources are regularly idle.

1. This one works well with machines, but people are another matter. Not only can we easily look busy, we might be busy with low-value work.
2. Work piling up in front of them is a pretty good indicator. This is another way of saying, "examine the queues and find the ones that are the longest".
3. A machine that is after a bottleneck will look idle. As I noted in #1, humans are great at looking busy.

1. The time spent in the incoming queue.
2. The time to perform the acceptance test.
3. The time spent in the "redesign" and "fix bugs" queues (queues that are absent in the current map).
4. The time for the redesigned and fixed feature to wind its way through the process again.
5. The time spent in the incoming queue for the second time
6. The time to perform the acceptance test.
7. The time to repeat steps 3-6 if the feature still isn't right.

• Embedding a product owner in the team
• Having defined and regular interaction between the product owner and the developers
• Doing feature design with the product owner and the whole team to make sure everybody is on the same page
• Writing explicit acceptance criteria for features ahead of time.
• Using ATDD with tools like Cucumber/Gherkin/Specflow.
• The best solution is going to depend on your organization and how you structure development. I recommend focusing on a shared understanding of scenarios and acceptance criteria up front, as that will both prevent issues and streamline acceptance testing by the team. One more thought on this area; the situation here is an example of an handoff between teams. Handoffs are a great place to look for waste, because:
  • They aren't under one team's control, which means it would take cross-team effort – which most companies are poor at – to improve them
  • The incentives in place are usually on a per-team basis, and therefore they reinforce an "us vs. them" perspective.
  • Politics often show up.
  • There are often fixed contracts/SLAs
  In other words, it's a lot of effort to try to fix them from within a team and the success rate isn't great, so they persist. In lean terms, the teams on both sides will consider the handoff to be a constraint, and will subordinate (ie de-optimize) the rest of their process to try to deal with it as best they can. It is therefore very important to ask, "what would happen if this handoff did not exist?"

  Moving upstream

  Moving upstream, we next come to the test box. It has a lot in common with the owner acceptance scenario, but in thinking about it I realized that I'm missing some important detail in the drawing. For sake of discussion, I've collapsed the whole upper part of the diagram into a simple "Design & Code" box. Not only is there a queue before test, there are queues on the output of test. The first queue sits before the triage process, where we spend time talking about how bad each bug is and whether we want to expend time fixing it. This is typically done by managers and other senior (read as "expensive", both in monetary terms and opportunity cost terms) people. I put "1-30 days" in front of that as a guess, but in my experience triage happens every week or every few weeks. After that, bugs move into another queue, which is typically called the "bug backlog". The reason that the label says "1-N" is that the bug backlog is where many bugs go to die; unless you are fixing bugs faster than you are creating them, this list is going to continue to grow and you therefore will not fix all of the bugs. I missed another feature of the diagram: Because bugs come in faster than you fix them, the bug backlog will grow to be too big, so there is often a "retriage" step, where you look at the bugs you already triaged and get rid of the lower-priority bugs. This is an example of rework of the rework. What can we say about the effect of this on the time it takes a feature to flow through the system? If you agree that a bug is an indication that a feature has not met acceptance, then if you have bugs against that feature you are not done, and therefore most of your features are never done; you may have shipped them, but you are still engaged in rework on them after you ship them. How did we end up here? Well, this seems like a good time for a story that I call "FeatureLocks and the Three Groups"... FeatureLocks desperately wanted to go out into the world, but he was held up because he had bugs. He went and talked to the developers, and they said, "FeatureLocks, it's not our fault, it's the test team's job to find bugs, and they are too slow at it". Then he went and talked to the testers, and they said, "FeatureLocks, we are sorry there are bugs, but we are always busy testing and the developers are just writing too many bugs". Then he went to the triage group, and they said, "FeatureLocks, we know that you want to go out into the world, but you have 10 priority 1 bugs logged against you and you'll just have to wait until they are fixed". FeatureLocks was so sad. His goal in life was to deliver value, but he couldn't do that until he worked well and could visit customers. They did not live happily ever after. At this point, I feel compelled to talk a bit about charter. If we want to improve our development process, we need a set of goals. Something like:
  • Decreasing the time it takes features to move through the system
  • Decreasing the time it takes to fix bugs
  • Decreasing the number of bugs
  • Identifying more bugs earlier in the process
  • Reducing the average bug age
  Who owns the charter to make things better? Upper management certainly owns the overall charter, but they aren't really close enough to the problem, so they delegate it to the individual team managers. That doesn't work because each of the group only owns a part of the problem, and the handoff/interactions between groups aren't really owned by anybody. Which means we really have two high-level problems; the first is the one we identified in the diagram - we have queues and lots of rework - and the second is that nobody owns the overall problem. You have probably figured out that this is a long discussion that will lead us towards what is sometime called "combined engineering". We take the dev and test teams and we combine them into a larger "engineering" team. What do we get out of this?
  • Since we don't have a handoff between teams, our queue between coding and testing will be smaller. More importantly, the way that process works is owned by one team, which means they can improve it.
  • It is in the new team's best interest to make bugs easier to find and fix.
  My experience is that combined engineering is a significant improvement over individual teams. There is a natural extension to combining dev and test; the "devops" movement includes operations functions such as deployment into the same team, and has the same sort of benefits. The meta point – which I stole directly from the Theory of Constraints – is to always look at the interfaces between teams, because those are places where inefficiencies naturally accumulate. That pretty much covers the bottom part of the diagram; in my next post I'll be tacking the design / code part.

  Part 3: The Development Team (3)

Adding the Queues

1. The input queue, which I'm going to ignore for now.
2. Design Review: After I have finished a design, I send it out to the rest of the team for review.
3. Code Review: After I have finished the implementation, I send out the code to the team for review.
4. Code Submission: I submit my code to an automated system that will run all the tests and check in if they all pass.
5. Test: The feature moves to the test phase. This might be done by the development team, or there might be a separate test team.
6. Acceptance: Somebody - typically a product owner - looks at the feature and determines if it is acceptable

1. Design Review: 3 hours to 1 day.
2. Code Review: 3 hours to 1 day.
3. Code Submission: 30 minutes to 1 day.
4. Test: 1 day to 10 days
5. Acceptance: 1 day to 10 days.

• Unreadable code.
• Context dependent code - code that does not have fixed behavior but depends on things elsewhere in the system
• Communication issues between humans. A big one here is the lack of a ubiquitous single language between customers to code; he cited several examples where the name of a feature in the code and the customer visible name are different, along with a number of other issues.

• We've always done it this way
• I'm busy
• That won't get me promoted
• That's not my job

1. Extract system code into a separate method
2. Make that method static
3. Move that method to a new class
4. Change the signature of the method to take a parameter of it's own type.
5. Make the method non-static
6. Select the constructor for the new class in the caller, and make it a parameter
7. Introduce an interface in the new class
8. Modify the new class parameter to use the base type (interface in this case).

1. Figure out what a method does
2. Write a test for it
3. Repeat

• There is a lot of test code that takes time to write
• Because of the high coupling, changes in many areas of the system will require changes in the test
• The test is very hard to understand, and it's often not clear whether the test is actually testing what it says it is testing
• Developers do not feel like they are being successful in writing unit tests.

EmployeeSource class

public static EmployeeCollection GetEmployees(EmployeeFilter employeeFilter, EmployeeSource employeeSource)
{
    var employeeCollection = employeeSource.FetchEmployees();

    return employeeCollection.Filter(employeeFilter.Matches);
}

return employeeSource.FetchEmployees().Filter(employeeFilter.Matches);

var collection = employeeSource.FetchEmployees().Filter(employeeFilter.Matches);

WriteToConsole(collection);

• Creates a chunk of data
• Performs an operation on that chunk of data to create a second chunk of data
• Consumes that chunk of data

public class Pipeline
{
    public static void Process<T1, T2>(Func<T1> source, Func<T1, T2> processor, Action<T2> sink )
    {
        sink(processor(source()));
    }
}

Pipeline.Process(
    employeeSource.FetchEmployees,
    (employees) => employees.Filter(employeeFilter.Matches),
    WriteToConsole);

Pipeline.Process(
    employeeSource.FetchEmployees,
    employeeFilter.Filter,
    WriteToConsole);

• I don't have a way to test my EmployeeSource class, so there is probably a trip to P/A/S coming up.
• I'm not really in love with Pipeline; I have an alternate approach in mind that I think will be better.
• The implementation of EmployeeSource has that ugly retry logic in it; that is something algorithmic that I hope can be teased out of the fetching.

1. Enter Idea
2. Triage by priority
3. Feature estimation
4. Release planning
5. Create
6. Ship

Enter idea

Traige by priority

Feature estimation

Release planning

Create

Ship

Where's the bottleneck?

• The later a change request is injected in the cycle, the closer it is to shipping and therefore the shorter the time period is before it will show up as value to our customer, so we want to get the change requests in late.
• The later a change request is injected in the cycle, the more disruptive it is to the development process, and the higher the priority that the ship date will slip, the change request will not be well-implemented, and the overall product will suffer more quality issues, so we want to get them early.

Quality

Talkin' 'bout Optimization

Risk

• Schedule risk - the risk that we will not be able to finish the project on the estimated schedule
• Market risk - the risk that the features that we developed are not longer relevant to our market.

• They aren't quite done with their features.
• They have a lot of bugs to fix
• All their processes were designed to work in a two-month-long ship step.
• They have absolutely no idea how to ship in two weeks.

1. Delay shipping.
2. Cut corners