Two people are talking about agile stuff…

Person 1: So, what kind of process do you use in your group?

Person 2 (apologetically, hanging her head low): Well, we’re not very agile. We are doing iterations and retros, but our planning isn’t very good, we don’t pair, and we rarely do TDD.

I’ve seen this interaction quite a few times, and I’ve finally decided that I need to say something.

Stop apologizing about how agile you aren’t

Seriously. Yes, I know that it’s frustrating to run into people who are more experienced and/or in environments where they can try things that you haven’t/can’t.

Knowing what those groups do is great, as it gives you data about what you might want to try in the future. . *Comparing yourself* to those groups is not. Focus on the positive and the achievable within your own environment. If you are experimenting and making progress, you are doing the right thing.

Welcome to Eric’s Code Shack.

New posts will be coming as soon as I find some free time to write.

For the time being, I recommend viewing some classic posts from the past…

In this post, I’m going to talk about what I’m calling the inner loop, what some people call “Ring Zero”; it is basically the simple act of writing code and validating it, over and over. It is focused on a single developer.

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…

 

Wait time	Impact
None	No impact
<15 seconds	Slight impact on flow
60 seconds	Some context lost. Annoying.
2 minutes	Most context lost; hard to keep focus
> 2 minutes	I’m doing something else

 

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.

 

 

Finally, we make our way to the heart of the development team and the design & code phase.

 

That is the top part of this diagram.

The design/review/finish and code/review/finish chunks are very similar. The developer does some work, submits it for review, perhaps does rework based on the review, and then finally finishes and moves onto the next step.

As is common with an overall map, we’re some important details:

In this detailed view, we see that the “Submit Review” step leads to a queue in front of three other developers, which then leads back to a queue for the original developer.

Let’s explore a different perspective, switching from a value stream view to a sequence diagram. This diagram shows what each person is doing during a typical code review cycle:

I tried to pick what feels like a reasonable example; of course there are some reviews that are simpler and some that are much more complicated.

What we see here looks very nice from a utilization perspective; everyone is busy all the time, and the time from submission to finish for Miss Blue looks pretty good.

This diagram is missing a bit of detail, so let’s add that in:

What we were missing is what the manufacturing world calls “setup time”, which is the time it takes to switch a machine from doing one task to doing another. Obviously, this time is time that the machine is not doing useful work, and therefore reducing the time spent on setup is a major factor in optimizing throughput (article).

In our world, we call this period a context switch. This isn’t a strict downtime, but there is a time of reduced efficiency, where developers are both less productive and more likely to make mistakes. How long it is depends upon the complexity of the mental model required; it might be only a couple of of minutes for a simple task, or it might be thirty minutes for a complex debugging scenario. Every time we switch from working on one thing to another, there is a loss of productivity and quality.

Note that this is the diagram for one review; the actual sequence diagram for a developer can be much worse; there may be multiple reviews going on at the same time, and there are other interruptions as well; meetings, breaks, and the all important lunch.

Looking at the diagram, the queues show up as lag time between when Miss Blue submits her code/design for review, and when the other developers start to review it. If we we can reduce that lag time – reduce the time spent in the queue – then we will improve the throughput of the system. So we  set up a policy that doing code reviews is our highest priority.

If you’ve made the choice in the picture, you know what happens; the process that you set to high priority runs faster, and you slow down the rest of the system. If you are lucky, you might even hang the system.

The same thing happens in our process; when a code review shows up, it gets handled quickly, which would seem to be what we want. But it also shifts the required context switch from “when I have some free time” to “right now”, and that change makes the interrupt more costly. The teams I’ve been on that tried this absolutely hated it.

Is there another solution? Well, we know from our earlier examples that getting rid of a queue is far better than just reducing its time – especially if there is a handoff, as we have here – so how can we get rid of a queue?

Well, perhaps we can limit the number of developers who are required for a given code review, so that there is only one review queue per task. It would look something like this:

That is better from an interruption perspective, but we probably lose some quality, Mr. Red has more time and feels more responsible to do a good code review, but he doesn’t have any more context than he did before, and he’s still going to feel that the time he spends on code review is taken away from his “real work”.

How can we improve that? What if he worked with Mrs. Blue during the design and coding, so that he understands *deeply* what is being done and can give immediate context. That would change the review process to be something like this:

And yes, we have invented pairing. One of the benefit of pairing is obvious if you look at the sequence diagram and the long periods of work devoted to a single task; that is clearly going to have less waste than the previous diagram.

The groups that I have been on that paired heavily ended up with a slightly different workflow; when the code is done, a code review is always sent out – so that others can see what is going on if they wish – and the pair has the option to ask for review from somebody else if they think they need it.

This has worked very well in the groups that I’ve worked in.

 <aside – the hidden cost of context switches>

If you are currently in a world like the one I described – one with lots of context switches – you might be saying to yourself, “self, I can see how it would be nice to have fewer interruptions, but I’m pretty good at context switching and still being productive.”

Such a feeling is nearly universal, and almost always, it is wrong. When we are in a world that requires context switching – especially one where those context switches are higher priority – it’s hard not to be in a reactive, firefighting mode. What I sometimes call tactical mode.

If you are in tactical mode, it’s very hard to engage your thoughts around bigger strategic questions, such as whether the approach that you are taking is an efficient one. Instead of “being productive” meaning “making good choices about how things are done so as to maximize output”, it becomes “responding quickly to interrupts and doing a decent job of juggling all the things on my plate”.

Or, to put it another way, in a results-driven, context-switching environment, the chance that you are spending any time at all thinking about your efficiency is pretty low.

To go back to my processor/operating system analogy, everybody knows what happens to background tasks when your CPU is at 100%, and strategic level thinking is a background task.

I honestly think that if you can get rid of code review interruptions, that alone saves enough time to make pairing equal to single-developer work, without going to any of the other advantages.

</aside>

Checkin Queue

Finally, we get to the checkin queue, where you submit code into a system that builds it runs all the tests, and then checks the code in if it was successful. This is often known as “gated checkin”.

Let’s add in some of the missing detail.

When we submit our code, we jump into a queue where a separate machine will get the changes, make sure that everything builds, run the tests, and then either check the code in if it works or send an email to us telling us what the problems are.

Let’s analyze the situation here…

First, I see that there is a queue for the developer to wait for this process to happen. That will require them to figure out something else to do during this time period. The wait is going to depend on how big the queue before “build” is, and how long the build and test run is going to take.

What effects does this approach have on the organization?

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.

The incentives typically push these systems into becoming bottlenecks, and they also lock the behavior of the team around the system; the team cannot experiment in this area because the system prevents it. My experience is that this can work okay with small codebases, but it rarely turns out good in larger codebases.

Is there an alternative that gets rid of the bottleneck? Consider the following:

When the developer is finished, they check in their changes. There is a separate build/test machine that notices the changes, does the build, and runs the tests. If things are okay, it doesn’t do anything. If there is a failure, it emails the team (and in some teams, turns on a flashing red light in the team room).

What are the differences?

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.

The one downside of this system is that the build is sometimes broken. The teams I’ve worked on have adopted a “5 minute rule”; if you can fix the breakage in 5 minutes – say you just forgot to include a new file when you checked in – then you are allowed to do it. If it’s going to take longer than that, you revert your checkin so that nobody else is blocked.

Which finally brings us to the end of the developer team section. Up next… the individual developer…

Next: Part 4 – The Inner Loop

(Note: The first version of this was a very random draft rather than the first part that I wrote. I blame computer elves. This should be a bit more coherant)

This episode of the series will focus on the development team – how a feature idea becomes a shippable feature.

A few notes before we start:

• I’m using “feature” because it’s more generic than terms like “story”, “MVP”, “MBI”, or “EBCDIC”
• I picked an organizational structure that is fairly common, but it won’t be exactly like the one that you are using. I encourage you to draw your own value stream maps to understand how your world is different than the one that I show.

In this episode, we will be looking at the overall development team. I’m going to start by looking at a typical development flow:

Green lines are forward progress, red lines show that we have to loop back for rework.

That’s the way it works for a single developer, and across a team it looks something like this:

I’ve chosen to draw the case where features are assigned out by managers, but there are obviously other common choices. Hmm… there are already a ton of boxes in the diagram, and this is just the starting point, so I’m going to switch back to the single-developer view for now.

What are we missing?

Adding the Queues

There are several queues in the process:

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

Now, let’s put some times on the time spend in the queue. The numbers I’m listing are from my experience for a decent traditional team, and they are typical numbers.

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.

Here’s an updated diagram with the numbers on it:

At this point, we would typically try to put numbers on all of the blue boxes, but because our features sizes vary so much, the numbers were all over the place and weren’t very useful.

We can, however, try to put some numbers on the red rework lines. I’d like you to think about what the numbers are in your organization, and we’ll pick it up from there.

Part 3: The Development Team (2)

Agile Open Northwest uses a different approach for running a conference. It is obviously around agile, and there is a theme – this year’s was “Why?” – but there is no defined agenda and no speakers lined up ahead of time. The attendees – about 350 this year – all show up, propose talks, and then put them on a schedule. This is what most of Thursday’s schedule looked like; there are 3 more meeting areas off to the right on another wall.

I absolutely love this approach; the sessions lean heavily towards discussion rather than lecture and those discussions are universally great. And if you don’t like a session, you are encouraged/required to stand up and go find something better to do with your time.

There are too many sessions and side conversations that go on for me to summarize them all, but I’ve chosen to talk about four of them, two of mine, and two others. Any or all of these may become full blog posts in the future.

TDD and refactoring: Did we choose the wrong practice?

Hosted by Arlo.

The title of this talk made me very happy, because something nearly identical lived on my topic sheet, and I thought that Arlo would probably do a better job than I would.

The basic premise is simple. TDD is about writing unit tests, and having unit tests is a great way to detect bugs after you have created them, but it makes more sense to focus on the factors that cause creation of bugs, because once bugs are created, it’s too late. And – since you can’t write good unit tests in code that is unreadable – you need to be able to do the refactoring before you can do the unit testing/TDD anyway.

Arlo’s taxonomy of bug sources:

• 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.

I think that the basic problem with TDD is that you need advanced refactoring and design skills to deal with a lot of legacy code to make it testable – I like to call this code “aggressively untestable” – and unless you have those skills, TDD just doesn’t work. I also think that you need these skills to make TDD work well even with new code – since most people doing TDD don’t refactor much – but it’s just not evident because you still get code that works out of the process.

Arlo and I talked about the overall topic a bit more offline, and I’m pleased to be in alignment with him on the importance of refactoring over TDD.

Continuous Improvement: Why should I care?

I hosted and facilitated this session.

I’m interested in how teams get from whatever their current state to their goal state – which I would loosely define as “very low bug rate, quick cycle time, few infrastructure problems”. I’ve noticed that, in many teams, there are a few people who are working to get to a better state – working on things that aren’t feature work – but it isn’t a widespread thing for the group, and I wanted to have a discussion about what is going on from the people side of things.

The discussion started by talking about some of the factors behind why people didn’t care:

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

There was a long list that were in this vein, and it was a bit depressing. We talked for a while about techniques for getting around the issue, and there was some good stuff; doing experiments, making things safe for team members, that sort of thing.

Then, the group realized that the majority of the items in our list were about blaming the issue on the developers – it assumed that, if only there wasn’t something wrong with them, they would be doing “the right thing”.

Then somebody – and of course it was Arlo – gave us a new perspective. His suggestion was to ask the developers, “When you have tried to make improvements in the past, how has the system failed you, and what makes you think it will fail you in the future?”

The reality is that the majority of developers see the inefficiencies in the system and the accumulated technical debt and they want to make things better, but they don’t. So, instead of blaming the developers and trying to fix them, we should figure out what the systemic issues are and deal with those.

Demo your improvement

Hosted by Arlo.

Arlo’s sessions are always well attended because he always comes up with something interesting. This session was a great follow-on to the continuous improvement session that I hosted.

Arlo’s basic thesis for this talk is that improvements don’t get done because they are not part of the same process as features and are not visibly valued as features.

For many groups, the improvements that come out of retros are either stuck in retro notes or they show up on the side of a kanban board. They don’t play in the “what do I pick up next” discussion, and therefore nothing gets done, and then people stop coming up with ideas it seems pointless. His recommendation is to establish second section (aka “rail”) on your kanban board, and budget a specific amount to that rail. Based on discussions with many managers, he suggested 30% as a reasonable budget, with up to 50% if there is lots of technical and/or process debt on the team.

But having a separate section on the kanban is not sufficient to get people doing the improvements, because they are still viewed as second-class citizens when compared to features. The fix for that is to demo the improvements the same way that features are demo’d; this puts them on an equal footing from a organizational visibility perspective, and makes their value evident to the team and to the stakeholders.

This is really a great idea.

Meta-Refactoring (aka “Code Movements”)

Hosted by me.

In watching a lot of developers, use refactoring tools, I see a lot of usage of rename and extract method, and the much less usage of the others. I’ve have been spending some time challenging myself to do as much refactoring automatically with Resharper – and by that, I mean that I don’t type any code into the editing window – and I’ve developed a few of what I’m thinking of as “meta-refactorings” – a series of individual refactorings that are chained together to achieve a specific purpose.

After I described my session, to friend and ex-MSFT Jay Bazuzi, he said that they were calling those “Code Movements”, presumably by analogy to musical movements, so I’m using both terms.

I showed a few of the movements that I had been using. I can state quite firmly that flipchart is really the worst way to do this sort of talk; if I do it again I’ll do it in code, but we managed to make it work, though I’m quite sure the notes were not intelligible.

We worked through moving code into and out of a method (done with extract method and inlining, with a little renaming thrown in for flavor). And then we did a longer example, which was about pulling a bit of system code out of a class and putting it in an abstraction to make a class testable. That takes about 8 different refactorings, which go something like this:

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).

Resharper is consistently correct in doing all of these, which means that they are refactorings in the true sense of the word – they preserve the behavior of the system – and are therefore safe to do even if you don’t have unit tests.

They are also *way* faster than trying to do that by hand; if you are used to this movement, you can do the whole thing in a couple of minutes.

I asked around and didn’t find anybody who knew of a catalog for these, so my plan is to start one and do a few videos that show the movements in action. I’m sure there are others who have these, and I would very much like to leverage what they have done.

 

I’ve been working on a talk on ways to make unit testing easier. I has not been going well; I’d come up with an approach I liked, do most of the slides for it, come back to it, and be unhappy with what I had written.

This happened – and I am not exaggerating – 4 times in a row.

In the 5th try, as I was working through the techniques I was going to talk about, I realized something. But let me back up a bit first…

Pretty much every introduction for unit testing starts with a very simple scenario using a very simple class; the flow is something like:

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

Or, if you are doing TDD, you swap the order and write the test before the method.

With a small class and small problem space, this works well; it’s simple for developers to understand, and you can therefore present it to a group and they walk out thinking that they understand unit testing.

There is one *tiny* problem, however.

It doesn’t work on real classes – and by that, I mean the classes that are in most real system. We all know that the testability of existing codebases is low, and we also know that most developers do not have the design skills or experience to take that sort of code and write good unit tests for it.

But that is what developers try to do, because THAT IS WHAT OUR INTRODUCTION TOLD THEM UNIT TESTING IS ABOUT.

So, they take their existing code, add in a bunch of interfaces so they can inject dependencies, pull in their favorite mocking library, shake it around a bit, and end up with a unit test.

Hmm…

• 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.

And – AND – there is very little chance of the test driving improvements, which is one of the main reasons we are advocating for a unit-testing approach.

We have been going about this in the wrong way.

We should focus on teaching developers how to look at code and figure out what different units are lurking in a single class, and also teaching them how to extract those units out so that simple tests can be written for them.

And… we should lighten up on the “you should write tests for everything”, because these expensive complex tests aren’t doing anybody any good.

Welcome back to You Suck at TDD. Today’s code will show up in the Improvements-Phase-3 branch if you would like to follow along.

In our last episode, we concentrated mostly on the Employee fetching and filtering. Things are better, but we still have a problem…

Well, actually, we have a number of problems, but we’ll start with the first one that I see, which is that Yucky.GetEmployees() is trying to do too many things – it both fetches data and filters it.

This is one of the most common things I see when I look at code – methods that try to do too much, and they are generally quite challenging to test because they do too much.

EmployeeSource class

So, we’ll start by working on that. I pulled the fetching code out into a new EmployeeSource class and then I pulled the creation of the EmployeeSource instance out of GetEmployees().

At this point, GetEmployees() only has a couple of lines in it:

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

    return employeeCollection.Filter(employeeFilter.Matches);
}

We can easily simplify it with an inline of employeeCollection:

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

At this point, there is no reason to have a separate method, so I inlined FetchEmployees() and did some cleanup in Main(), and deleted class Yucky.

This does make Main() more complex than I would like it, but I did the inline with a purpose; I find that my thoughts are often constrained by the abstractions that are present, and the act of getting rid of a bad abstraction sometimes makes it easier to find the right abstraction.

Spending a little time with Main(), it isn’t testable code, and I’d like to get some tests in there…

If we abstract out writing the output to the console, we get something like this:

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

WriteToConsole(collection);

That’s decent code; since it doesn’t have any conditionals, it’s likely code that either works all the time or never works at all. Not my first target when I’m looking to make things more testable, but if we wanted to cover it, how can we do it?

Well, the most obvious thing to do is to extract the IEmployeeSource interface, create a method that calls that, create a mock (or a simulator using P/A/S), and then write a test or two. That adds some complexity; you have another interface hanging around, and you need to create and maintain the mock and simulator. It will definitely work, but it is a fairly heavyweight approach.

I’m instead going to go with a different option. One of the things I’ve noticed with developers is that while they may pick up on the opportunity to create a data abstraction – or not, given the title of this series – they rarely pick up on the opportunity to create an algorithmic abstraction.

So, I’m going to go in that direction.

From an abstract perspective, this code does the following:

• 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

Our current implementation implements this through procedural statements, but we could consider doing this differently. Consider the following code:

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

This is an abstraction of the pattern that we were using; it’s general-purpose code that takes a source, processor, and sink, and knows how to wire them up to pass data between them. The version of Process() that I wrote handles this specific case, but you can obviously write versions that have more than one processor.

Doing that with our code results in the following:

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

I like the first and third arguments, as they are just method names. I would like the second line also to be a method name.

We can recast this by adding a Filter() method to EmployeeFilter, which then yields:

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

That makes me fairly happy; Pipeline.Process() is tested, and the code to use pipeline is almost declarative.

Take a few minutes and look through the current implementation, and write down what you like and what you don’t like.

Looking at the code, it is improved, but there are still a number of things that I don’t really like:

• 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.

Those will come up next time.

 

In the last post, I talked about some of the analysis used in lean. Now, let’s see how we can apply the same principles to the software world.

I will start by trying to create a value-stream map. But where to start? There are three levels at which we can do the mapping:

1. The entire organization, starting at the idea stage and going through the stage where we ship bits to customers.
2. The overall implementation of a feature/MBI/whatever by the development team.
3. The inner developer loop for a single story/feature/whatever by a single developer or pair.

I decided to start at #3, spent a lot of time writing, and then realized that the inner loop discussion depends a lot on the other ones, so I put that aside, and decided to start at the outside, with level #1. First issue solved…

I also need to decide what process I am going to map. I’ve worked in a number of different teams with very different processes, many of which I don’t remember perfectly. So, instead of choosing a real team, I’ve decided to invent a composite team that is a bit more on the old-school side of things. Perhaps “old-school interested in building a new school” is a decent description.

The actual process of creating one of the is done through an interactive discussion with a number of people and is very customized to the way a particular organization works. Some of what I create here will not apply to you, so I’ve come up with two ways that might make it more useful for you.

First, I encourage you to draw your own diagram as you read through what I am saying – or perhaps as you read through it a second time. And second, I’ve added a few Question sections where you might want to think a bit about what your answer is before you read mine.

That said, let’s dive in. Our organization builds software, and follows the following steps:

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

Let’s put those into our first diagram:

Here’s a little bit more information about the steps:

Enter idea

Somebody comes up with an idea, and we need to track it, so it gets entered into a system. This takes about 15 minutes per item.

Traige by priority

Once we have the ideas, we need to order them by priority so we know what the most important ones are. This is typically done by a small group of people who meet for a couple of hours every month.

Feature estimation

To be able to determine which features fit into the release, we need to know how big they are. We will ask the engineering team to do the estimation for us. Since we ask them to give us good estimates, it takes anywhere from 1 to 8 hours of work for them to do each estimate.

Release planning

Before we can start working on a new release, we need to figure out what features we are going to put into it. This is a very important function as it determines what we will invest in for the next release and there are lots of competing priorities, so we set up a release planning group and they spend a lot of time figuring out a draft plan, getting it reviewed, modifying it, rolling it out, etc. It takes about 100 hours for the group to do this, spread over a period of a couple of months.

Create

The engineering team takes the list of features, works through them, and creates the new version of the software. We are using a 2 year ship cycle, so this phase is that two years minus the overhead of planning and shipping, or about 20 months.

Ship

There are things that we need to finish at the end of the cycle; we do acceptance testing and fix the leftover important bugs and go through all the mechanics of putting together the full release. This takes about 2 months.

Let’s add those times to the chart:

Something you may have noticed is that there are different kinds of times here; “Enter idea” and “Feature estimation” are per-item times, “Triage by priority” is a periodic short meeting, and the rest are just time boxes.

What’s next? Well, those steps just don’t feed directly from one to the next; we have inventory. And wow, we have a lot of it; so much that we give each set of inventory a name so we can tell them apart. We’ll add the inventory with names and counts to the diagram:

These are numbers that I’ve seen for a small product group; one with 20 or so developers on it. They are conservative; I’ve seen examples where the number of prioritized features is well about 500.

There may be a way to express the timeboxed sections properly using the value-stream mapping nomenclature, but I don’t know of one, so I’ll stick to this approach.

Since we will be trying to determine how long it takes a item to move through our system, periodic processing poses a bit of a challenge; one item might get lucky and arrive on the day of the meeting, and another might get unlucky and arrive right after a meeting and have to wait for the next one. We’ll therefore choose an average that is half of the cycle time.

Here’s a little table of our average end-to-end time:

Step	Inventory count	Elapsed time
Enter Idea	0	15 minutes
Triage by Priority	200	2 weeks
Feature Estimation	200	5 weeks
Release Planning	100	2 months
Create	50	20 months
Ship	50	2 months
Total		26 months

A little over two years for an idea to come in and make it through our system, which is about what we would expect for a 2 year cycle.

Where’s the bottleneck?

Question: In our current system, where do you think the bottleneck lies?

Answer: Our analysis is a little complicated by the fact that many of our operations aren’t full time, so let’s focus on the Create and Ship steps. Let’s express them as rates:

Create: 2.5 features/month

Ship: 25 features/month

So, creation is our bottleneck, which means I’ve spent a lot of words coming to the result everybody was expecting. Let’s continue building the rest of our model to see if we get some more insight into what is going on.

Because of the long lead time – 26 months – this approach cannot be very responsive to changes in business climate or to needs that are discovered as part of the development process. There are different ways to address this; for this part of the discussion, we are going to choose one traditional way of getting there.

We’ve added a change request process, whereby we identify a small number of very-high-priority features and feed them into the middle of the create process, thereby getting better responsiveness. This feels like an agile and responsive thing to do, but in reality, it really doesn’t help us a ton, for a couple of reasons:

• 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.

Or, to put it in ToC terms, adding additional destabilizing work to the bottleneck does not yield great results, for obvious reasons.

Pretty much everybody hates change requests, but they are viewed as a necessary evil under this model.

Quality

Our group creates a lot of bugs. Let’s add them into our design:

This world of bugs is a complex one. There are weekly triages, bug bars, ship rooms, customer support interactions, service-level agreements, quick fix engineering, hotfixes, service packs, etc. Lots of process.

I spent a while diagramming the process with all of the detail, and quickly realized that if I wanted to include it in this diagram, I had to simplify it considerably, so this is a bare-bones representation. I am not including verification time for bugs, and I’m also not including loop-back for retriaging existing bugs or bugs that were marked as fixed but aren’t, nor am I including regressions. All of which increase the time significantly.

<aside>

If you have a lot of bugs, I highly suggest trying to diagram what happens with bugs including the customer and all the teams that deal with bugs and all the steps you go through. I think it will be enlightening on how much those bugs are costing you.

</aside>

So, anyway, we have two sources of bugs: bugs from the customer that come in from the left and bugs out of the current create step that come in from above. We triage them, and then they are fixed by our development team – the same group that is trying to do the create work. Then we either ship them as fixes to our previous product, or we feed the fixes back into the product under development. Or both.

The time the development team spends fixing bugs is time that they aren’t spending on product work, or – in ToC terms – we have added another stream of work to our bottleneck. How does it affect the total time for an item to go through the system?

Well, if our team fixes 1000 bugs during this long cycle – and that’s not an outlandish number – and they each take 4 hours to fix – also not an outlandish number, then we are spending 500 days of time on bug-fixing. A 20 month project for 10 developers is on the order of 4000 days, so it looks like we are spending around 12% of our time on bugs. Which really doesn’t seem that bad, so we can probably ignore it.

However, if we have a lot of bugs external to the create step, there are also a lot of bugs internal to that step, which are costly. I’ll defer that to the next post.

That completes the creation of our initial diagram. It’s a simplification of the real world, but should server our purposes.

Talkin’ ’bout Optimization

Now that we have identified the constraint, the first thing that we need to do is to exploit it. Which is one of the topics for the next post.

The problem that we have at this level is that we have a significant and very obvious problem – the problem of bugs – but we *have not* attempted to deal with it in the proper way, from within the development team.

Question for the class: Viewed through the lens of ToC, how would you describe the bug triage and fix system? What step does it belong to?

Answer: The whole bug flow is about both subordinating and elevating the constraint.

We have de-optized the process for our non-developers by introducing a new stream of work, which is handled by bug-fix orchestration technicians. They are in charge of looking at bugs, deciding which ones are important enough to fix, deciding when we should fix them, deciding who should fix them, tracking our progress on fixing them, predicting when all of them will be fixed. etc. etc. etc.

And, as this process is obviously costly, we introduce bug-fix orchestration *optimization* technicians, who are responsible for making the process faster. And bug-fix customer advocates, who liaison with the orchestration technicians to make sure that the customer impact of bugs is properly considered. And, in the ship step, we have bug fix coordination meetings, where we get together to make sure we are fixing the right bugs.

And much of this process is handled by our management, which means we are using our high-value assets towards this goal, instead of deploying them towards more strategic concerns.

And I’m still not done…

In my diagram, there’s the section of bugfixing where we are fixing bugs in products that have already been shipped. This takes time away from the developers working on the current product, so we can elevate our constraint by setting up a separate process to handle bugfixing for shipped products.

We create a separate group that is only responsible for bugfixes in our current shipping product, staff it up, and that makes our main process faster. We call this group “sustained engineering”. They are responsible for creating and managing service pack releases to fix the worst of our bugs.

Question: How effective are these approaches at reducing the effect of our constraint?

Answer: Not very. In fact, it can be argued that they make the problem worse.

To draw a manufacturing analogy…

We are building cars, and we have been noticing that when they come out of the paint shop, the doors consistently have some glops of paint on them. We also know that our customers are complaining about the door paint flaking off.

The paint shop is busy creating new doors and also has a lot of doors that need to be redone piled up outside, so we set up a group that looks at all the doors that failed our QC check, takes the ones that we think aren’t bad enough to really bother the customer, and puts them back on the assembly line. Only the very bad doors go back to the paint shop. We look through this whole pile every few months, or at least most of it.

We have customers complaining about gloppy and flaking paint. We tell them that we don’t normally fix those sorts of issues, but if they buy the special extended warranty package, they can get special access to our door fix technicians. We have a separate group who is in charge of doing that – once again, so we don’t slow down the paint shop.

 Question: Taking the ToC perspective, what are we missing?

Answer: We missed exploiting the constraint.

It is pretty clear that we have a massive rework problem; we are spending vast amounts of effort fixing issues that are come out of the create process, and it’s so bad that we are ignoring a number of issues that our customers care about. But instead of exploiting the constraint, we are trying to mitigate it from the outside.

I will expand on the bug discussion in a future post, but I think there is an interesting parallel in the manufacturing world.

In the 1950s and 1960s, the US auto industry was quite successful at building products that didn’t have great quality. Then, a bunch of imported cars showed up – cars built by companies that had adopted a very different view on quality, a view that, ironically, came mostly from a number of people from the US.

You probably know how that turned out; the US auto industry learned a painful lesson about trying to compete with companies that are using a better system.

The parallel to software is pretty obvious.

Enough about bugs for now, time to switch topics…

Risk

Risk is something that those of us in development teams really don’t think too much about; that is mostly a topic for the business people. But I’d like to spend a little time on it here.

The two big risks that we need to worry about are:

• 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.

We have a problem with our 2-year cycle; that is quite a long period of time and it is well known that software estimates aren’t very good, so we run a pretty good chance of either being late or having to cut features. Or both. And we have the list of bugs we need to fix, and regressions to fix, and we know that the time to fix bugs isn’t terribly predictable.

There is also a large market risk; we know from our diagram that it takes 26 months for most features to make it through the pipeline, and though we can use change requests to be faster, it will generally only cut our cycle in half, and that’s still fairly long.

This long cycle also requires a fairly large investment of money; a development team of 10 can easily cost $200,000 per month, so over the two years, we’ve put 24 * $200,000, or $4.8 million into the release.

In the theory of constraints, the bottleneck in the system isn’t always something physical; sometimes it’s a policy or choice made by the business. The 24-month cycle is an arbitrary one, and it is something that we could change.

If we pick a shorter approach, we can cut our schedule risk, cut our time to market, and cut our investment. Which seems like a total win to do.

How can we make this change the easiest and safest way? Well, we know how to plan in 2 months, and we know how to ship in 2 months. How about if we shrink the “create” part of the process down to two months, which will give us a 6-month overall cycle (2 months planning, 2 months create, 2 months shipping). Once we do a couple of those and get used to them, we will naturally find ways to cut down the planning part, and we can probably make the shipping part a bit faster as well.

This is a great idea, but it turns out that pretty much nobody does this, and here’s the reason.

In our old system, we were devoting 20 months out of 24 to creating features (well, and bug fixes, but I’m ignoring that for now). If we go to a 6 month cycle, we only get two months worth of “create” in each cycle, for a total of 8 months every two years.

8 months of feature creation is much less than 20, so there is no chance we are going to adopt that approach, despite the fact that it gives us far faster time to market, much less risk, and is something that our team can demonstrably do successfully.

What teams generally do is go for a linear reduction; our 6 month cycle becomes 2 weeks planning, 20 weeks coding, and 2 weeks shipping. Planning works fine – it’s easier, in fact – and development starts the way it always done, and then – as the ship portion approaches – the team realizes the following:

• 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.

Their options are:

1. Delay shipping.
2. Cut corners

The team generally picks a combination of these. And then, they turn right around and start the next cycle.

There’s another bit of organizational psychology going on here. In a lot of US companies, being bold, aggressive, and making big bets is encouraged and rewarded. That is why we see re-org after re-org; we are doing something significant.

This makes little sense to me. From the agile and lean perspective, we want continuous improvement, because that is the path where the risk is minimized. There is very likely a path that takes the team from a 2-year cycle to a great 6-month, 3-month, or even 1-month cycle, but it’s very clear that you cannot get there just by decree.

That’s all for this post. In the next one, we are going to crack open the create process and see what’s inside.

Part 3 – The Development Team (1)

I’ve been spending some time over the past few months exploring the lean side of the house and looking for things I can adapt into the agile side of the house. The most interesting thing I found was the theory of constraints.

After spending some time writing this, I realized that I need to split this into two separate posts; one where I talk about the theory of constraints in general, and the second where I talk about how I think it applies to software.

For this section, I’m going to talk about manufacturing, partly because that’s where the theory was originally applied, and partly because it’s more approachable. Trust me that what I write here will apply to software development.

The classic work on this is “The Goal” by Goldratt, which I highly recommend.

Let’s make it better

From: Plant Manager
To: Component assembly;welding;painting;packaging
Subject: Improvement

It’s time to kick off our 2016 improvement process; I would like each of you to get together with your teams and figure out what your improvement targets are going to be for next year. 

Signed,

Plant Manager

If a business does improvement – and many do not – this is a pretty typical approach. And, if the word, “poorly” pops into your head, you have already figured out how well this sort of approach works. If you think we are better in the software business, you are mistaken. In general, we are quite a bit worse.

Why do these programs fail? It’s very simple…

To make a process faster, you must first determine why it is slow.

I’m hoping somebody is saying, “I know why my process is slow”. And you may be right, but I am also very convinced that you are also quite wrong. And that gives me a chance to introduce my first thought:

Thought 1: To improve a system, you must first understand the whole system. 

If you do performance optimization of programs, you may know the first law of optimization, which is, “The part of the program that is making things slow is never the part that you think it is”. If you go around optimizing the parts of the program that you think are slow, it doesn’t really get much faster.

Hmm. Isn’t that exactly what is happening with groups trying to get faster?

To make a program run faster, you use a tool to analyze the behavior of the whole system – a profiler. And, to make a manufacturing plant run faster,you need a similar tool.

That is what the theory of constraints can give us – a way to look at the whole system.

Our goal

When doing optimization, we need some sort of metric, or goal. In performance optimization, it’s execution time.

What should our goal be? I’ve already asserted that “improve the output of each section” is an ineffective way to look at things, and I’ve said that I want to look at the overall system, so how about “improve the output of the whole plant” as a goal?

That seems good. We know how to measure it (units shipped per month), and everybody can work together to make it better. And I agree that shipping more units per month would be a good thing, but as a goal, it is an utter failure, because of a couple of simple problems.

It gives us absolutely no guidance on how to actually improve the current state; it does not tell us why the current system is slow.

It’s also a bad goal for another reason; if we all pull together, we can ship a lot of really crappy product in a short period of time.

So, we need a better measure, and luckily, there is a very good one; we track the elapsed time it takes us to produce a product, from order to ship. Let’s walk through how we are going to track it.

Define our process

Here is the process for our plant:

We are trying to figure out how long it takes from when we start manufacturing to when it goes out the door, so, we go out and do some measurements of how long each process takes, and add them to our diagram.

And now we know that it takes 60+45+60+30 = 195 minutes to make one item, and we can go off and start optimizing. It probably makes sense to start with component assembly and painting, since they take the longest.

Wrong.

In this scenario, the current end-to-end time for a specific item is on the order of 10 days.

Wait, what? How can it be 12 days if the process takes only a little over 3 hours to complete?

I’d like you to ruminate on the situation. There is something missing in the diagram that I drew, and it’s something that I could easily have measured when I went out and measured the time each individual step. What is missing?

(spoiler space)

 

 

 

 

 

 

 

 

 

 

 

An improved picture

What I was missing was the concept of inventory. Whenever there is a handoff between two steps, there may be an accumulation of inventory. That is where the extra time is; we have 100 items waiting to be welded, so each item will have to wait for the 100 items in front of it to be processed first. That will take 4500 minutes, or about 75 hours. There are two items waiting for painting, so that is 1 hour of time there, and the 5 items waiting to be packed add 2.5 hours of time.

So the total time is 75 + 1 + 2.5 = 78.5 hours of lag + 3 hours of processing = 81.5 hours, or a little over 10 days.

<aside>

Why is inventory bad?

Inventory is bad, let me count the ways:

1. It ties up a lot of capital; we have invested money in the raw materials and the labor to create the intermediate state. If the inventory was lower, we could deploy that capital elsewhere or we could improve our return on investment.
2. We don’t know how good it is. If our component assembly starts producing poor-quality items, it will be nearly 10 days until we find out, and we will have to throw away/rework a lot of expensive items.
3. The items in inventory represent things that we think we need, but our plans may change during that time, leaving us with intermediate items that are of little use.
4. We have to pay to store it, track it, move it around, etc.

</aside>

The diagram I’ve created is a very simple version of a value stream map. And the times I used are pretty conservative; it isn’t uncommon for the end-to-end time of a single item to be measured in months.

We’re going to put aside the amount of inventory for a moment, and focus on the steps.

Now, can we make it better?

Do we know where we are slow?

The answer is “yes”. Looking at the diagram, we can tell that we have an issue with the welding process. You can do it mathematically; component assembly is capable of creating 3 items per hour, but welding can only do 1.3 items per house. Or, you can just look for the places where inventory piles up in the factory.

Saying welding is slow is really a misstatement; there may be absolutely nothing wrong with the welding process; what we have discovered is that the welding process is a bottleneck in our system. Because it is the slowest step, it is constraining the output of the system to be – at best – one item every 45 minutes.

That concept is why the theory of constraints is named what it is – we have a constraint, and it controls the output of the whole system.

Let’s now cast our minds back to my earlier assertion that “everybody get better” programs don’t work, and see whether this diagram can shed any light on the situation. What happens to the system if we improve the speed of component assembly, painting, or packaging?

That’s right, pretty much nothing. Which leads to:

Thought #2: If you aren’t addressing the bottleneck, you won’t improve the overall system performance

Given that few groups know what their bottlenecks are, it’s not surprising that their attempts at optimization don’t improve their system performance.

Now that we have a target, it’s time to talk about ways to address it. There are a few options:

1. Increase capacity (buy more equipment). This is the go-to option in most cases, because most groups don’t know how to optimize. It’s also the priciest.
2. Optimize the bottleneck. Look at the process of the bottleneck in detail, and see if there is a way to optimize it. This might mean creating a separate value stream map for the bottleneck. Can we get better utilization out of the machine?
3. Subordinate the other parts of system to the constraint.

Let’s talk about the third one, because it’s the least obvious and therefore most interesting one. Looking at the picture again, we have a lot of excess capacity in component assembly. We do a little investigation, and discover that about 20 minutes of the welding work isn’t welding work, it is “getting ready to weld” work. Let’s modify our process again.

We pulled 20 minutes of work out of welding and added it to assembly. What is our total end-to-end time?

Well, our welding queue is now 100*25 = 2500 minutes, or 42 hours, but the painting queue has gotten bigger, and now has 20 hours. Total is 42 + 20 + 2.5 = 64.5 hours, a significant improvement. Note that the queue sizes I chose were for purpose of illustration.

And how did we do that? We did it by moving work to the component assembly step; it is now a full 20 minutes slower than it was before, but it being slower had no impact on the overall system, because it still produces items faster than the welding step can consume them.

Thought 3: Sometimes the best way to improve a bottleneck is to de-optimize the steps around it. 

This is another “wait, what?” moment; one step got slower, and the overall system got faster. It’s a little easier to see if I add in a little table:

Step	items/hour before	items/hour after
Component assembly	3	2.5
Welding	1.3	2.4
Painting	2	2
Packaging	2	2

It’s pretty obvious why we had a big queue in front of welding in the first case. It’s still slower than component assembly, but it is now faster than painting, so we are seeing a queue show up there. And we pushed our overall system performance up to 2.4 items per hour.

Which leads to another question. Is welding still the bottleneck?

The answer is obviously “no”; painting and packaging are the bottlenecks now. So, that is where we would work next.

Rework

I made a simplifying assumption for the earlier diagrams; I assumed that all of our processes were perfect. But, in reality, they aren’t, so it would be good to add that to our diagram.

This is one way of expressing rework. It says that 10% of the time, we spend an extra 30 minutes on welding to fix issues from the previous step, so that bumps the average welding time up to 28 minutes. There can also be inventory before or after the rework part; as you might expect, this can bump up the end-to-end time significantly.

What to do with inventory

After improving our throughput, we still had one big queue and we were starting to accumulate another. In a perfect world – where every step was matched in capacity – our queues would be a fixed size, but that never really happens; there is always a bottleneck one place or another. And, as we notice, our end-to-end time is heavily dominated by the time due to the queues.

If we let the system parts free-run, that will lead to the accumulation of a ton of inventory over time, which is very bad. The most common approach is to switch from a push system – where the previous step just runs flat-out – to a pull system – where the previous step runs when the next step needs more items. This is commonly known as a “just in time” approach, and in it’s simplest incarnation, you set an inventory trigger point (say, of a few hours inventory) that lets the previous step know it should start running again.

If we take our queues down to half a day, we would only end up with 1.5 days worth of inventory in the system, taking the overall end-to-end time down to around 2 days. That cuts our inventory cost down to about 20% of what we had before, and our agility is up a similar amount.

But… this change isn’t free. Since we are carrying less inventory, we need to watch the system a little more closely to make sure that we don’t run out.

Because our component assembly step has excess capacity, we will need to run it slightly slower so inventory doesn’t build up. This approach flies in the face of traditional management philosophy; we are specifically telling a group to either slow down or do something else during the down time.

Setup time

It’s pretty common for a single machine/team to do multiple things. In that case, there is a setup time overhead to switching from doing one thing to another. Implementing pull systems will tend to drive the batch size down, and it’s important to remember that there is overhead to consider.

Summary

After writing this, I went out and looked at the defined steps of the theory of constraints again, to see how I did. Here are the steps:

1. IDENTIFY the system’s constraint.Okay, I covered that.
2. EXPLOIT the constraint”Exploit” here means “do whatever you can to optimize within the constraint”. I forgot this one initially but went back and added it.
3. SUBORDINATE everything else to the constraintThis is the part about being willing to de-optimize another area to speed up the constrained part.
4. ELEVATE the constraint.If it’s still a constraint, this is when you throw money at the situation; buy new machines, that sort of thing.
5. PREVENT INERTIA from becoming the constraint.If you addressed one constraint, your new constraint moved somewhere else. You’ll need to think of that next.

 

That’s all for the introduction to the theory of constraints. In my next post, I’ll take the techniques that I talked about and apply them to the software world.

Part 2 – The Development Cycle