Kanban systems can be used when the work generally follows some kind of common workflow.  For many software development projects, this is an easy constraint.  Most software involves some kind of problem definition, some kind of design activity, and some kind of verification.  Most software development also involves the use of a limited set of technology applied to a limited set of systems within a limited problem domain, so that most of the work done falls into a few general types.

Most is not all, however, and sometimes work will appear that doesn’t fit neatly into any well-defined type.  Or maybe it does have a type that we haven’t defined yet.  If the work is non-value-added, we can chalk it up to overhead.  If the work is value-added, we will want to track it like all of the other value-added work.  A useful trick for managing that kind of work is to reserve a swimlane for ad-hoc workflow:

If you find yourself using the ad-hoc swimlane frequently, you might want to map the value stream of some of the work items to see if you can discover some latent or emerging workflow.

The simplest kind of kanban system is the CONWIP system, for CONstant Work In Process.  The simplest kind of CONWIP system is no more than our fundamental kanban element:

The simplest CONWIP cardwall is a classic Agile cardwall with a limit on work-in-process:

An equally intuitive interpretation of CONWIP defines capacity simply as the number of people available to work, so that each person is the kanban:

CONWIP is a rule about work items, not a rule about workflow.  We are free to define any workflow we like as long as we observe the global limit.  This can be a helpful approach when we want to observe the flow of work, but expect a lot cycling between states, perhaps in an exploratory design mode:

The earliest Scrumban design was just such a CONWIP workflow, which we can represent directly in a simple cardwall.  Here we have no limits on any specific column, but the total number of work items is limited by the yellow kanban cards, which are returned to the “Free” box when the task they contained is complete:

Breaking out a workflow implies a sense of sequence.  If there is no real sequence, we could define a global “busy” state and reduce specific activities to a checklist.  Exposing the workflow to visual control might suggest a need to level resource utilization or that we value reducing backflows as an improvement goal.  Pooling the WIP limit suggests that the visual control is enough feedback to help the team achieve leveling.  My personal preference is to work in this way where possible.

If a person can be the kanban, then how about a pair? If we combine pairing and kanban with a workflow or checklist, then we get something like Arlo Belshee’s Naked Planning:

I think the most interesting kind of CONWIP system is the Bucket Brigade. Everybody is allowed one card at a time, which can either be moving upstream or downstream. People can either pair (at the cost of some queueing) or work solo as much as they prefer. Constant WIP, zero queueing, full utilization, soft specialization, balanced workflow…what’s not to love?

The model can be edited and simulated in PIPE2.

If we can assign tasks and limit work by pairs, then why not do this with whole teams? That is essentially the approach of Microsoft’s Feature Crews with their quality gates.

Some value streams are too long or too complex to effectively manage with a single WIP limit.  Many development teams exist within some larger organization that requires coordination between teams and competition for resources.

Where do features come from?  Smaller shops may interact directly with the customer, but larger shops may have a more complex process for dealing with a large number of customers or with highly sensitive customers. If we peer inside the black box of the Product Owner, we might discover a whole team of people working to understand customers and define requirements: business analysts, product managers, usability researchers, product designers. Work-in-process on the business analysis side can just as easily go off the rails as development work.  Have you ever seen an epic requirements specification or a bottomless product backlog and wondered where it came from?  Your product owner might represent another group of people who feel pressure to produce and look busy.  Value stream thinking encourages us to take an interest in what those people are up to and why.

Where do features go after we’ve built them? A large enterprise may have complex deployment requirements that involve integrating code into a manufacturing process or provisioning a datacenter. This work probably involves a different team than the development team, but they are still part of the value stream and their throughput affects everybody.  Operations teams often have to deal with long lead times and different natural batch sizes than the development teams that feed them.  Each group can benefit from understanding the status and availability of the other.

A small team may be able to self-regulate with visual control, but a long value stream may need more explicit control. Kanban gives us an easy solution by chaining together pooled segments. Within each segment, we manage by visual control, but between segments, we manage by kanban:

CONWIP systems allow us to regulate team workflow with a lighter touch.  Pooling kanban across closely related functions and zooming out by one level of scale makes it easier to think about using kanban to manage large systems.

Any kanban-controlled workflow system can be described by combinations and variations¹ of a basic pattern:

Sometimes we can simplify the diagram by replacing the kanban backflow with a simple capacity parameter², but often it is better to show the flow of kanban explicitly.  Many of the software development kanban systems we’ve seen are simple workflow systems composed by chaining this basic element:

I would like to think that any professional software engineer would be able to think up more interesting workflows than just a linear cascade.  Then again, I would also like to think that any professional software engineer would understand the value of keeping things simple.  I have personally come to prefer a more symmetrical call-stack style of flow for software development, because I believe that any person who requests custom work should also be responsible for approving the completion of that work. Consumers pull value from producers, not the other way around:

Petri nets are ideal for describing workflow systems because they are a) concurrent; b) formal, simulable, and sometimes even verifiable; and c) relatively easy to read by humans.  Any Petri net that can be drawn without crossing edges can easily be made into a “card wall” for visual control³:

Sometimes a different workflow is needed, depending on the kind of thing being made:

Some tasks can be done in parallel by specialized resources:

When we split tokens, we may need to keep track of their common ancestor so that we can merge them again.  Colored Petri nets let us associate composite work items across branches:

Sometimes a large work item can be decomposed into smaller work items of a similar type.  We might think of a branching workflow to model this, but that is hard to do if we don’t know how many component work items will be created.  Petri nets allow us to take another approach by generating new tokens in-place and then executing them concurrently on the same workflow branch:

When all of the unit work items are complete, they are integrated into their parent work item:

While that might look a little complicated, in practice it’s as simple as the “2-tier” style (or n-tier) cardwall that is often used for project management:

A state transition is a black box that may have some internal process.  We might expose that process with a hierarchical model.  Alternately, we might want to collapse extraneous diagram detail into a single supertransition.  Hierarchy is a simple syntax extension to any workflow model.

Feedback should be considered implicit to any creative process, but it can complicate these models without much benefit to understanding⁴.  In practice, kanban systems regulate feedback very well, because the limits serve as a ratchet function that gracefully responds to feedback and damps oscillation.  A process operating at capacity will not accept new work, and a process operating over capacity will also not accept new work.  Again, it’s awkward to model “over capacity” so we have to be mindful to treat our models for what they are: models.

Once we understand some of these basic design elements, we can use them to describe or design a wide variety of product development processes.  A computer scientist armed with Petri nets and a bit of knowledge about queueing, networks, and processor scheduling has some wicked tools at his disposal for Value Stream Analysis.

A staple of software engineering research is the effectiveness of design reviews and code inspections for discovering defects.  Methodologists love inspections, but they seem to be difficult to sustain in practice. I’ve seen a few typical reasons for this:

1. Inspection is a specific skill that requires training and discipline.  Naive, unstructured “code review” is worse than useless and eventually self-destructs.
2. Inspection is quick to be dropped under acute schedule pressure, and slow to restart as a habit once it has been broken.
3. Inspection works well for frequent small batches and badly for infrequent large batches.

Reason 1 is a matter of skill, and can be solved with education.  Reasons 2 and 3 are process issues.

The “inspection gap” illustrates a curious aspect of human nature.  There are certain behaviors that a group of people will agree should be practiced by its members.  Individual members of the group, when asked, will say that they believe that members of the group should practice the behavior.  But then, in practice, those same individuals do not practice that behavior or practice it inconsistently.  If you point this out to them, they may agree that they should do it, or even apologize for not doing it, and then continue to not do it anyway.

In my mind, this is a good part of what Lean thinking has to offer.  Lean methods like Visual Control recognize this aspect of human nature and provide people with enough structure and context to act in a way that is consistent with their own beliefs.  If people using a Lean process agree that code inspections are a good idea, then it will not be hard to get them to agree to incorporate inspections into the process in a way that is hard to neglect.  Lean strives to make it easier to do the right thing than do the wrong thing.  Lean helps people align their actions with their values.

Checklists

One practice that works well in most workflow systems is the simple checklist.  Human attention is a delicate thing.  People get distracted, make mistakes, and overlook things even when they know better. A checklist is a simple device to keep your intentions aligned with your actions. Doctors who use checklists deliver dramatically improved patient outcomes.  Would you get on an airliner with a pilot who didn’t use a pre-flight checklist?  Would you get on an airliner controlled by software that was written without using checklists?

Checklists and kanban are highly complementary because you can attach a checklist directly to a kanban ticket and make the checklist part of the completion transaction.  Checklists improve confidence and trust, and expose tacit knowledge.  Checklists relieve anxiety and reduce fear.  Can you think of any part of your development process where you’d sleep better at night knowing that all of the important questions were answered correctly by somebody you trust?

Workflow

Checklists work well for individual activities that do not require specific sequencing, but they don’t work as well for activities that require collaboration from people who have competing commitments.  We can raise the stakes for everybody if we elevate our checklist item to the workflow and subject it to the pull discipline.  That makes your problem everybody’s problem and gives your peers sufficient incentive to collaborate.

Inspections are a typical example at the scale of a single developer, but there are other practices and scales that we might consider.  Failure Mode and Effects Analysis (FMEA) is another highly effective technique that many people agree with in principle but find difficult to implement in practice.  FMEA is a systemic method and often targets components or subsystems that are much larger than “user story” scope .  Security lifecycle and regulatory compliance activities may also fall into this category.  An advantage of using composite workflow is that you can schedule activities that apply to different scales of work.

Process retrospectives can also be attached to workflow in this way.  Compared to a more open-ended periodic retrospective, a workflow-bound retrospective asks a more specific question:  How could we have created this work product more effectively? Such a workflow-based retrospective directly implements Deming’s Plan-Do-Study-Act cycle.

Are there any practices you would like to see your team use consistently, but have trouble fitting in to your schedule?

This confusion stems from a misunderstanding of the relationship between basic principles of production at Toyota and kanban as a technique to help implement those principles.

– Shigeo Shingo, A Study of the Toyota Production System ]]> http://leansoftwareengineering.com/2009/05/27/shingo-on-cargo-cult-kanban/feed/ 1 http://leansoftwareengineering.com/2009/05/13/more-conference-coverage/ http://leansoftwareengineering.com/2009/05/13/more-conference-coverage/#comments Wed, 13 May 2009 21:59:20 +0000 Corey Ladas http://leansoftwareengineering.com/?p=1355

I should be more enthusiastic, but I bit down on a stone in my food and broke a tooth yesterday.

Notice that “match the number of people currently available” is not one of the ways.  You’re trying to discover how many people you need not how many you have.

Feature Crews

The namesake attribute of a Feature Crew is the cross-functional workcell. A crew should contain most of the capability that it needs to fully specify, design, verify, and integrate a complete product feature. Typically that means a Program Manager (in the Microsoft sense of the title), a handful of developers, and a couple of testers. Depending on the type of feature, there may also be a product designer or other specialist, but we’ll see how such a resource might be shared across workcells.

Branch-by-Feature

One of the primary problems that Feature Crews address is the difficulty of maintaining the integrity of very large code bases under development (imagine 1000 developers coding against a 10,000,000 line system). FC poses the problem as the tension between a) keeping the main branch as current as possible, and b) keeping the main branch as robust as possible. The FC solution is to make features an atomic transaction. A feature is either 0% complete or 100% complete, and a feature is not 100% complete until it can be demonstrated that it satisfies the same quality criteria as the rest of the main branch.

Features-in-process are not allowed on the main branch. The FC alternative is branch-by-feature. A crew takes a branch when it takes possession of the feature kanban. The crew is responsible for forward-integrating any changes that are checked into main while their feature is in process. That is, if another crew integrates and breaks your feature-in-process, it’s your responsibility, not theirs. When your feature is finally complete AND you have integrated with all changes on main AND you pass all of the quality gates, THEN you can reverse integrate your feature into the main branch, and everybody else will have to forward integrate your changes.

From a Lean perspective, even without the scale issue, there is an argument to be made for atomic MMFs and branch-by-feature. MMFs are business-valued. User stories are merely user-valued. Customers demand features. Users merely request them. Allowing features-in-process on the main branch exposes the product to inventory and market risk. A Lean deployment model should be transactional at the granularity of business value, not user value. The purpose of the MMF practice is to bring those values as close together as possible.

Quality Gates

While things like unit tests and customer acceptance tests are necessary to meet the quality criteria of a multi-million-line codebase, they are certainly not sufficient. A Minimum Marketable Feature, taken as a whole, has properties that are more than the sum of its component tasks. There are certain kinds of design and quality control activities that have a larger natural granularity than user stories.

Quality Gates ensure that all of the systemic work gets done that does not fit naturally into the functional development process. A good deal of security and reliability engineering has nothing to do with the intended functionality of a feature, and everything to do with how components of a system will behave in the presence of other components of that system under a wide range of operating conditions. Quality Gates also facilitate the sharing of specialized resources, like a security engineer or a system architect, that are impractical to include on every crew.

Feature Crew is only a process framework

The Feature Crew model treats both features and workcells essentially as black boxes. Like Scrum, the Feature Crew method is not prescriptive about workflow. What happens inside the crew is somewhat up to the crew.  One could imagine that a crew implements a mini-phase/gate (and some do, though we know that isn’t wise). A crew could choose to implement an off-the-shelf process like Cleanroom or XP for their internal workflow, and many do. Some crews will use an ad-hoc local body of practice, or git-r-done cowboy craft. The quality gates set the bar, how you meet the bar is your concern.

Since Feature Crews address a different scale than Scrum, we can even combine Feature Crews with Scrum. If a crew takes on a 6-week feature, that crew could then overlay a 1-week timebox within those 6 weeks and decompose the feature into Scrum-like work items and goals, which are then implemented according to local custom. Again, this is not uncommon practice.

Kanban hierarchy and matrix workcells

Since we already understand Feature Crews as a type of kanban system, and we see how can we can overlay Scrum as a secondary planning process, then it follows that we can use something like Scrumban as the internal scheduling process. We don’t have to do this, but doing so buys us some symmetry between the layers of planning hierarchy, and allows us to share metrics, tools, and terminology between workers and management. The cumulative flow report for your team is of a similar kind to the cumulative flow report for the project as a whole.

Feature Crews are an effective method of capacity calculation, but they are also a blunt instrument for that purpose. Making a team self-contained can result in boom-and-bust duty cycles for individual team members. The beginning of a feature branch may be heavy on analysis and high-level design, and the end of a feature branch may be heavy on system testing and bug fixing. A UI designer might find herself very busy for the first couple of weeks and mostly idle for the last couple of weeks. Such a team may be able to learn a certain amount of task leveling, but only so much before running into other problems.

The complementary dysfunction of individual bursting is poor total resource utilization. If every crew employs their own UI designer at 50% capacity, then you’ve hired twice as many designers as you really need to get the work done. As much as we love flow, that is a high price to pay for it. Introducing a second level of kanban granularity gives us access to a finer set of controls that we can use to dial in a better balance between availability and utilization.

Nested Scrumban is enough to give us a more consistent process between the whole and its parts, but it doesn’t help us directly with our utilization problem. For that, we will appeal to a little bit of queueing theory. For some development functions we may be able to share resources between workcells with the same low delay as more dedicated resources, but at much higher utilization. Software development costs are overwhelmingly dominated by labor costs, so paying at least some attention to labor utilization is worth our consideration. Lean cares a lot about labor utilization, it just cares about it in the right order.

Given these tools, we can design a hybrid feature crew / matrix organization. Some resources can be feature-aligned and dedicated to their workcells. Other resources can be function-aligned and pooled across workcells. A product group with 10 workcells probably doesn’t need 10 security engineers, 10 user researchers, 10 architects, and 10 database administrators. But there is some right ratio of each of these functions to each other, and those ratios can be determined by value stream analysis, theory of constraints, and other heuristics.