If you’ve built a kanban system, or you’ve tried to put a WIP (work in progress) limit on a phase in your workflow, you’ve probably asked or been asked this question. And very often, then answer is “I don’t know. How about we try * n*.” where n is a guess. Usually an educated guess like:

- 2 x the number of developers
- 1.5 x the number of people on the team
- Number of people involved + 1

And these are all ok places to start if you have no data, but with a little data, we can stop guessing and set our WIP limits with some empirical information and at the same time start building a system that will satisfy one of the assumptions required for us to use Little’s Law properly. There are two things that we need to have in order to use this super easy method:

- Data about average time in state for work items
- CFD (cumulative flow diagram)

Ok, I guess we don’t *need* the CFD if we have the data, but it sure is nice to visualize this information. 😉 We do need to have some data about the way that work passes through our system and we need the data that would be required to create a CFD. For the purposes of this post, lets assume that we are capture the time in state for each work item. Entire time in the system is often called lead time. Time in between any two phases in the system can be cycle time but we’re interested in cycle times for a single state at a time as our objective is to determine the WIP limits for each column in our kanban system.

Let’s use a simple approach to measuring average time in state in days. On our simple kanban system above, we have a Ready State, Development state and a Done state. Each day, we count the # of items that have cross a state boundary and put those numbers on our CFD chart. After several weeks, we have enough data to start calculating a couple new metrics from our CFD.

With even just a couple weeks of data that visualizes how work moves through our system, we can now start measuring Average Arrival Rate (AAR) and Average Departure Rate (ADR) between any two states in our system. AAR and ADR are simply represented as the slope of a line. If we calculate the rise (x-axis) over the run (y-axis) values, we get the slope.

It is the relationship between the two values that is interesting to us and will allow us to more empirically set the WIP Limit values in the system. Based on our understanding of Little’s Law, we are striving for a average rate of divergence between the two of near 0.

A negative divergence suggests the WIP limit is to low and that we are under utilized.

A positive divergence rate suggests the WIP is too high and we are overburdened.

Since ADR (the rate at which we finish work) represents our current capability, the value of ADR should be considered a great candidate for the WIP limit for this state. With the the right WIP limit in place, AAR should match ADR and we will find an average divergence rate of 0. As your team’s capability changes, our divergence will go either positive or negative and will provide an indication of when our WIP limits should change and what they should change to.

And there you have it! When the rate of divergence between AAR and ADR is near zero, we know that our WIP limit is right and that we’re satisfying one of the assumptions required to make Little’s Law work for us!

Hi Dave, I appreciated your thoughts on this.

However, while I agree that we should strive to match ADR to ACR, that alone, unfortunately, will not tell us what our WIP Limit(s) should be: http://corporatekanban.com/no-littles-law-will-not-tell-you-what-your-wip-limits-should-be/

Hi Dan! Thanks for your insightful response! I agree with you 100%

WIP limits are definitely a on-going conversation that needs to happen within a team. I hope that people use this as a technique for identifying potential WIP limits based on observed behaviour rather than guesses.

[…] https://agileramblings.com/2013/01/13/what-should-my-wip-limit-be-super-easy-method-to-find-out/ […]

[…] appear. Predicting the right limit for each section is usually some combination of calculation, educated guessing, and […]