One Change at a Time
One of the twelve Agile principles states “At regular intervals, the team reflects on how to become more effective, then tunes and adjusts its behaviour accordingly” . Many Agile teams have retrospectives every two weeks that result in concrete actions that are executed in the next time period (sprint). There are many types of tuning and adjustments that a team can do. Examples are actions that improve the flow of work, automation of tasks, team coorporation.
Is it good habit for retrospectives to focus on the same type of improvement or should the team alter the type of improvements that will be done? In this blog I will look into the effect of multiple consecutive actions that affect the flow of work.
The simulation is inspired by the GetKanban game .
Ideally one would have a set-up for an experiment in which two exactly equivalent teams are compared. One team that would perform consecutive actions to improve the flow of work. The other team would just make one adjustment to improve the flow, and focus subsequent improvements on other areas than flow. At the same time the flow is measured and verified after a certain period of time whether the first or the second team has achieved better results in terms of flow.
Such an experiment is in practise (very) difficult to perform. In this blog I will study the same experiment by making use of simulations.
For the purpose of simulation I consider a team consisting of three specialists: 1 designer, 1 developer, and 1 tester. The team uses a kanban process to achieve flow. See the picture below for the begin situation.
In the simulation, at the beginning of each working day it is determined how much work will be completed by the team. During the simulation the average cycle time is measured. The initial Work in Progress (WiP) limits are set to 3 for each column and is indicated by the red ‘3’s in the picture.
The average amount of work done by the ‘team’ and the average effort of one work item are such that on average it takes one card about 5,5 days to complete.
At the end of each work day, cards are pulled into the next columns (if allowed by the WiP limits). The policy is to always pull in as much work as allowed so the columns are maximally filled. Furthermore, the backlog is assumed to always have enough user stories ready to be pulled into the ‘design’ column. This very much resembles developing a new product when the backlog is filled with more than enough stories.
The system starts with a clean board and all column empty. After letting the system run for 75 simulated work days, we will trigger a policy change. Particularly the WiP limit for the ‘design’ is increased from ‘3’ to ‘5’. After this policy change the system runs for another 100 work days.
From the chart showing the average cycle time we will be able to study the effect of WiP limit changing adjustments.
The simulation assumes a simple uniform distribution for the amount of work done by the team and the effort assigned to a work item. I assume this is OK for the purpose of this blog. A consequence of this, is that the result probably can’t be scaled. For instance, the situation in which a column in the picture above is a single scrum team is not applicable since a more complex probability distribution should be used instead of the uniform distribution.
The picture below shows the result of running the experiment.
After the start it takes the system little over 40 work days to reach the stable state of an average cycle time of about 24(*) days. This is the cycle time one would expect. Remember, the ‘ready’ column has a size of ‘3’ and the other columns get work done. So, one would expect a cycle time of around 4 times 5,5 which equals 22 days which is close to 24.
At day 75 the WiP limit is changed. As can be inferred from the picture, the cycle time starts to rise only at day 100 (takes about 1 cycle time (24 days) to respond). The new stable state is reached at day 145 with an average cycle time of around 30(**) days. It takes 70 days(!) to reach the new equilibrium.
The chart shows the following interesting features:
- It takes roughly 2 times the (new) average cycle time to reach the equilibrium state,
- The response time (when one begins to notice an effect of the policy change) is about the length of the average cycle time.
(*) One can calculate (using transition matrices) the theoretical average cycle time for this system to be 24 days.
(**) Similar, the theoretical average cycle time of the system after policy change is 31 days.
In this blog we have seen that when a team makes adjustments that affect the flow, the system needs time to get to its new stable state. Until this state has been reached any new tuning of the flow is questionable. Simulations show that the time it takes to reach the new stable state is about 2 times the average cycle time.
For scrum teams that have 2-week sprints, the system may need about 2 months before new tuning of flow is effective. Meanwhile, the team can very well focus on other improvements, e.g. retrospectives that focus on the team aspect or collaboration with the team’s environment.
Moreover, don’t expect to see any changes in measurements of e.g. cycle time within the time period of the average cycle time after making a flow affecting change.
To summarise, after making flow affecting changes (e.g. increasing or decreasing WiP limits):
- Let the system run for at least the duration of the average cycle time so it has time to respond to the change,
- After it responds, notice the effect of the change,
- If the effect is positive, let the system run for another duration of the average cycle time, to get to the new stable state,
- If the effect is negative, do something else, e.g. go back to the old state, and remember that the system needs to respond to this as well!
 Agile manifesto, http://agilemanifesto.org/principles.html
 GetKanban, http://getkanban.com