As readers of this occasional blog know, this blog has been less of a 'web log' and more a series of small essays on the topic of development analytics. I have decided to start writing less formal entries more frequently and have realized I would be comfortable doing that on my own web site, murraycantor.com.
I want to be entirely clear. IBM has in no way looked over my shoulder in the writing of the blog and has been very generous in providing me a forum. Nevertheless, I will be freer sharing my opinions when there is no opportunity of confusing my often idiosyncratic opinions of those of the company's.
Since the last entry introducing the concept of liability, I have had the opportunity to discuss it on several occasions colleagues in IBM, In the course of this discussions I formulated what seems to be a useful way to explain the idea. In particular, I presented this idea at the Managing Technical Debt Workshop held on October 9. The following is a preview of what I will present as a lightning talk at the Cutter Consortium Summit next week.
Imagine an insurance agent comes into your office with the following offer: "Our company will indemnify your code against the following risks:
Excess support costs (above some deductible)
The policy will only cost $X a year. You realize that code insurance is much like auto liability insurance. In the auto case, the insurance protects you financially against the possible unfortunate outcome of driving the car, in the code case the insurance protects you against against some unfortunate outcome of running the code. So code liability insurance is like automobile liability insurance. This leads to the definition:
'Technical Liability' is the financial risk exposure over the life of the code.
(Thanks to my colleague, Walker Royce, for this crisp definition.)
Note auto insurance and code insurance have some significant differences.
The context for driving - city streets, highways, parking lots, ... - is more limited than the range of contexts that code can operate. Software is truly everywhere from which being embedded in an avionics system to Angry Birds on a smartphone.
The risk for auto insurance is spread among small numbers of large relatively homogenous populations: young drivers, safe drivers, high-risk drivers, etc. So rates can be computed from population experience. We have no such insurance markets for software.
Generally, firms faced with assuming a liability have a choice: Either they buy a policy indemnifying them against the risk or they self-insure. When they self insure, it is often reported in the annual reports.
If you ship software, you are assuming a liability. As far as I know, code insurance is either rare or nonexistent. If it did, the cost of the policy would be charged against the financial value code. So we are left with self-insuring,
Here is the main point. In order to truly assess the economic value of the code, one should, as best one can, estimate the technical liability and a fair price, X, for the indemnification. Even a rough estimate of X is better than ignoring the liabilities assumed by shipping code.
So how to estimate X? My first observation should be of no surprise to readers of this blog. Since technical liability involves the future, there are a range of outcomes of future exposure, each of which has some probability. Technical liability has a probability distribution and so is a random variable. X is a statistic (perhaps the mean) of the distribution.
As suggested above, code liabilities comes in flavors: There are exposures resulting from security, reliability, integrity, and so on. Each of these flavors is characterized by its own random variable. The overall liability is the sum of the liabilities that apply to the particular code. As I mention in a previous entry, this sum of random variables is also a random variable found using Monte Carlo simulation.
Now, reasoning about code liability is not unprecedented. Car manufacturers estimate warrantee exposure, telephone switch manufacturers reason about the economic value of going from .99999 reliable to .999999 reliable. There are Bayesian models of the likelihood of a security breach. To estimate technical liability, we need to agree upon the taxonomy of flavors of liability, not a daunting task, and then assemble good enough models of each into an overall framework.
Over the last couple of years I have been more or less following the technical debt community's discussion on what exactly is technical debt. Some ague that technical debt is limited to what it would cost to address deficiencies such as those found by code inspection tools such as Sonar. Other writers such as Chris Stirling introduce aspects or kinds of technical debt: quality debt, design debt; ....
My interpretation of the Ward Cunningham metaphor on incurring debt by shipping is broader, including the wide range of after-delivery costs. This entry is continue that discussion and suggest one path forward.
I argued that technical debt should reflect the fact that the very act of shipping software incurs all sorts of possible liabilities, any one of which may incur some future cost.
Future service costs
Executives getting on planes to deal with critical situations
Fines resulting from privacy violations
Loss of business from failing a compliance audit
Loss of intellectual capital due to security flaws
The nature of the liabilities very from domain to domain. Shipping the next rev of a mobile game like angry birds entails much less liability that next rev of avionic software for a commercial jet.
The costs of fixing the code may be the least of it and under-estimates the assumed labilites. Reasoning about whether these liabilities outweigh the benefits of shipping the code is key to the ship decision.
Since I wrote that entry I have been watching the technical debt space and see that I may be the minority, but not alone, with this perspective, Some people argue that technical debt is solely the cost of addressing shortfalls in the code. Others adopt a broader definition. In fact, in a conversation I had with Capers Jones, a long-time expert in software measurement, he shared a conversation he had with Ward discussing the same points. I have seen others make a distinction between software debt and technical debt. I have decided not to weigh in on this argument, but suggest we call all of the liabilities, (wait for it ...) technical liability.
There is a key difference between standardly-defined technical debt and technical liability: Technical debt involves code quality and can be determined. The liabilities involve possible future events and so entail predictions of the future. Some might even consider technical debt knowable and technical liability unknowable.
Readers of this blog know where I am going. Technical liability, unlike the more limited technical debt, involves a range of future possibilities and so each of the components of liability should be specified as a random variable with a probability distribution. The security violation might or might not occur. But if it does, the possible expense could sink the company. Reasoning about the risk takes some advanced techniques like setting the price of an insurance policy.
Finally, the economic decision if it makes sense to ship a piece of software, one needs to balance the value expected from the ship against the assumed liabilities. Note that the future value is also a random variable. In that case the decision to ship should be based on the techniques found here. I will elaborate the reasoning ibehind technical liability n a future blog (promise).
In summary then, technical liability gives a more complete picture of the economics of shipping a piece of code than technical debt, but it requires more sophisticated analysis.
An ongoing theme of this blog is that development processes differ from other business processes in that there is a wide range of uncertainty inherent in the efforts. It follows that tracking and steering development efforts entails ongoing predicting, from the evolving project information, when a project is likely to meet its goals.
Late last year, Nate Silver author of the Fivethrityeight blog and well know predictor of elections published The Signal and the Noise, a text for the intelligent layperson on how prediction works. I was impressed by the book as it explained the principles behind the sort of Bayesian analytics we need for development analytics without any explicit math. However, I felt for the folks in our field would greatly benefit by having the mathematical blanks filled in. So I decided to write a series of papers introducing the topics to folks who had some statistics and maybe some calculus in college, but not a solid background in prediction principles.
In my last blog, I laid out a vision of how a project lead and her stakeholders might use the predictive analytics to drive to better project outcomes. As I mentioned in the entry, IBM Rational is work on such a tool. A demonstration of this tool is found here: Agile Development Analytics Demo. The video was created by Peri Tarr, the lead architect on the project.
Some of you might notice that the terminology and development process described in the demo is at odds with your understanding of Agile. We do understand that and currently we are working on making a robust tool that accommodates a wide range of processes for what might some might call 'pure Agile' to various hybrids we are discovering in the market place.
In the next blog, I will explain more on how the tool works.
To now, this blog has been a series of essays on the theoretical considerations underlying the analytics of development. With this entry, I want to start changing the emphasis to the practicalities of building analytic tools. Going from theory to practice raises all kinds of issues: data content and formats, robustness of algorithms, reinforcing agile practices, .... To start that discussion, lets start with an epic on how an analytic tool for agile teams might work:
A lead of an agile team, call her Shirley. has been asked to deliver a mobile application, with a specified set of features, in time for the next world games, which is one year away. Understanding that the future is uncertain, Shirley treats the time to complete as the random variable. Before committing to the project, she needs an initial distribution of the time to complete the project. With such a distribution, she has a view of the probability of achieving the goal. It is the area under the distribution curve that lies to the left of the target date in Figure 1.
Figure 1. Probability distribution of delivering the Shirley’s Mobile app project
Fortunately she has tool called 'ARaVar' to help her build and maintain this distribution. This tool is federated with her OSLC agile project environment, Agilista (a fictional product). To use ARaVar, the team estimates the level of effort required for each feature using planning poker. In particular, for each feature’s level of effort the leadership team agrees on three values to enter in Agilista:
The low (best case) – Assumes all the stars align and the feature comes together easily to meet requirements.
The high (worse case) – Assumes Mr. Murphy S. Law and Ms. May Hem unexpectedly join the team and inject unexpected challenges and obstacles.
The nominal (most likely) – Assumes level of effort has the expected mix of good fortune and bad luck.
Behind the scenes, the ARaVar finds these inputs to in Agilista and uses them to define triangular probability distributions. In particular, AraVar interprets these effort inputs as saying
There is zero probability that the level of effort will be less than the best case.
There is zero probability that the level of effort will be greater than the worst case.
The greatest probability of the level of effort will be at the expected case.
So ARaVar sets the distributions to be zero below the low value and above the high value, with a peak at the expected case. Figure 2 show the resulting triangular distribution, setting the high and low to zero and setting the peak (expected case) so that the total area of the distribution in one.
Figure 2: Typical triangular distribution for each feature.
In the parlance of Bayesian reasoning, this technique provides the subject matter experts a means of arriving at an honest prior, based on current information and informed belief. If the difference between the low and high of the distribution of a feature is large, then the team is expressing its uncertainty of the effort required to deliver the feature. This gives Shirley’s team the opportunity to focus the team on resolving the uncertainties early, progressively de-risking the project.
With this prior estimate in place, Shirley has an idea of how likely it is she can make the commitment and she negotiates the content. What-if analysis in ARaVa provides her with capability to compute the impact of adding, changing or dropping one or more features from the program. Luckily, she does find that one of the relatively uncertain features is more of a nice-to-have than a must-have and adds considerably more risk than value. So she negotiates that feature out of scope for a firmer commitment to an earlier delivery in 11 months as illustrated in Figure 3.
Figure 3: The negotiated delivery commitment: earlier and more predictable.
So Shirley now is in a good place. She has agreement on the scope of the project between her team and her stakeholders. She feels her team has a good chance of delivering on time.
In the Agile fashion, work proceeds by establishing work items to deliver the features. These work items are scheduled for iterations/sprints, on an ongoing basis. As the team completes work items, they not only have less work to complete, but also have a track record of the actual time it takes the team to complete work (called team velocity). From a Bayesian perspective, these constitute important evidence of how well the project is actually executing. ARaVa queries Agilista for the completion status of the features, the work item burndown history, and updated effort-to-complete estimates for the remaining features. ARaVa uses modern predictive algorithms to update the time to complete distribution.
With these ongoing predictions, Shirley can discuss with her team, and external stakeholders, whether the odds of meeting the commitment are improving (as they should) or degrading. If the later is the case, she can use ARaVa to predict the impact of managing content (decommitting features) or adjusting resources. For example, the tool revealed that one feature was very much at risk. In discussion with the stakeholders, it was decided that this feature was necessary and so it was decided that for the next sprint there should be more resources focused on the this feature. Some staff were assigned to the team for just that sprint. With ARaVa, all stakeholders can have a more honest and trustworthy discussion on how best to proceed.
ARaVa does not yet exist, but it is not a dream. IBM Rational and Research are now in the process of developing such a tool for a possible delivery next year. We are calling the project AnDes (for Analytics of Development). AnDes uses state of the art learning algorithms. We do have working versions federated with Rational Team Concert (We did show a preview at last year’s Rational Innovate). In addition to consideration of automating the data collection, we are exploring how it can be applied across a wide range of projects:
Large to small
Innovative to complex
Fully or partially agile.
We are looking for design partners now! Interested? Please let me know at firstname.lastname@example.org.
In my previous couple of blog entries, I used triangular distributions for examples. For many who suffered through (or maybe enjoyed) their stat classes (what are the odds?), this might be a surprising choice. They were taught the default choice would be a Gaussian distribution. For those more attuned with modern business analytics, they are likely to be familiar with triangular distributions. In this entry, I'll briefly the reasoning beyond each of them.
First, as you hopefully recall, both are distributions associate with random variables (Those who don't recall migh benefit from the series of tutorials at The Khan Academy site). Each are non-negative functions with integral (area under the curve) one. (There are fancier mathematical definitions, but no matter.) Each describes the likelihood of each of set of possible outcomes of some random variable. The difference in shape between Gaussian (aka Normal) and triangular distributions reflects the nature and use of the random variables.
Briefly, normal distributions are often arise as the histogram of a set of measurements. They have some central value (called the mean) and some dispersion (called standard deviation) around the mean. Anyone who took a stat class studied these distributions. They show up in a many contexts:
The distribution resulting from tabulating the histogram of repeated, but imprecise measures of some quantity and then divided the entries by the sum of the measures is often assumed to be normal. The mean of the distribution is the estimator of the actual measure.
Statisticians like the normal distribution for several reasons. First, it is easy to parameterize. If you know the mean. mu (μ), and the standard deviation, sigma (σ), you have completely characterized the distribution. For example, the likelihood of a measurement occurring is often characterized as being within some number of σ's from the mean. Figure 1 shows how this works.
The likelihood of a value falling in a range is given by the area under the curve. For example, the probability of a value of the normally distributed random variable falling within one standard deviation of the mean is 68.2%.
Normal distributions have one really cool feature called the Central Limit Theorem, which states that under remarkably general conditions, the sum of a set of random variables will be close to normal. Notice, in the previous blog entry, when we added two triangular random variables, the sum appeared smooth and in fact started to look normal.
All that said, I do have have a pet peeve. Normal distributions are overused. Most things in nature and economics are not normally distributed. For example, as as documented in Wikipedia, these phenomena are nowhere near normal, but are closer to a Pareto distribution:
The sizes of human settlements (few cities, many hamlets/villages)
File size distribution of Internet traffic which uses the TCP protocol (many smaller files, few larger ones)
Hard disk drive error rates
The values of oil reserves in oil fields (a few large fields, many small fields)
The length distribution in jobs assigned supercomputers (a few large ones, many small ones)
The standardized price returns on individual stocks
Fitted cumulative Pareto distribution to maximum one-day rainfalls
Sizes of sand particles
Sizes of meteorites
Areas burnt in forest fires
Severity of large casualty losses for certain lines of business such as general liability, commercial auto, and workers compensation.
Getting back to our topic, let's turn to triangular distributions. They are not used to describe a set of measured outcomes from an experiment. They are used to describe what we know or believe about some unknown random variable. For example, the sales of a new product one year after delivery generally can not be determined by measuring the sales of a bunch of new products. As pointed out by Douglas Hubbard, treating the future sales as a single fixed variable is unreasonable (although all too common). What is more reasonable is setting the low (L), high (H) , and most likely (E) values of the future sales. As I wrote in an earlier entry, these are the values that specify a triangular distribution. I.e. triangular distributions are set to zero below a given low value, L, and above the high value, H, and peaks at the expected value E. The distribution is then a describe be a triangular curve so that the total area is 1. Here is the distribution for L = 1, E=6, and H=7.
Some would argue there is a 'real' distribution of the future sales random variable and it is unlikely to be triangular. My response is for all practical purposes, it does not matter. The triangular distribution is a good-enough approximation to whatever the real distribution might be. By 'good enough' I mean they may be used to support decision making: they are a big improvement over using single values. They are also practical as they easy to specify and there is no assumption of symmetry, No wonder they are common in business analytics.
To wrap up, normal distributions are occasionally useful to describe outcomes of measurements while triangular distributions are useful for giving rough estimates of one's belief of the liklihood of outcomes based on the evidence on hand. More generally, normal distributions are useful in frequentist statistics and triangular in Bayesian statistics. See this Wikepedia article for a discussion of the kinds of statistics. Much of what we do in development analytics is more Bayesian than frequentist. I hope to write more about that in the near future.
When I started this experiment in blogging, I wrote that I am not a natural blogger. I am not the affable, chatty web presence who on a daily or weekly basis shares one's thoughts. I have learned since then what kind of blogger I am. My style is to write little essays that might take weeks to prepare, given the priorities of my day job. I have also found I do enjoy writing the blog as it gives me a chance to share some issue that is top of mind. So here goes:
A few weeks ago, my good colleague displayed a chart in one of his PowerPoint decks entitled something like "How to Understand Murray." The chart was an explanation of probability distributions. It was both flattering and a bit of a wakeup call. As Arthur mentioned and the readers of this blog know, much, if not all, of my writing assumes an understanding of probability and probability distributions (aka probability densities). My experience in discussions with folks from our industry is that most of them have vague memories from some stat class in college and so can generally follow the discussions, but most could use a refresher. I could simply refer readers to a good Wikipedia article, but, instead, let me given a domain-specific example.
Let's go with a topic I wrote about in a previous entry: time to ship. For explanation purpose, let's take a fictional example. Suppose you are starting a project expected to ship in 110 days. That said, we cannot be 100% certain of being ready to ship on exactly that day, no sooner, no later. In fact, it is very unlikely we will exactly hit that day. Maybe we will be ready the day before, or maybe the day before that. Since being ready on or any day before day 110 is success, we can sum up the probabilities on being ready on any of those days to get the probability we really care about. All that said, the probabilities for each of the days matter because we need them in order to get the sum. The set of probabilities for each of those days is the probability distribution or density that is our topic.
Let's look at a simple example.
In this example of a triangular distribution there is 0% probability that the product will be ready before day 91 and we are 100% certain we will be ready before day 120. We think the days become more probable and reach a peak as we approach day 110 and fall off after that. This graph then shows the probability, day by day, of being ready on exactly that day. You might notice that the peak is less than 0.07 (actually 0.06666...). This makes sense since we are assuming that the project may be complete on any of 30 different days and so the densities would be in the neighborhood of 1/30 = 0.03333. In our case, some are above and some below.
These distributions are the basis of calculating the likelihood of outcomes. The principle is very simple: The probability of being ready within some range of dates is the sum of the probabilities of being ready on exactly one of those dates, i.e. , we add up the density values for those days. As I explained above, if we want to compute the probability of being ready on or before day 110, we would add up all of the densities for days 90 to 110 to get 0.7. Using the same reasoning the probability of being ready on some day before day 120 is the sum of all the densities which comes to exactly 1.0, which was one of our going in assumptions. In fact the property that the sum of densities for all possible outcomes equals 1 is a defining property of distributions. Those who want to try this out at home could use this spreadsheet.. For example, can you find on what day, being ready on or before that day is an even bet?
For most development efforts, the overall state of the program (some would say 'health') is characterized by the shape of the distribution, This shape changes every day. Every action the team takes changes the shape. So, one key goal of development analytics would be to track the shape of the distribution throughout the lifecycle, a daunting task. More on this (probably ) in future postings.
One of the common criticisms of estimation methods is that
the calculation is no better than the assumptions: garbage in, garbage out (affectionately known as GIGO). That is, if you make poor or
dishonest assumptions then you will get misleading forecasts. It is especially egregious
that occasionally someone might take advantage of the system by gaming the system
and intentionally feeding assumptions that lead to false forecasts to get a
desired business decision.
However, estimation is an essential part of any disciplined
funding decision process (such as program portfolio management). The funding decision
relies on estimates of the costs and benefits. But for reasons just described,
estimation is suspect.
So, what to do? I suggest the answer is not to abandon
estimation; the answer is the not input garbage, or if you do, detect it as
soon as possible to minimize the damage.
First note that the future costs and benefits are uncertain,
so any serious approach to the GIGO
problem is to treat the assumptions as random variables with probability distributions and work from there. Generally, this allows one to use the limited information at hand to enter the assumptions and calculate
Douglas Hubbard, in How to Measure Anything, gives us one way to proceed. Briefly, when an uncertain
value is needed, ask the subject matter expert (sme) to give not one but three values:
low, high, and expected. The three values may be used to specify random
variables with triangular distributions [ref].
In this case, the greater the difference between the high
and low values, the wider the triangular distribution of the estimate reflecting
the uncertainty of the sme who is honestly making the assumptions.
One can use the random variables as values in the estimation
algorithm using MonteCarlo by repeatedly
replacing the single values with sampled values of the triangular distributions
and assembling the distribution of the estimated value. Note the estimate is
again just as good as the assumptions, however we assess our faith in the
estimate by the width of the 10%-90% range of its distribution.
For example, one might estimate to the total time for
completing s project by a project by entering, for each task, the least time,
the most time, and the most likely time. Then one could apply Monte Carlo
simulation or more or
more elementary methods to rollup the estimates to compute the distribution
of the time to complete.
Hubbard goes further by suggesting that as actuals in the
assumptions come available to review if they fall within the 10% -90% range of
the initial distributions.If they do,
fine. If they don’t, questions are asked about the underlying reasoning and
beliefs. Over time the organization becomes more capable and accountable at
making good assumptions.
Further, we can also deal with the garbage in garbage out
problem by using actual data whenever possible. There are at least two techniques.
In the first, as actuals in the assumptions become available
in the, they can used to replace the distributions. For example if there are
month-by-month sales projections captured as triangular distributions to
forecast sales volumes, the distributions are replaced by the actual sales numbers.Also, one should update the remaining triangular
distributions reflecting the actual sales trends. The resulting estimate will usually
have a narrower distribution.
A second technique is Bayesian trend analysis. In this case
we use actuals for evidence of the estimate. For example, if a project were on
track, then we can expect that certain measures, such as burn down rate and
test coverage reflect that. If a project were to ship on time, the number
unimplemented requirements would be going to zero, Similarly, the code coverage
measure would be trending towards the target. So these measures are evidence of a healthy
project. Using Bayesian trend analysis, we
can turn the reasoning around and update the initial (prior) estimate of the
time for completion using the actuals as evidence for an improved estimate. The
result is an improved probability distribution of the time to complete the
project. As more actuals become available, the distribution becomes narrower,
increasing the certainty of the forecast.
This way one can detect early if the system is being gamed
and at the same time, use the actuals to estimate the likelihood of an on-time
So generally, one can use actuals to not only improve the
estimation process as Hubbard suggests, but also to apply Bayesian techniques,
to improve the estimates of the program variables.
In the previous entry, I introduced a probabilistic view of a commitment. The main idea is that when you commit to deliver something in a future, you are making a kind of bet. The odds of winning the bet is the fraction of the distribution of the time=to-deliver before the target date. For example, in the following example, the project manager has a 47% likelihood of winning the bet.
The raises a couple of questions. First, how is the distribution of time-to-complete determined? There are variety of methods to estimate time to complete of an effort. I am not taking a position on what method to adopt. The important point is that the estimation method should not return a number but a distribution! The major estimation vendor have this capability even if it not always surfaced. I will expand on this point in the next blog entry. For now, the key point is that you should be working not with point estimations, but with the distributions.
Second is how the project manager affects the shape and position of the distribution and therefore affects the odds. Some of the techniques are intuitive, some not so much, There are two things one might do: move the distribution relative to the target date, and change the shape if distribution typically narrowing it so that more of it .lies within the target date.
In the first, one can either move the target date out, so that the picture looks like this
This is, of course, intuitive - moving out the date lowers the risk. Another intuitive thing a project manager might do is the descope the project - commit to deliver less functionality. This may have two effects on the distribution: It will move it to the left as there will be less work to do. Depending on the difficulty of the descoped feature, the descoping may also narrow the distribution. By removing a difficult to implement feature. one is more certain of delivery, narrowing the distribution, removing risk resulting in this diagram:
Now comes the unintuitive part. Suppose the target date and content are not negotiable. What is a project manager to do then? The idea is to take actions that will narrow the distribution in Figure 1 so that it looks like
How is this done? Many project managers, in the name of making progress choose the easiest functions to implement first, "the low hanging fruit". However, by doing this the shape of the curve in figure in minimally affected, The less intuitive approach, Following the principle of the Ration Unified Process, is to work on the most difficult, riskiest requirements first! These are the requirements of which
the team has the least information and so should tackle first in order to have time to gain the information needed to succeed. Putting off the riskier requirements and doing the easy stuff first gives the appearance of progress, but by putting off the riskier requirements, one will run out time to do the riskier requirements and fail to meet the commitment.
All this has to be while ensuring their is sufficient time to fulfill all the requirements, risky or not. So in the end, one must account for both the time to complete tasks and their uncertainty to meet commitments. Some techniques for doing that will be discussed in the next blog entry.
I know. It has been a long time since my last posting. Over the last few months, nothing much happened that prompted an entry. I was thinking of writing something sort of philosophical on the nature of estimates, but never got to it. Then there was the tsunami and the associated nucler reactor failures at the Fukushima power plant. Suddenly, the topic became more urgent. This is relevant to this blog, because our domain includes the engineering and economics of safety critical systems. Presumably the nuclear reactor industry uses the state of the art methods. I have been exploring what is going on there and while, I am far from an expert, I have found out some things worth sharing in a blog.
We have been told that reactor failure is a 1 in over a hundred thousand year event. Sounds reasurring.Yet, in my lifetime, there have been three that I know of: Three mile island, Chernoble, and now Fukushima. Discounting Chernobyl which apparently was greatly under-engineered, something must be wrong for the there to be two meltdowns in what has been estimated to be a one in over 100,000 year event. Apparently, there have been many near misses, e.g. the loss of coolant at the Brown's Ferry plant, something one would not expect from such safe systems. This raised some questions. What does a 'one is N year event' mean? Does it mean that we should not expect the event until N years has passed or that we can be certain one will occur within N years? More importantly, if there are K systems, each having a 1 in N year safety rating, what is the rating of the poputation? As I have pointed out in previous blogs, we do not need estimates, we need probablility distributions to get any practical understanding.
Here are some of the sources I found. This New York times article helps explain what is going on. A few points in the article caught my attention. First, they carried out 'deterministic' risk analysis (see this NRC page) because probabilistic methods are "too hard. A good summary of the difficulty of the problem and the history of how it is addressed is found in Uncertainty: A Guide to Dealing with Uncertainty in Quantitative Risk and Policy Analysis by M. Granger Morgan and Max Henrion. Briefly there is little to go on to estimate the likelihood of individual events and their dependencies in event chains. So the distribution of estimated time to failure must have huge variance. This article by M. V. Ramana summarizes and criticizes the current practice including probablistic risk assessment (pra). The key idea is that failure results from a sequence of component failures. Each component is reliable and so the probability of a system failure is the joint probability of the component failures which is very low. This assumes that the component failures are independent events. However, as parts of thr system, the joint probabilities are hard to estimate. For example, one component may fail which results in a second component running out of spec which might result in a number of other component failures. Getting the joint probabilies right entails a very faithful system model and data collected from thousands of simulations with varying inputs and Monte Carlo methods to take into account the variability of the components. The output of such a simulation could be used to improve the system design.
I want to focus on another pra challenge: estimating the likelihood of the devastating single cause event, such as earthquake and the tsunami. Clearly some sort of data are needed for the estimate, but what sort of data? As pointed out in the New York Times article, there was a deep historical data search of the size and frequency of the earthquakes in the relevant geographic region. That led to the conclusion that planning for 18 feet of water was sufficient. Recall that the reactor was inundated by 40 feet of water. So past performance was not predictive. In retrospect, that is not surprising given that tectonic plate movements are hardly a stationary process. An alternate approach is to use modern geologic models and plate measurements. Then one could and should run simulations to get a distribution of the flood depths. One could argue that this approach is also suspect, since they depend on the quality of the models which introduces a subjective element. However, using historical data also is based on an assumption that earthquake generation is based on a stationary process, a very dubious model and its adoption is equally subjective. To be fair, presumably one could not run the needed simulations in the 1960's and so the frequency model may have been the best available. It can be argued that earthquake prediction is notoriously difficult, especally pinpointing when an earthquake will occur. However, using Monte Carlo methods and simulations, it seems reasonable that one can create a probility distribution of the time to an earthquake above a certain size and use this to estimate the likelihood of the event over the lifespan of the plant.
The point of this discussion is that frequency model data is no more 'objective' than data used to build and apply models. Both involve subjective assumptions of the validity of the model. Note Baysian data analysis methods can be used to validate the various models and so we can assess their usefulness in the estimation process.
Finally, these safety estimates are used to set policy and, in particular, to make economic decisions about nuclear energy. The cost of a failure is huge. For example, an estimate of the cost of Fukushima failure is $184 Billion. The proponants of nuclear energy argue that they make economic sense, assuming they are safe enough and the new designs are much safer. Maybe so. But knowing they are safe enough will take much better analytics than we have seen to date.
It should not be a surprise that I have been following the
BP oil spill with much interest. In fact, as I starting typing this entry, I
was watching the grilling of the BP CEO, Tony Heyward, by Congress. Rep. Stupak
is focusing on the BP’s risk management.
Some of you have read my earlier posting on my thoughts of
the BP decision process that led to the Deepwater Horizon blowout. So far,
information uncovered since that posting is remarkably consistent with my
earlier suppositions. In this entry I would like to step back a bit and discuss
what broader lessons might be learned from the incident. While it is all too
easy to fall into BP bashing, I would rather use this moment to reflect more
deeply on risk taking and creating value. (BTW, some of you might now that my
signature slogan is ‘Take risks, add value’.)
In our industry we create value primarily through the
efficient delivery of innovation. Delivering innovation, by definition,
requires investing in efforts without initial full knowledge of the effort
required and the value of the delivery. This incomplete information results in
uncertainties in the cost, effort, schedule of the projects and the value of
the delivered software and system, i.e. cost, schedule and value risk.
Deciding to drill an oil well also entails investing in an
effort with uncertain costs and value. In this case, the structure of the
subsystem and productivity of the well cannot be know with certainty before
drilling. As I pointed out in an earlier blog, a good definition of risk is
uncertainty in some quantifiable measure that matters to the business. So in
both our industry and oil drilling we deliberately assume risk to deliver
So, what can we learn from the BP incident? Briefly, one
creates value by genuinely managing risk. One creates the semblance of value
for a while by ignoring risk.
Assuming risk, investing in uncertain projects, provides the
opportunity for creating value. That value is actually realized by investing in
activities that reduce the risk. The model that shows the relationship is described
in this entry. So, reducing risk has economic value, but reducing risk
takes investment. In the end, the quality risk management is measured with a
return on investment calculation. This in turn requires a means to quantify and
in fact monetize risk.
I wonder what was there risk management approach was
followed by BP. A recent Wall Street Journal article suggested they used a risk
map approach – building a diagram with one axis a score of the ‘likelihood of
the risk’ and the other a score of the ‘severity of a failure’. So with this
method, they would score the risk of a blowout as very low (based on past
history) with a very high consequence. So, such a risk needs to be ‘mitigated’.
(Some actually multiply the scores to get to some absolute risk measure.) Their
mitigation was the installation of a blow-out preventer. They could then
confidently report they have executed their risk management plan. Note these
scores are at best notionally quantified and not monetized.
Paraphrasing my good colleague, Grady Booch (speaking of
certain architecture frameworks), risk maps is the semblance of risk
management. As pointed out by Douglis Hubbard in The Failure of Risk Management (and in
an earlier rant in this blog), this sort of risk management is not only
common, but dangerous: It is a sort of business common failure mode that leads
to bad outcomes. Also, Hubbard points out, useful risk management entails
quantification and calculation using probability distributions and Monte Carlo
analysis. I would add that since risk management in the end is about business
outcomes, risks need to be monetized as well as quantified. I am willing to bet
a good bottle of wine that BP did no such thing. Any takers? The business
common failure mode was over-reliance on the preventers, even though there are
several studies showing they are far from ‘failsafe’.
Further, it appears BP assumed risk by consistently taking
the cheaper, if riskier. design and procedure alternative, the one with greater
uncertainty in the outcome, even when the cost of an undesired, if unlikely,
outcome was possibly catastrophic.The laundry list of such decisions is long; some outlined in Congressman
Waxman’s letter to Tony Hayward.CEO’s of Shell and Exxon testified before congress that their companies
would have used a different, more costly designs and followed more rigorous
procedures. According the congressional and journalistic reports, this behavior
is BP standard operating procedure. So BP assumed risk by drilling wells but did not invest in reducing
For quite a while they got away with the approach of
assuming but really reducing risk, and appeared to be creating value as
reflected in stock and value and dividends to the investors. The BP management
raised the stock price from around $40/share in 2003 to a peak of around of
$74/share prior to the Deepwater Horizon incident. At this writing the stock is
trading at $32/share and the current dividend has been cancelled. Investors
might rightly wonder if there is another latent disaster and so discount the
apparent future profitability with the likelihood of unknown liabilities. The
total loss of stockholder value is over $100B, which is in the ballpark of the
eventual liability of BP. So, whatever approach BP used to manage risk failed.
BTW, some may recognize this same pattern in the management
of financial firms that participated in the subprime mortgage market. In that
case, they ‘mitigated risk’ by relying on the ratings agencies. Those who
actually built monetized models of the risk realized there was a great
opportunity to bet against the subprime mortgage lenders and made huge fortunes
(See, e.g. The Big Short: Inside the
Doomsday Machine by Michael Lewis .).
Readers of the blog will notice a recurrent theme is some of
the postings. It is essential that we assume and manage risk. To repeat a
favorite quote, “One cannot manage what one does not measure.” The risk map,
score methods, while common are insufficient to the needs of our industry; they
do not measure, nor really manage risk. We as a discipline need to step up to
quantifying, monetizing, and working off risk in order to be succeed as drivers
of innovation. We need to step up to the mathematical approach found in the
Douglas and Dan Savage’s (see
this posting) texts.
I came to this same realization probably a decade ago. I
held off at first because I had not deep enough understanding of how to
proceed, and I knew I would encounter great skepticism. I tested the waters in
2005 and posted my first
paper on the subject in 2006. I indeed received a great deal of skepticism
and resistance, but enough acceptance to go forward. I have learned some
important lessons from all that. In my next blog, I will share my experiences
of bringing more mathematical thinking to risk management for SSD.
This blog experiment seems to be working. The entries are
gietting around 100 visits and growing - good enough to keep at it. I have
found that writing the entries has given me the opportunity to clarify and
express my thoughts. This entry is a case in point.
We are deploying a BAO solution for the level 3 support
organizations in our IBM India Software Labs. That deployment provides a case
study in how to integrate two concepts I introduced in earlier blogs. This
entry is longer than the others. I hope you find it worth the wait and effort
In those previous entries, I discussed two frameworks for reasoning
These frameworks address different aspects of the problem of
using measures to achieve business goals by measuring the right things and
taking actions to respond to the measurements. In fact, these frameworks fit
together hand and glove.
that level 3 support teams provide fixes to defects found in delivered
code.Each of the teams deals with
an ongoing series of change requests (aka APARs, PMRs). An organization goal is to reduce the time to and cost of completion of these
requests. To achieve the goal, they are adopting some Rational-supported practices
and supporting tools. So the questions
that need to be answered are:
is the time trend of the time to complete of the change requests?
is the time trend of the cost to complete of the change requests?
each case how would I know that some improvement action resulting in
significant improvement in the trends?
comes the hard part: determining the measures
that answer the questions. The change requests come arrive somewhat
unpredictably. Each goes through the fix and release process and presumably
gets released in a patch or point release.So at any given time there is a population of currently open
and recently closed releases. The measures that answer the question are a time
trend of some statistic on the population on some population of change requests.
of the change requests requires different amount of time and effort to
complete. So to measure if the outcome is being achieved, one must reason
statistically: defining populations of requests, building the statistical
distribution of say time to complete for that population, defining the outcome
statistic for the distribution.So
we need to do things to define the measure:
the population of requests for each point on the trend line
the statistics on that population
keep it simple (as least as simple as possible), lets form the population by
choosing the set of change requests closed in some previous period, say the
previous month or quarter. To choose a statistic, one needs to look at the data
and pick the statistic that best answers the question. Most people assume the
mean of the time (or cost) to complete is the best choice.However, that choice is appropriate
when the shape of the histogram of the time to complete is centered on a mean
as is common in normally distributed data.
of the advantages of working in IBM is that we have lots of useful data. Inspection
of some APAR data of the time to complete from one of our teams in the IBM
Software Lab in India shows the distribution is not centered on a mean. and so reduction of the mean time to
complete is not the best measure of improvement.
have looked at literally tens of thousands of data points for time to complete
of change requests across all of IBM and have found the same distribution. For
you statistics savvy, it appears to be a Pareto Distribution, but statistical
analysis carried out by Sergey Zeltyn of IBM Research’s Haifa lab shows that
this distribution does not well fit any standard distribution. A possible
explanation is that is the time required to fix the defects is Pareto
distributed, but since the resources available to fix them is limited, the
actual time to complete is not pure Pareto. In any case, a practical way to
proceed is to choose a simple (non-parametric) measure: width of the head, i.e.
the time it takes to complete 80% of the distributions.
with this analysis in place, the organization decides to precisely specify the goal such as a 15% reduction in time and
cost to complete 80% of the requests closed each month.So the outcome measures are the time it took to close and costs of 80% of
the requests closed each month.
chosen this measures, we are ready to identify the data sources and instrument the measures. So far so good. But wait, we still need
to answer questions 3.
I mentioned, in order to improve the outcome measure and achieve the goals, the
lab teams have agreed to adopt appropriate Rational practices and tools to automate certain processes. The
practices were selection using the Rational MCIF Value Tractability Trees (a
development causal analysis methed). Adopting and maturing the practices and
their automations are the controls. Some
control examples are automating the regression test and build process, and the adoption
of a stricter unit test discipline to reduce time lost in broken builds. There
are control mechanisms with associated control
measures such as time-to-build, regression test time-to-complete, percent
of code unit-tested, and a self-assessment by the team of their adoption of testing
and build practices.
answer question 3, we need statistical analytics
to determine if the changes in the control measures have had a significant impact
on the outcome measures. Our Research staff has settled on those analytics, but
I will discuss that in a later entry. This entry is already too long.
case study is both reasonably straightforward and far from trivial. It does
show as promised that GQM(AD) and Outcome and Controls work together. I leave
you all with a thought problem. How would you apply the pattern to teams
developing new features to existing applications?
I am not a natural blogger, but do like to share thoughts with a broad community of folks with shared interests.
My longterm passion is how to enable organizations develop and deliver software and systems more effectively. I am thoroughly convinced that well organized, governed organizations not only deliver value to the businesses, but also enhance the lives of the staff. In such organizations, people get to work on cool things, innovate, work well with colleagues, and build a legacy by being a part of making things of value.I also am a mathematician by training. I am especially energized by the opportunity to apply mathematical reasoning to the improvement of software and system organizations. My current assignment is to lead the Business Analytics and Optimization (BAO) strategy for Rational.
So, with this blog, from time to time, I will share my thoughts on BAO for software and system organizations. I hope this blog will be a catalyst for building a community. I especially look forward to comments and conversations.