Somewhere, right now, a technology executive tells their directors: “we
need a way to measure the productivity of our engineering teams.” A working
group assembles to explore potential solutions, and weeks later, proposes
implementing the metrics: lead time, deployment frequency, and number of
pull requests created per engineer.
Soon after, senior engineering leaders meet to review their newly created
dashboards. Immediately, questions and doubts are raised. One leader says:
“Our lead time is two days which is ‘low performing’ according to those
benchmarks – but is there actually a problem?”. Another leader says: “it’s
unsurprising to see that some of our teams are deploying less often than
others. But I’m not sure if this spells an opportunity for improvement.”
If this story arc is familiar to you, don’t worry – it’s familiar to
most, including some of the biggest tech companies in the world. It is not uncommon
for measurement programs to fall short when metrics like DORA fail to provide
the insights leaders had hoped for.
There is, however, a better approach. An approach that focuses on
capturing insights from developers themselves, rather than solely relying on
basic measures of speed and output. We’ve helped many organizations make the
leap to this human-centered approach. And we’ve seen firsthand the
dramatically improved understanding of developer productivity that it
provides.
What we are referring to here is qualitative measurement. In this
article, we provide a primer on this approach derived from our experience
helping many organizations on this journey. We begin with a definition of
qualitative metrics and how to advocate for them. We follow with practical
guidance on how to capture, track, and utilize this data.
Today, developer productivity is a critical concern for businesses amid
the backdrop of fiscal tightening and transformational technologies such as
AI. In addition, developer experience and platform engineering are garnering
increased attention as enterprises look beyond Agile and DevOps
transformation. What all these concerns share is a reliance on measurement
to help guide decisions and track progress. And for this, qualitative
measurement is key.
Note: when we say “developer productivity”, we mean the degree to which
developers’ can do their work in a frictionless manner – not the individual
performance of developers. Some organizations find “developer productivity”
to be a problematic term because of the way it can be misinterpreted by
developers. We recommend that organizations use the term “developer
experience,” which has more positive connotations for developers.
What is a qualitative metric?
We define a qualitative metric as a measurement comprised of data
provided by humans. This is a practical definition – we haven’t found a
singular definition within the social sciences, and the alternative
definitions we’ve seen have flaws that we discuss later in this
section.
Figure 1: Qualitative metrics are measurements derived from humans
The definition of the word “metric” is unambiguous. The term
“qualitative,” however, has no authoritative definition as noted in the
2019 journal paper What is Qualitative in
Qualitative Research:
There are many definitions of qualitative research, but if we look for
a definition that addresses its distinctive feature of being
“qualitative,” the literature across the broad field of social science is
meager. The main reason behind this article lies in the paradox, which, to
put it bluntly, is that researchers act as if they know what it is, but
they cannot formulate a coherent definition.
An alternate definition we’ve heard is that qualitative metrics measure
quality, while quantitative metrics measure quantity. We’ve found this
definition problematic for two reasons: first, the term “qualitative
metric” includes the term metric, which implies that the output is a
quantity (i.e., a measurement). Second, quality is typically measured
through ordinal scales that are translated into numerical values and
scores – which again, contradicts the definition.
Another argument we have heard is that the output of sentiment analysis
is quantitative because the analysis results in numbers. While we agree
that the data resulting from sentiment analysis is quantitative, based on
our original definition this is still a qualitative metric (i.e., a quantity
produced qualitatively) unless one were to take the position that
“qualitative metric” is altogether an oxymoron.
Aside from the problem of defining what a qualitative metric is, we’ve
also encountered problematic colloquialisms. One example is the term “soft
metric”. We caution against this phrase because it harmfully and
incorrectly implies that data collected from humans is weaker than “hard
metrics” collected from systems. We also discourage the term “subjective
metrics” because it misconstrues the fact that data collected from humans
can be either objective or subjective – as we discuss in the next
section.
| Type | Definition | Example |
|---|---|---|
| Attitudinal metrics | Subjective feelings, opinions, or attitudes toward a specific subject. | How satisfied are you with your IDE, on a scale of 1–10? |
| Behavioral metrics | Objective facts or events pertaining to an individual’s work experience. | How long does it take for you to deploy a change to production? |
Later in this article we provide guidance on how to collect and use
these measurements, but first we’ll provide a real-world example of this
approach put to practice
Peloton is an American technology company
whose developer productivity measurement strategy centers around
qualitative metrics. To collect qualitative metrics, their organization
runs a semi-annual developer experience survey led by their Tech
Enablement & Developer Experience team, which is part of their Product
Operations organization.
Thansha Sadacharam, head of tech learning and insights, explains: “I
very strongly believe, and I think a lot of our engineers also really
appreciate this, that engineers aren’t robots, they’re humans. And just
looking at basic numbers doesn’t drive the whole story. So for us, having
a really comprehensive survey that helped us understand that entire
developer experience was really important.”
Each survey is sent to
a random sample of roughly half of their developers. With this approach,
individual developers only need to participate in one survey per year,
minimizing the overall time spent on filling out surveys while still
providing a statistically significant representative set of data results.
The Tech Enablement & Developer Experience team is also responsible for
analyzing and sharing the findings from their surveys with leaders across
the organization.
For more on Peloton’s developer experience survey, listen to this
interview
with Thansha Sadacharam.
Advocating for qualitative metrics
Executives are often skeptical about the reliability or usefulness of
qualitative metrics. Even highly scientific organizations like Google have
had to overcome these biases. Engineering leaders are inclined toward
system metrics since they are accustomed to working with telemetry data
for inspecting systems. However, we cannot rely on this same approach for
measuring people.
Avoid pitting qualitative and quantitative metrics against each other.
We’ve seen some organizations get into an internal “battle of the
metrics” which is not a good use of time or energy. Our advice for
champions is to avoid pitting qualitative and quantitative metrics against
each other as an either/or. It’s better to make the argument that they are
complementary tools – as we cover at the end of this article.
We’ve found that the underlying cause of opposition to qualitative data
are misconceptions which we address below. Later in this article, we
outline the distinct benefits of self-reported data such as its ability to
measure intangibles and surface critical context.
Misconception: Qualitative data is only subjective
Traditional workplace surveys typically focus on the subjective
opinions and feelings of their employees. Thus many engineering leaders
intuitively believe that surveys can only collect subjective data from
developers.
As we describe in the following section, surveys can also capture
objective information about facts or events. Google’s DevOps Research and
Assessment (DORA) program is an excellent concrete
example.
Some examples of objective survey questions:
- How long does it take to go from code committed to code successfully
running in production? - How often does your organization deploy code to production or
release it to end users?
Misconception: Qualitative data is unreliable
One challenge of surveys is that people with all manner of backgrounds
write survey questions with no special training. As a result, many
workplace surveys do not meet the minimum standards needed to produce
reliable or valid measures. Well designed surveys, however, produce
accurate and reliable data (we provide guidance on how to do this later in
the article).
Some organizations have concerns that people may lie in surveys. Which
can happen in situations where there is fear around how the data will be
used. In our experience, when surveys are deployed as a tool to help
understand and improve bottlenecks affecting developers, there is no
incentive for respondents to lie or game the system.
While it’s true that survey data isn’t always 100% accurate, we often
remind leaders that system metrics are often imperfect too. For example,
many organizations attempt to measure CI build times using data aggregated
from their pipelines, only to find that it requires significant effort to
clean the data (e.g. excluding background jobs, accounting for parallel
jobs) to produce an accurate result
The two types of qualitative metrics
There are two key types of qualitative metrics:
- Attitudinal metrics capture subjective feelings, opinions, or
attitudes toward a specific subject. An example of an attitudinal measure would
be the numeric value captured in response to the question: “How satisfied are
you with your IDE, on a scale of 1-10?”. - Behavioral metrics capture objective facts or events pertaining to an
individuals’ work experiences. An example of a behavioral measure would be the
quantity captured in response to the question: “How long does it take for you to
deploy a change to production?”
We’ve found that most tech practitioners overlook behavioral measures
when thinking about qualitative metrics. This occurs despite the
prevalence of qualitative behavioral measures in software research, such
as the Google’s DORA program mentioned earlier.
DORA publishes annual benchmarks for metrics such as lead time for
changes, deployment frequency, and change fail rate. Unbeknownst to many,
DORA’s benchmarks are captured using qualitative methods with the survey
items shown below:
Lead time
For the primary application or service you work on,
what is your lead time for changes (that is, how long does it take to go
from code committed to code successfully running in production)?
More than six months
One to six months
One week to one month
One day to one week
Less than one day
Less than one hour
Deploy frequency
For the primary application or service you
work on, how often does your organization deploy code to production or
release it to end users?
Fewer than once per six months
Between once per month and once every six months
Between once per week and once per month
Between once per day and once per week
Between once per hour and once per day
On demand (multiple deploys per day)
Change fail percentage
For the primary application or service you work on, what
percentage of changes to production or releases to users result in
degraded service (for example, lead to service impairment or service
outage) and subsequently require remediation (for example, require a
hotfix, rollback, fix forward, patch)?
0–15%
16–30%
31–45%
46–60%
61–75%
76–100%
Time to restore
For the primary application or service you work on, how long
does it generally take to restore service when a service incident or a
defect that impacts users occurs (for example, unplanned outage, service
impairment)?
More than six months
One to six months
One week to one month
One day to one week
Less than one day
Less than one hour
We’ve found that the ability to collect attitudinal and behavioral data
at the same time is a powerful benefit of qualitative measurement.
For example, behavioral data might show you that your release process
is fast and efficient. But only attitudinal data could tell you whether it
is smooth and painless, which has important implications for developer
burnout and retention.
To use a non-tech analogy: imagine you are feeling sick and visit a
doctor. The doctor takes your blood pressure, your temperature, your heart
rate, and they say “Well, it looks like you’re all good. There’s nothing
wrong with you.” You would be taken aback! You’d say, “Wait, I’m telling
you that something feels wrong.”
The benefits of qualitative metrics
One argument for qualitative metrics is that they avoid subjecting
developers to the feeling of “being measured” by management. While we’ve
found this to be true – especially when compared to metrics derived from
developers’ Git or Jira data – it doesn’t address the main objective
benefits that qualitative approaches can provide.
There are three main benefits of qualitative metrics when it comes to
measuring developer productivity:
Qualitative metrics allow you to measure things that are otherwise
unmeasurable
System metrics like lead time and deployment volume capture what’s
happening in our pipelines or ticketing systems. But there are many more
aspects of developers’ work that need to be understood in order to improve
productivity: for example, whether developers are able to stay in the flow
or work or easily navigate their codebases. Qualitative metrics let you
measure these intangibles that are otherwise difficult or impossible to
measure.
An interesting example of this is technical debt. At Google, a study to
identify metrics for technical debt included an analysis of 117 metrics
that were proposed as potential indicators. To the disappointment of
Google researchers, no single metric or combination of metrics were found
to be valid indicators (for more on how Google measures technical debt,
listen to this interview).
While there may exist an undiscovered objective metric for technical
debt, one can suppose that this may be impossible due to the fact that
assessment of technical debt relies on the comparison between the current
state of a system or codebase versus its imagined ideal state. In other
words, human judgment is essential.
Qualitative metrics provide missing visibility across teams and
systems
Metrics from ticketing systems and pipelines give us visibility into
some of the work that developers do. But this data alone cannot give us
the full story. Developers do a lot of work that’s not captured in tickets
or builds: for example, designing key features, shaping the direction of a
project, or helping a teammate get onboarded.
It’s impossible to gain visibility into all these activities through
data from our systems alone. And even if we could theoretically collect
all the data through systems, there are additional challenges to capturing
metrics through instrumentation.
One example is the difficulty of normalizing metrics across different
team workflows. For example, if you’re trying to measure how long it takes
for tasks to go from start to completion, you might try to get this data
from your ticketing tool. But individual teams often have different
workflows that make it difficult to produce an accurate metric. In
contrast, simply asking developers how long tasks typically take can be
much simpler.
Another common challenge is cross-system visibility. For example, a
small startup can measure TTR (time to restore) using just an issue
tracker such as Jira. A large organization, however, will likely need to
consolidate and cross-attribute data across planning systems and deployment
pipelines in order to gain end-to-end system visibility. This can be a
yearlong effort, whereas capturing this data from developers can provide a
baseline quickly.
Qualitative metrics provide context for quantitative data
As technologists, it is easy to focus heavily on quantitative measures.
They seem clean and clear, afterall. There is a risk, however, that the
full story isn’t being told without richer data and that this may lead us
into focusing on the wrong thing.
One example of this is code review: a typical optimization is to try to
speed up the code review. This seems logical as waiting for a code review
can cause wasted time or unwanted context switching. We could measure the
time it takes for reviews to be completed and incentivize teams to improve
it. But this approach may encourage negative behavior: reviewers rushing
through reviews or developers not finding the right experts to perform
reviews.
Code reviews exist for an important purpose: to ensure high quality
software is delivered. If we do a more holistic analysis – focusing on the
outcomes of the process rather than just speed – we find that optimization
of code review must ensure good code quality, mitigation of security
risks, building shared knowledge across team members, as well as ensuring
that our coworkers aren’t stuck waiting. Qualitative measures can help us
assess whether these outcomes are being met.
Another example is developer onboarding processes. Software development
is a team activity. Thus if we only measure individual output metrics such
as the rate new developers are committing or time to first commit, we miss
important outcomes e.g. whether we are fully utilizing the ideas the
developers are bringing, whether they feel safe to ask questions and if
they are collaborating with cross-functional peers.
