Improving Incident Retrospectives

This post was originally published on

man standing in front of waterfall with rainbow

Photo by Jared Erondu on Unsplash

As a Site Reliability Engineer (SRE) at Indeed, I often participate in the retrospective process that follows an incident. Retrospectives—in use at Indeed since late 2015—are a meaningful part of our engineering culture. I have never questioned their importance, but recently I was struck by shortcomings I saw in some retrospectives. For example:

  • A retrospective meeting might use only ~30% of the allotted time.
  • What is discussed might be gleaned from reading the incident ticket and retrospective document instead of attending the meeting.
  • Too much focus is devoted to the conditions that “triggered” the incident.
  • Signals used for deciding to hold a retrospective tend to direct focus toward incidents with high impact or high visibility.
  • Were participants actually learning anything new? It became apparent to me that we were not using every incident to realize our full potential to learn.

I decided to explore why so that we could improve our process.

The typical retrospective

Retrospectives at Indeed are usually a one-hour discussion including up to several dozen participants. The meeting is open to anyone in the company, but usually participants have either been involved in the incident response or have a stake in the outcome.

Facilitators follow a prescribed process:

  1. Review the timeline.
  2. Review the remediation items in the template.
  3. Find owners for the remediation items.
  4. Open the room for questions.

Spotting opportunities for improvement

In Summer 2018 I visited one of our tech sites and was invited to several local retrospective meetings to discuss some recent incidents. As an SRE it wasn’t unusual for me (or members of my team) to be invited. I also had subject matter expertise in a technology related to the incidents.

The facilitators took about 5 minutes to review the timeline, spent 8-10 minutes reviewing the remediation items, and concluded with questions related to the specific technologies involved in the causal chain. I didn’t learn anything new. I could have gained the same information from reading the incident ticket and retrospective document. This was a rare opportunity when a unique and eager group of people gathered in a conference room ready to collaboratively investigate. Instead, we never achieved the full potential.

This result is not uniform across retrospectives. I’ve been present in retrospectives where the participants offered such rich detail that the conversation continued well beyond the one-hour time limit, culminating with a huddle outside of the conference room.

The facilitators for these particular retrospective meetings followed the process faithfully but had only utilized ~30% of the time. It was clear to me that the retrospective process itself needed improvement.

Nurturing a safety culture

To understand potential changes, I first solicited viewpoints on why we conduct retrospectives at Indeed. Reasons I heard are likely familiar to most software organizations:

  • Find out what caused the outage
  • Measure the impact
  • Ensure that the outage never happens again
  • Create remediation items and assign owners

These goals also reflect Indeed’s strong sense of ownership. When someone’s service is involved in an incident, there’s a concern that we were closer to the edge of failure than we thought we were. Priorities temporarily change and people are more willing to critically examine process and design choices.

It’s important to use these opportunities to direct efforts toward a deeper analysis into our systems (both people and technical) and the assumptions that we’ve made about them. These approaches to a different safety culture at Indeed are still relatively new and are evolving toward widespread adoption.

Recommendation: Decouple remediation from the retrospective process

One process change I recommend is around the creation of remediation items. The retrospective process is not necessary as a forcing function for driving accountability of finding and owning remediation items.

I consistently observe that the creation of remediation items occurs organically after Production is stabilized. Many fixes are obvious to teams in the hindsight following an incident.

I see value in decoupling these “after action” activities from the retrospective process for many reasons.

  • The search for remediation items is often a tacit stopping point that halts further or deeper investigation.
  • The accountability around owning remediation items should be tightly coupled to incident ownership.
  • The retrospective process should be an optional activity. By making the retrospective process optional, teams that decide to engage in it are doing so because they see value in it rather than as an obligation or a checklist item.
  • Participants are freed up to conduct a deeper investigation unencumbered by the search for remediation items and shallow explanations.

Recommendation: Lighten up the retrospective template

Another useful change is with the retrospective template itself.

Using retrospective templates can be a lot like filling out forms. The focus is directed toward completion of an activity rather than free exposition. A blank document encourages a different kind of sharing. I have witnessed incidents where responders were so motivated to share their thoughts and descriptions that they produced rich and detailed analysis simply by starting with a blank document.

If every incident is shaped like a snowflake, it’s impossible to develop a template that is equipped to capture its unique characteristics. A template constrains detail and triggers explanations through close-ended questions. A blank canvas is open-ended. A template is yet another tacit stopping point that hinders deeper analysis. I recommend that we apply templates to incident analysis, but that we use blank documents for the retrospective process.

Driving organizational change

I have learned a lot by working to drive change at Indeed as we’ve grown quickly. My efforts have benefitted from my tenure in the company, experience participating in hundreds of incidents, and connection to the literature. I have made headway but there is still a lot to do.

I attribute some of my progress so far to finding other advocates in the company and remembering to communicate.

Find advocates

Advocates are colleagues who align closely with my goals, acknowledge where we could be doing better, and share a vision of what could be. I had no trouble finding these advocates. They are colleagues who are willing to listen, have an open mind and have the patience to consider another perspective. I held numerous 1:1s with leaders and stakeholders across the organization. I found opportunities to bring these topics up during meetings. I gave tech talks and reached out to potential advocates whenever I visited one of our global Engineering offices.


As much as I might think that I was communicating what I was working on, it was never enough. I found I had to constantly over-communicate. As I over-communicated and leveraged multiple media, I may have sounded repetitive to anyone in close proximity to my words. But this was the only way to reach the far edges of the organization who might not have otherwise heard me. Not everybody has time to read every email or internal blog post.

Looking ahead

Response to these changes has been largely positive. The focus during retrospectives is still anchored to the technological factors, when more attention could be paid to the human factors. I’m exploring different avenues for increasing the reach and effectiveness of these efforts. This includes working with our instructional design team to create a debrief facilitator program, communicating more often and more broadly, making more process changes, continuing to help teams produce and share high quality write-ups, and focusing on producing educational opportunities. At this point we’ve only scratched the surface and I’m looking forward to what we will accomplish.

About the author

Alex Elman is a founding member of the Site Reliability Engineering team at Indeed. He leads two teams: one that focuses on Resilience Engineering and one that supports the flagship Job Search product. For the past eight years Alex has been helping Indeed adopt reliability practices to cope with ever increasing complexity and scale. Follow Alex on Twitter @_pkill.

Cross-posted on Medium.

Tweet about this on TwitterShare on FacebookShare on LinkedInShare on RedditEmail this to someone

D-Curve: An Improved Method for Defining Non-Contractual Churn with Type I and Type II Errors

Businesses need to know when customers end their business relationships, an act called “churn.” In a subscription business model, a customer churns by actively canceling their contract. The company can therefore detect and record this churn with absolute certainty. But when no explicit contract exists, churn is more passive and difficult to detect. Without any direct feedback from the customer, companies cannot determine whether the customer has lapsed temporarily or permanently.

Until now, detecting churn in such non-contractual relationships has been mostly arbitrary and more art than science.

Various analysts deal with the non-contractual churn conundrum in different ways. One popular approach is to assume the customer has churned if they lapse for a sufficiently long consecutive period of time. A problem with this approach, apart from it being guesswork, is that the chosen threshold for the length of the lapse period is often too high. This causes the business to wait too long to identify any churn problems. In Prediction of Advertiser Churn for Google Adwords, the authors are only able to measure churn after 12 months! Such a long wait period reduces the value of churn detection and the business’s ability to address problems. In analyses that estimate the churn period as a specified percentile of a distribution of buy cycles—time between successive customer purchases—choosing an optimal percentile (90th, 95th, 99th, etc) is difficult.

In this blog post, we present an improved scientific approach for defining non-contractual churn. Our approach avoids the struggle of choosing an optimal percentile by minimizing a well-defined objective function of type I and II errors.


Churn period (d) is the minimum length of consecutive silent (no transaction) periods beyond which a customer is considered to have ended their business relationship. Companies commonly partition a book of business into active and churned customers. Where customer relationships are non-contractual, any specified d will have associated type I & II errors. Therefore we should choose a definition that minimizes an objective function of these errors. In our approach, we specify the function to be a weighted average of the errors.


  • e1(d) is the expected type I error associated with churn definition d; Type I error is labeling the customer as churned when they are active;
  • e2(d) is the expected type II error associated with churn definition d; Type II error is labeling the customer as active when they have churned;
  • w is the weight the analyst places on type I errors relative to type II errors; it can be interpreted as the relative costs of the errors.

The optimal churn definition, denote d*, therefore minimizes F(d). We call F(d) the d-curve.

To compute the error functions, e1(d) and e2(d), we need to introduce another set of notation:

  • ci represents the true churn status of customer i, 0=Active, 1=Churned;
  • li represents the number of consecutive periods customer i has lapsed.

With the above definitions, e1(d) and e2(d) are derived as follows.

From (2) and (3), we see that e1(d) is the overall proportion of active customers mislabeled as churned. Similarly, e2(d) is the overall proportion of churned customers mislabeled as active.

Implementing the theory

Suppose you have data that has recorded the periods associated with all customer transactions from time S to T.

To determine the optimal churn definition, complete the following experiment:

  1. Specify the minimum number of periods, D, beyond which you are almost sure that the customer has truly churned. You can do this by empirically examining distributions of customer buy cycles (the difference in periods between successive customer transaction dates) and choosing a sufficiently high percentile. We’ll call D the validation period. This means that the subjects of the experiment have to be limited to the subset of customers who have at least one transaction prior to T-D; else we cannot calculate the customer’s true churn status, ci. Also, the length of the entire data (T-S) should be long enough to allow you to evaluate the selected domain of churn definitions for the d-curve, F(d). For example, if the domain is {d:d<K+1), then T-S must exceed K+D.
  2. If you are only interested in voluntary churn, remove all customers otherwise terminated involuntarily by the company.
  3. For each customer i, determine the last purchase period as of time T:
    Calculate lapse period as of time T:
    And calculate the true churn status:
  4. For each customer i, calculate the last purchase period as of time T-D:Calculate lapse period as of time T-D:
  5. Select the domain of churn definitions, {d:dK}, on which you want to minimize F(d).
  6. For each churn definition in the selected domain, d =0, 1, 2…K, predict churn status for each customer as of time T-D, and measure the type I and II errors (e1(d) and e2(d)). Notice that e1(d) and e2(d) can be calculated from the data as follows:
  7. Select an appropriate weight, w.
  8. For d=0, 1, 2, …K, derive F(d) using (1).
  9. Choose the d that minimizes F(d) as your optimal d.

Results from real world application

We identified one of Indeed’s non-contractual products—job sponsorship—and applied both the percentile and d-curve methods to defining its churn period. We used monthly transaction data from September 2016 (S) through September 2019 (T).

Note that while the trends and insights we share are consistent with actual findings, we adjusted the actual results to protect the security of Indeed’s data.

Percentiles method

In this approach, we calculate the buy cycles for each customer. We can then represent each customer by a summary statistic (mean, median, and max) of their buy cycles. We then generate the distribution of the summary statistic across different customers:

Quantiles Mean Median Max
0 1 1 1
0.2 2 2 2
0.4 2 2 2
0.6 3.5 3 5
0.8 4.7 3 9
0.9 6.2 5 13
0.95 8 7 17
0.99 15 15 25
1 38 38 38
All figures illustrative


These results illustrate the analytic dilemma associated with the percentiles method. The distribution varies by the choice of summary statistic. Even with a given summary statistic, it’s not clear which percentile (90th, 95th or 99th) is optimal. Apart from that, any reasonable choice of percentile results in unnecessarily high churn definitions. For example, the 95th percentile of the distribution of mean buy-cycles is 8 months, while that of maximum buy-cycles is 17 months! And we will see in the next approach that while such longer definitions have lower type I errors, they have higher type II errors.

The d-curve approach deals with all of these problems by choosing the churn definition with the minimum weighted sum of the type I and II errors.

D-curve approach

We parameterized our model as follows:

  • w=0.5
  • D=12
  • S= 09-2016
  • K=12
  • T=09-2019
  • T-D=09-2018
Churn period Type I error (%) Type II error (%) Weighted error (%)
0 100.0 0.0 50.0
1 43.8 6.4 25.1
2 33.0 13.1 23.1
3 26.6 19.0 22.8
4 21.9 24.8 23.3
5 17.8 30.8 24.3
6 14.7 36.8 25.7
7 12.2 42.4 27.3
8 10.4 46.7 28.6
9 8.9 50.8 29.9
10 8.0 54.1 31.0
11 6.9 58.2 32.5
12 5.8 62.6 34.2

Using the d-curve, we choose 3 months as our optimal churn definition. A hypothesis test at 1% level of significance rejects the null hypothesis that the error for d=3 equals that of d=4.

More applications for the d-curve

We have formulated a framework for optimally selecting thresholds. While we apply the approach to define churn periods for non-contractual relationships, our approach has many other real world applications, chief of which is determining threshold probabilities in classification.


We are particularly grateful to Trey Causey, Ehsan Fakharizadi and Yaoyi Chen for their review and excellent feedback. We are, however, responsible for any mistakes in the post.

Cross-posted on Medium.

Tweet about this on TwitterShare on FacebookShare on LinkedInShare on RedditEmail this to someone

Unthrottled: How a Valid Fix Becomes a Regression

This post is the second in a two-part series.

In a previous post, I outlined how we recognized a major throttling issue involving CFS-Cgroup bandwidth control. To uncover the problem, we created a reproducer and used git bisect to identify the causal commit. But that commit appeared completely valid, which added even more complications. In this post, I’ll explain how we uncovered the root of the throttling problem and how we solved it.

Photo of highway at night

Photo by Jake Givens on Unsplash

Scheduling on multiple CPUs with many threads

While accurate, the conceptual model in my prior post fails to fully capture the kernel scheduler’s complexity. If you’re not familiar with the scheduling process, reading the kernel documentation might lead you to believe the kernel tracks the amount of time used. Instead, it tracks the amount of time still available. Here’s how that works.

The kernel scheduler uses a global quota bucket located in cfs_bandwidth->quota. It allocates slices of this quota to each core (cfs_rq->runtime_remaining) on an as-needed basis. This slice amount defaults to 5ms, but you can tune it via the kernel.sched_cfs_bandwidth_slice_us sysctl tunable.

If all threads in a cgroup stop being runnable on a particular CPU, such as blocking on IO, the kernel returns all but 1ms of this slack quota to the global bucket. The kernel leaves 1ms behind, because this decreases global bucket lock contention for many high performance computing applications. At the end of the period, the scheduler expires any remaining core-local time slice and refills the global quota bucket.

That’s at least how it has worked since commit 512ac999 and v4.18 of the kernel.

To clarify, here’s an example of a multi-threaded daemon with two worker threads, each pinned to their own core. The top graph shows the cgroup’s global quota over time. This starts with 20ms of quota, which correlates to .2 CPU. The middle graph shows the quota assigned to per-CPU queues, and the bottom graph shows when the workers were actually running on their CPU.

Multi-threaded daemon with two worker threads


Time Action
  • A request comes in for worker 1. 
  • A slice of quota is transferred from the global quota to the per-CPU queue for CPU 1.  
  • Worker 1 takes exactly 5ms to process and respond to the request.
  • A request comes in for worker 2. 
  • A slice of quota is transferred from the global quota to the per-CPU queue for CPU 2.

The chance that worker 1 takes precisely 5ms to respond to a request is incredibly unrealistic. What happens if the request requires some other amount of processing time? Multi-threaded daemon with two worker threads

Time Action
  • A request comes in for worker 1. 
  • Worker 1 needs only 1ms to process the request, leaving 4ms remaining on the per-CPU bucket for CPU 1.
  • Since there is time remaining on the per-CPU run queue, but there are no more runnable threads on CPU 1, a timer is set to return the slack quota back to the global bucket. This timer is set for 7ms after worker 1 stops running.
  • The slack timer set on CPU 1 triggers and returns all but 1 ms of quota back to the global quota pool.  
  • This leaves 1 ms of quota on CPU 1.
  • Worker 2 receives a long request. 
  • All the remaining time is transferred from the global bucket to CPU 2’s per-CPU bucket, and worker 2 uses all of it.
  • Worker 2 on CPU 2 is now throttled without completing the request.  
  • This occurs in spite of the fact that CPU 1 still has 1ms of quota.

While 1ms might not have much impact on a two-core machine, those milliseconds add up on high-core count machines. If we hit this behavior on an 88 core (n) machine, we could potentially strand 87 (n-1) milliseconds per period. That’s 87ms or 870 millicores or .87 CPU that could potentially be unusable. That’s how we hit low-quota usage with excessive throttling. Aha!

Back when 8- and 10-core machines were considered huge, this issue went largely unnoticed. Now that core count is all the rage, this problem has become much more apparent. This is why we noticed an increase in throttling for the same application when run on higher core count machines.

Note: If an application only has 100ms of quota (1 CPU), and the kernel uses 5ms slices, the application can only use 20 cores before running out of quota (100 ms / 5 ms slice = 20 slices). Any threads scheduled on the other 68 cores in an 88-core behemoth are then throttled and must wait for slack time to be returned to the global bucket before running.

Resolving a long-standing bug

How is it, then, that a patch that fixed a clock-skew throttling problem resulted in all this other throttling? In part one of this series, we identified 512ac999 as the causal commit. When I returned to the patch and picked it apart, I noticed this.

-       if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
+       if (cfs_rq->expires_seq == cfs_b->expires_seq) {
               /* extend local deadline, drift is bounded above by 2 ticks */
                cfs_rq->runtime_expires += TICK_NSEC;
       } else {
                /* global deadline is ahead, expiration has passed */
                cfs_rq->runtime_remaining = 0;

The pre-patch code expired runtime if and only if the per-CPU expire time matched the global expire time (cfs_rq->runtime_expires != cfs_b->runtime_expires). By instrumenting the kernel, I proved that this condition was almost never the case on my nodes. Therefore, those 1 milliseconds never expired. The patch changed this logic from being clock time based to a period sequence count, resolving a long-standing bug in the kernel.

The original intention of that code was to expire any remaining CPU-local time at the end of the period. Commit 512ac999 actually fixed this so the quota properly expired. This results in quota being strictly limited for each period.

When CFS-Cgroup bandwidth control was initially created, time-sharing on supercomputers was one of the key features. This strict enforcement worked well for those CPU-bound applications since they used all their quota in each period anyway, and none of it ever expired. For Java web applications with tons of tiny worker threads, this meant tons of quota expiring each period, 1ms at a time.

The solution

Once we knew what was going on, we needed to fix the issue. We approached the problem in several different ways.

First, we tried implementing “rollover minutes” that banked expiring quota and made it usable in the next period. This created a thundering herd problem on the global bucket lock at the period boundary. Then, we tried to make quota expiration configurable separate from the period. This led to other issues where bursty applications could consume way more quota in some periods. We also tried returning all the slack quota when threads became unable to run, but this led to a ton of lock contention and some performance issues. Ben Segall, the author of the CFS scheduler, suggested tracking the core-local slack and reclaiming it only when needed. This solution had performance issues of its own on high-core count machines.

As it turns out, the solution was actually staring us right in the face the whole time. No one had noticed any issues with CFS CPU bandwidth constraints since 2014. Then, the expiration bug was fixed in commit 512ac999, and lots of people started reporting the throttling problem.

So, why not remove the expiration logic altogether? That’s the solution we ended up pushing back into the mainline kernel. Now, instead of being strictly limited to a quota amount of time per period, we still strictly enforce average CPU usage over a longer time window. Additionally, the amount that an application can burst is limited to 1ms for each CPU queue. You can read the whole conversation and see the five subsequent patch revisions on the Linux kernel mailing list archives.

These changes are now a part of the 5.4+ mainline kernels. They have been backported onto many available kernels:

  • Linux-stable: 4.14.154+, 4.19.84+, 5.3.9+
  • Ubuntu: 4.15.0-67+, 5.3.0-24+
  • Redhat Enterprise Linux:
    • RHEL 7: 3.10.0-1062.8.1.el7+
    • RHEL 8: 4.18.0-147.2.1.el8_1+
  • CoreOS: v4.19.84+

The results

In the best-case scenario, this fix enables a .87 increase in usable CPU for each instance of our affected applications, or a corresponding decrease in required CPU quota. These benefits will unlock increased application density and decreased application response times across our clusters.

Decrease in required CPU load

How to mitigate the issue

Here’s what you can do to prevent CFS-Cgroup bandwidth control from creating a throttling issue on your systems:

  • Monitor your throttled percentage
  • Upgrade your kernels
  • If you are using Kubernetes, use whole CPU quotas, as this decreases the number of schedulable CPUs available to the cgroup
  • Increase quota where necessary

Ongoing scheduler developments

Konstantin Khlebnikov of Yandex proposed patches to the Linux kernel mailing list to create a “burst bank.” These changes are feasible now that we have removed the expiration logic, as described above. These bursting patches could enable even tighter packing of applications with small quota limits. If you find this idea interesting, join us on the Linux kernel mailing list and show your support.

To read more about kernel scheduler bugs in Kubernetes, see these interesting GitHub issues:

Please also feel free to tweet your questions to me @dchiluk.

Cross-posted on Medium.

Tweet about this on TwitterShare on FacebookShare on LinkedInShare on RedditEmail this to someone