13. Site Reliability Engineering (SRE)¶
In the VisIt project, members of the development team are frequently called upon to respond to a variety of inquiries often originating directly from users. Some of these may relate to the use of VisIt such as
- How do I do a surface plot?
- How do I compute the mass in a given material?
- How do I get client/server to TACC working?
and some may relate to an operational aspect of either the VisIt software itself such as
- A botched managed VisIt installation.
- An update to host profiles to address site access changes.
- A missing database reader plugin.
or, the underlying computing infrastructure upon which VisIt depends such as
- An incompatible graphics driver.
- A downed file system or network.
- A trip in the security environment.
Typically, such inquiries originate from users in the midst of using VisIt and are encountering some kind of difficulty. In highly effective software projects, the work involved in handling such inquiries does not end with fixing this one user’s problem and sending them on their way. When one user encounters a problem, there are probably others who have encountered the same problem. Furthermore, often the problems users encounter are suggestive of minor, easily fixed deficiencies in either the software itself or its associated processes and artifacts.
The continuous investment of effort to craft and carry out small corrective actions in response to such inquiries is a best practice. It represents a fusion of aspects of Google’s Site Reliability Engineering (SRE) process (sometimes also called Systems Reliability Engineering or Services Reliability Engineering) and aspects of User Experience Regression Testing and/or User Experience Driven Development (UXDD).
For mature DOE software projects with wide reach and many users, SRE activity represents a brand of effort wholly different from conventional software product development, planning and execution. Like most DOE software projects, VisIt has no dedicated SRE resources. Instead, developers themselves must also support SRE work. Nonetheless, managing SRE work effectively and efficiently is an essential part of maintaining the overall productivity and sustainability of the software as well as the productivity of both users and developers of the software alike.
This document describes how the VisIt project manages its SRE activities. Some of the goals of this process are…
- To maintain a reputation for timely and quality response to customer inquiries.
- To develop a practice of routine housekeeping quality improvements to the VisIt software and associated processes and artifacts impacting user and/or developer productivity.
- To load balance SRE work in an equitable way across the development team.
- To reduce SRE interruptions for the team as a whole.
- To log, track and evolve a database of SRE activity and effort to help inform ongoing development plans and resource allocation.
- To identify and document escalation paths for major incidents.
- To aim for a four hour response time.
While many aspects of SRE are under the direct control of VisIt developers, some are not and involve collaboration with other teams in resolving. In most cases the extent of the VisIt team’s involvement in the operations is confined primarly to the VisIt software itself; its development, testing, release and deployment which includes installations the VisIt team directly manages, hosted binary downloads for common platforms and the tools and resources to build from sources. Operational issues impacting VisIt but outside of this scope are typically delegated to other teams who are responsible for the associated processes and resources.
13.1.1. Business Hours¶
In the IT world where companies like Google, Apple and Amazon have whole teams dedicated to SRE activity, coverage is 24/7 and response time is measured in minutes. For the VisIt project where the majority of funded development takes place at Lawrence Livermore National Lab, coverage is during normal West Coast business hours, 8am-12pm and 1-5pm (GMT-8, San Francisco time zone), Monday through Friday excluding LLNL holidays and response time may be as much as four hours due to team members having to multi-task among many responsibilities.
13.1.2. Developer Away Notifications¶
As an aside but nonetheless related to the SRE response time goal of four hours, developers agree to notify the team if, during normal business hours, they will unexpectedly be away from their desk for periods longer than four hours. This is true whether working on or off site. Being away for periods shorter than four hours does not require such notification.
13.2. The Basic Process¶
SRE work is allocated and rotated among developers in one-week shifts. During a shift, one developer’s role is to serve as the Primary SRE contact and a second developer’s role is to serve as a Backup. Except for escalations, all other developers are free of SRE responsibilities for that week.
The role of the Primary is to respond within the response time goal, to each inquiry. Ideally, all SRE activity during the week is handled and resolved solely by the Primary. However, escalations, which we hope are rare, will wind up engaging the Backup and may even engage other developers. In addition, any active SRE dicussions that remain unresolved at the end of the week are formally handed off to the next Primary.
Active SRE discussions will be logged and tracked in a separate GitHub, issues-only repository within the visit-dav GitHub organization. Upon resolution of serious incidents, the Primary will prepare a brief postmortem to inform a discussion at the next project meeting of possible changes in practices to avoid repeating such major incidents.
Because SRE work tends to be interrupt driven, there is always the chance that the Primary will have no active discussions. At these idle times, the Primary shall use their time to address general housekeeping or other low-hanging fruit type work. In particular, there shall be no expectation that a developer serving as Primary can get any other work done beyond their active or idle SRE obligations. In slow weeks, its conceivable they can. But, there can be no implied assumption or expectation that this will be the case.
A schedule of the Primary and Backup assignments going out several months is periodically negotiated by the team and posted in the form of a shared calendar. Primary and Backup responsibilities are rotated so as to balance the load among team members.
The preceding paragraphs describe VisIt’s SRE processes at a basic level and in the ideal. Nonetheless, several terms here (those that are links or in italics in the paragraphs above) require elaboration. In addition, there are also many practical matters which can serve to complicate the basic process. These details are addressed in the remaining sections.
The Primary’s role is to respond, within the response time goal, to each inquiry that occurs during that week including those that come in during the preceding weekend/holiday. The Primary’s goal is to resolve all inquiries by the end of their week.
The Primary has the sole responsibility for responding to inquiries and opening and resolving SRE issue tickets. When the Primary needs help to resolve an SRE issue, s/he should first enlist the Backup. This is an escalation. Nonetheless, the Backup (or other developers for that matter) are called into action only by explicit request of the Primary. Note that enlisting additional resources for help is part of escalation and is not the same as a handoff.
If the Primary’s schedule changes such that the response time goal may not be met, the Primary may temporarily delegate his/her role and responsibilities to the Backup. To the extent possible, such temporary delegation from Primary to Backup should be handled formally and by mutual agreement. Temporary delegation of the Primary’s role is also not the same as a handoff.
Ideally, the Primary is able to handle all SRE activity and no other developers are engaged. Thus, other developers are free to ignore customer inquiries as well as redirect customers who may contact them directly via email, phone or walk-in. It is a best practice to handle such redirections with a formal, three-way handoff confirming that the customer indeed makes contact with the Primary.
13.4. SRE vs. Product Development¶
Part of the reason for formalizing this process is the recognition of a different category of work, Site Reliability Engineering (SRE), that is essential part of maintaining the overall quality of a software product as well as the productivity of both developers and users of the software alike. Nonetheless, SRE work is very different from conventional product development type work where bug fixes, technology refreshes and feature enhancements are estimated and prioritized, methodically planned and resources are assigned to hit target release dates.
Issues that impact one user’s productivity often impact others. Likewise for developers. When such issues come to our attention, whenever possible it is often helpful to identify two kinds of actions; a short-term constructive correction and a longer-term comprehensive solution.
|Constructive Correction||Comprehensive Solution|
|Short term||Longer term|
|Faster response||Slower response|
|Low cost/benefit||Higher cost/benefit|
|Low risk||Higher risk|
A constructive correction has value only when it represents a step towards the comprehensive solution, can sufficiently reduce the impact of the issue and can be rolled out to users significantly sooner and with lower cost than the comprehensive solution. Ordinarily, a constructive correction is something the Primary handles as part of their SRE activity. The comprehensive solution, which often involves more planning and resource allocation, is handled as part of normal product development activities.
Constructive corrections can wind up falling through the cracks of traditional software project management and planning processes. However, such work also often represents low cost high benefit improvements in quality of either the software itself or the development or deployment processes supporting it. We refer to issues of this nature as general low-hanging fruit type issues.
Apart from acknowledging their existence, a key part of this process is the allocation of a small fraction of our resources for the sole purpose of supporting SRE activities and developing a practice of continuously crafting constructive corrective actions arising from SRE inquiries.
Consequently, another key role of the Primary is to use any time not working active SRE issues to fix other low-hanging fruit issues from the product development backlog. As a rule of thumb, low-hanging fruit is considered to be anything that the team believes is fixable within a half-day’s (4 hours) worth of effort. When there are many such tasks in the system to work on, the Primary is free to use his/her judgment to decide which s/he can most productively address.
Part of the acknowledgment of this new category of work is the new issue tracker for tracking it. New SRE activity will start with an issue being added there. As an SRE incident unfolds it may result in either the same issue being moved to the product development issue tracker and/or new issue(s) being added to the product development tracker. Any new product development issues should be linked back to the original SRE issue that spawned them.
13.5. Active SRE Discussions¶
Active SRE issues will be logged and tracked as discussions in our GitHub Discussions within the visit-dav GitHub organization. For each new inquiry, a discussion will be created either by the primary or by the customer who started the discussion.
The primary will endeavor to capture all relevant information and communications in this issue. The use of GitHub Discussions for this purpose has a number of advantages over other options such as email including better search/browse as well as support for attachments. For the remainder of this document we simply use the term conversation to refer to the communication involved in an active SRE issue.
Upon receiving a new inquiry, if the inquiry did not start as a discussion (e.g. maybe it was a telephone hotline call, or a walk-in, the procedure is for the Primary to include the initial information (with the exception of inquries involving classified information) in a new GitHub discussion, attach the SRE label and from then on handle all communication through the conversation associated with that discussion.
For any work the Primary performs, even if it is a rather trivial amount of work to resolve, there should be an associated discussion for tracking that work. Tracking even the trivial tasks will help to build a database of activity we may be able to later mine to identify patterns and further process improvements.
An SRE discussion is answered when the associated inquiry is
resolved. Or, it is answered and labeled
wontfix if 21 days pass since the user last engaged in any conversation with
VisIt developers to reach a resolution. Because of the manner in which the
interface to GitHub discussions behaves, the natural chain of communication may
not always lend itself well to using GitHub’s check this reply as the answer
feature to indicate an SRE discussion is resolved. In such circumstances, the
adopted approach is to add a new, top-level comment with a link to whichever
previous, embedded comment best answers the issue and then tag that new, top-level
comment as the discussion’s answer.
13.6. Supported Methods of Contact¶
An SRE inquiry with the VisIt team begins with a first contact and may optionally be followed by ongoing conversation. These two kinds of communication have different requirements and can involve different processes. This is due to the fact that we need to balance two priorities; accessibility for users and productivity for developers.
To maximize accessibility for users, we should support a wide variety of methods of first contact. However, to maximize productivity for developers, we should restrict methods of ongoing conversations.
A key benefit of having the VisIt team co-located with our user community is that users can spontaneously make a first contact with any one of us by an office drop-in or a tackle in the hallway or parking lot. This can even occur on social media platforms such as Confluence, Jabber, MS Teams, etc. where users can wind up engaging specific VisIt developers that happen, by nothing more than coincidence, to also be using those platforms.
A challenge with these spontaneous methods of first contact is that they inadvertently single out a specific developer who is then expected to at least respond and possibly even to also resolve the issue. But, these actions and the effort they involve are the responsibility of the primary SRE. Consequently, spontaneous methods of first contact can wind up jeopardizing the goals of our SRE process by making it difficult to track, allocate and manage SRE effort.
Therefore, the methods of first contact we officially support are those which engage the whole team instead of singling out a specific member. This includes…
Whenever users attempt a first contact through something other than the supported methods listed immediately above, the receiving developer should make an effort to handoff the inquiry to the primary SRE as quickly and politely as practical.
What does it mean for a method of first contact to be supported? It means there is an assurance that the particular platform is being monitored by VisIt team members during normal business hours such that the response time goal can be maintained. In addition, supported methods are encouraged and promoted in any documentation where VisIt support processes are discussed.
Balancing the priorites of user accessibility with developer productivity involves a compromise on the number of platforms we make an assurance to monitor. Currently, this is limited to those listed above. However, the selected methods should be periodically reevaluated. If there is some platform which seems to be gaining popularity among users, it could either be added to the list of supported platforms or perhaps it could be integrated with email in the same way GitHub issue conversations have been.
13.7. Response Time and Response vs. Resolution¶
The response time goal of four hours was chosen to reflect the worst case practicalities of team members’ schedules and responsibilities. For example, if the Primary has meetings just before and just after the lunch hour break, there can easily be a four hour period of time where inquiries go unattended. Typically, we anticipate response times to be far less than four hours and certainly, when able, the Primary should respond as quickly as practical and not use the four hour goal as an excuse to delay a prompt response.
Since a majority of funding for VisIt is from LLNL and since VisIt developers are co-located with many of its LLNL users, certainly these users as well as their direct collaborators are accustomed to response times of less than four hours. For example, the VisIt project operates a telephone hotline and also frequently handles walk-ins. As an aside, after a recent small test effort to maintain a rapid response time, a noticeable up-tick in user email inquiries was observed suggesting that rapid response times have the effect of encouraging more user interactions.
It is also important to distinguish between response and resolution here. A key goal in this process is to ensure that customer inquiries do not go unanswered for a long time. However, responding to a customer inquiry does not necessarily mean resolving it. Sometimes, the only response possible is to acknowledge the customer’s inquiry and let them know that the resources to address it will be allocated as soon as practical. In many cases, an immediate response to acknowledge even just the receipt of a customer’s inquiry with no progress towards actual resolution goes a long way towards creating the goodwill necessary to negotiate a day or more of time to respond more fully and maybe even resolve.
Resolution of an SRE issue often involves one or more of the following activities…
- Answering a question or referring a user to documentation.
- Diagnosing the issue.
- Developing a work-around for users.
- Developing a reproducer for developers.
- This may include any relevant user data files as well as approval, where appropriate for world read access to such data as part of attaching to a GitHub issue.
- Identifying any low-hanging fruit type work that would address, even if only in part, the original SRE inquiry and then engaging in the housekeeping work to resolve it.
- Determining if the user’s issue is known (e.g. an issue ticket already exists).
- Updating a known issue with new information from this user, perhaps adjusting labels on the issue or putting the issue back into the un-reviewed state for further discussion at a VisIt project meeting.
- Identifying and filing a new product development type issue ticket.
To emphasize the last bullet, resolution does not always mean a customer’s issue can be addressed to satisfaction within the constraints of the SRE process as it is defined here. Sometimes, the most that can be achieved is filing a highly informative issue ticket to be prioritized, scheduled and ultimately resolved as part of normal VisIt product development activities. The SRE issue gets promoted to a product development issue. It is closed in the SRE issue tracker and new issue is opened in the product development issue tracker including a reference to the original SRE issue. Doing so does serve to resolve the original SRE issue that initiated the work.
13.8. Serious Incidents and Postmortems¶
Serious incidents are those that have significant productivity consequences for multiple users and/or require an inordinate amount of resources (either time or people or both) to diagnose, work-around and/or ultimately properly correct.
When such incidents occur, it is a best practice to spend some time considering adjustments in processes that can help to avoid repeating similar issues in the future.
When such incidents reach SRE resolution, the Primary will prepare a brief postmortem (often just a set of bullet points) explaining what happened and why, estimating the amount of resources that were needed to resolve the incident, describing key milestones in the work to resolve the incident and suggesting recommendations for changes in processes to prevent such incidents from being repeated. This postmortem will be used to guide team discussion during a subsequent weekly project meeting.
Our SRE processes involve two kinds of handoffs. One is the redirection of a customer who makes contact with a developer not serving as the Primary. The other is the handoff of unresolved SRE issues from one week’s Primary to the next.
To handle customer redirection handoffs, it is a best practice to use a three-way handoff giving the customer some assurance that their initial contact with someone is successfully handed off to the Primary. For example, for a call-in, it is a best practice to try a three-way call transfer. For some developers, the prospect of redirecting friends and colleagues with whom they may have long standing relationships may be initially uncomfortable. But it is important to recognize that this an essential part of achieving one the goals of this process, to reduce SRE interruptions for the team as a whole.
If an active SRE issue cannot be resolved within the week of a Primary’s assignment, it gets handed off to the next week’s Primary. Such handoffs shall be managed formally with a comment (or email) to the customer(s) and the next week’s Primary and Backup in the associated GitHub issue. The associated issue(s) in the SRE issues repository shall be re-assigned by the previous week’s Primary upon ending their shift. However, a preceding week’s Primary may be near enough to resolving an SRE issue that it makes more sense for him/her to carry it completion in the following week. In this case, s/he will leave such issues assigned to themselves.
SRE inquiries may escalate for a variety of reasons. The
technical expertise or authority required may be beyond the Primary’s
abilities or other difficulties may arise. For issues that the Primary does
not quickly see a path to resolution, the Backup should be enlisted first.
When developer expertise other than Backup is needed, the Primary should
try to engage other developers using the
@ mention feature in the associated
GitHub issue. However, where a Primary is responsible for maintaining the
response time goal, other developers so enlisted are free to either delay or even
decline to respond (but nonetheless inform the Primary of this need) if their
schedule does not permit timely response. Such a situation could mean that the
only remaining course of action for the Primary to resolve the issue is to
file a product development issue as discussed at the end of a preceding section.
If after investigation and diagnosis the work required to resolve an SRE incident remains highly uncertain or is not believed to be a low-hanging-fruit type task, the Primary should search the product development issues to see if this is a known issue and, if so, add additional information to that known issue about this new SRE incident (and perhaps remove the reviewed tag from the issue to cause the issue to be re-reviewed at the next VisIt project meeting) or submit a new issue to the product development issue tracker. Such action then resolves the original SRE issue.
13.10.1. Special Considerations for Classified Computing¶
Occasionally, incidents arise that may be specific to a classified computing environment. This is not too common but does happen and it presents problems for a geographically distributed team. In many ways, handling such an incident is just a different form of escalation.
On the one hand, customers working in a classified computing environment are accustomed to longer response times. On the other hand, such work is often a high priority and requires rapid response from a developer that is on site with classified computing access.
Our current plan is to handle this on a case-by-case basis. If neither the Primary nor Backup are able to handle a customer response incident requiring classified computing, the Primary should
- First determine the customer’s required response time. It may be hours or it may be days. If it is days. Its conceivable the issue could be handled in the following week by a new Primary/Backup pair.
- If customer indicates immediate response is required, the Primary should query the whole team to arrange another developer who can handle it.
13.11. Scheduling and Load Balancing¶
To balance the work load of SRE, the responsibilities of the Primary and Backup are rotated, round-robin among team members. For example, on a team of eight developers, each would serve as Primary only one week in eight or 12.5% of their time. However, a number of factors complicate this simple approach including percent-time assignments of team members, alternate work schedules, working remotely, travel, vacations, trainings, meetings, etc.
Round-robin assignment may lead to a fair load by head-count but isn’t weighted by percent-time assignments. From a percent-time assignment perspective, it might be more appropriate for a developer that is only 50% time on VisIt to serve as the Primary only half as often as a 100% time developer.
Since a majority of VisIt developers divide their time across multiple projects, we use 50% as the nominal developer assignment. Because of all the factors that can effect scheduling, the VisIt project has opted to manage scheduling by periodically negotiating assignments 1-3 months into the future and recording the assignments on a shared calendar. The aim is an approximately round-robin load balancing where contributors who are more than 50% time on VisIt are occasionally assigned an extra week. Either Primary or Backup can make last minute changes to the schedule by finding a willing replacement, updating the shared calendar and informing the rest of the team of the change.
Whenever possible, an experienced Backup will be scheduled with a less experienced Primary.
13.12. A Common Misconception: SRE is an Interruption to Programmatic Work¶
When faced with a long backlog of development tasks, team members can all too easily perceive SRE work as an interruption to those tasks. This is a common misconception. SRE is an important aspect to a successful product and project on par with any other major development work. It is part of what is involved in keeping the software working and a useful tool in our customer’s workflows not only here at LLNL, likely VisIt’s biggest customer, but wherever in DOE/DOD and elsewhere in the world VisIt is used.
Indeed, there are several advantages in having developers involved with SRE activities. These include..
- Learning what problems users are using the tool to solve.
- Learning how users use the tool.
- Learning what users find easy and what users find hard about the tool.
- Learning where documentation needs improvement.
- Learning where the user interface needs improvement.
- Learning operational aspects of user’s work that the tool can impact.
- Building collaborative relationships with other members of the organization.
- Learning how users operate in performing their programmatic work for the organization which helps to inform planning for future needs.
In short, the work involved in Software Reliability Engineering (SRE) and ensuring productivity of both users and developers of VisIt is programmatic work. The practice of having software development staff integrated with operations is more commonly referred to as DevOps. There is a pretty good video that introduces these concepts.