Key stats
- 4 critical tasks: deploy, rollback, schema change, and incident recovery
- 3 coverage counts: two or more performers, one performer, and none
- 2 rotations per quarter with a named observer
- 1 review of accepted concentration per service area
Key takeaways
- Define concentration by demonstrated task performance, not by team membership.
- Record partial capability separately so read-only coverage is not mistaken for full coverage.
- Use the review history as a knowledge artifact and require the reason in each change.
- Rotate the tasks where one absence would block a release or an incident response.
- Review accepted concentration with an owner and a mitigation rather than pretending it is gone.
Question and scope
This report examines knowledge concentration inside an offshore maintenance pod and the risk it creates for the buyer. The unit is one service area and the number of engineers who can change it safely without help. The scope covers application maintenance, release support, and incident response handled by a dedicated pod. It excludes new product development, where some concentration is expected while a feature is still being designed.
The question is whether a pod can show concentration risk before a departure or a leave period exposes it. A team can look fully staffed while one engineer holds the only working knowledge of a payment path. The report treats demonstrated capability as the evidence, not the number of people listed on a team chart. It looks for a review routine that makes that evidence visible.
Define knowledge concentration
Concentration exists when the safe completion of a task depends on one person's memory or access. The dependency can be technical, such as knowing which migration is safe to rerun, or organizational, such as knowing who can approve a database change. Both forms create a queue when that person is unavailable. The risk is not that the person is careless but that the system has no second path.
A useful definition also includes partial knowledge. An engineer may be able to read a service but not deploy it, or deploy it but not recover it. Record the capability separately rather than marking the service as covered. Partial coverage often looks sufficient until the task reaches the step that nobody else has performed. Record the highest step each engineer has performed so the gap is visible before an incident.
The artifacts that carry knowledge
Knowledge lives in runbooks, tests, deployment scripts, dashboards, and the review history. Each artifact can carry context only if someone can use it without asking the author. A runbook that references a personal script is not an artifact; it is a pointer to a person. Review the artifacts by asking a second engineer to complete a task using only what is written.
The review history is often the richest source and the least maintained. A pull request that explains why a retry was added preserves a decision that a runbook would miss. Make the explanation part of the change rather than a separate document that drifts. When the author leaves, the review thread remains as the closest thing to their reasoning. Ask authors to write the reason in the change description, not only in the ticket that will be closed.
Proxy measures for bus factor
Count the engineers who have completed each critical task in the last two quarters. Use deploy, rollback, schema change, and incident recovery as the task list. A count of one is a concentration finding, even when several people have read the code. The measure is a proxy, because it records demonstrated action rather than understanding. Use the deployment and incident records rather than asking people to estimate their own coverage.
Add the number of critical tasks that no one completed in the period. That number identifies areas that are either stable or untested, and the team should know which. Pair the count with the review coverage, so a service that nobody touched but several people reviewed is not treated as a total unknown. The measures together are more honest than a single bus factor number.
Rotation and pairing
Rotation should follow the critical task list, not a general desire to share work. Assign a second engineer to each high-risk task for one cycle and have that engineer perform the task while the primary observes. Record the result and the questions that came up. A pairing session that never reaches the actual command is a shadowing exercise, not a transfer.
Keep the rotation small enough to be real. Moving every engineer through every service produces shallow coverage and disrupts delivery. Choose the tasks where a single absence would stop a release or an incident response. Rotate those first, then revisit the list at the next quarter. Choose tasks by the cost of an absence, not by how interesting the service is to learn.
What to measure
Report the number of critical tasks with at least two recent performers, the number with one, and the number with none. Add the time to recover when the primary is unavailable, measured in a drill or a real event. That time is the practical consequence of concentration and is easier for a buyer to interpret than a coverage percentage. Keep the raw task list beside the summary.
Track whether the same engineer appears as the only performer across several tasks. A pattern of concentration around one person is a continuity risk that a single service review will miss. Record the concentration by person as well as by service, and review it with the pod lead. The measure should prompt a rotation decision rather than a performance comment.
A quarterly review protocol
Hold the review once a quarter with the pod lead, the buyer-side service owner, and one engineer who is not the primary. Walk through the critical task list and the evidence of who performed each one. Ask the non-primary engineer to explain the riskiest task without prompting from the primary. Record where the explanation stops. Keep the primary engineer present but let the second engineer answer first, so the gap is visible.
Choose two tasks for the next quarter and schedule the rotation. Assign a date, an observer, and a written output such as an updated runbook section. Review the result at the next meeting before adding more tasks. A quarterly plan with two completed transfers is more credible than a long list of intended rotations. Make the output concrete, such as a runbook change or a successful rollback in a safe environment.
Limits of the signal
A task count cannot show how well someone understands a system. An engineer may complete a deploy by following a script while missing the reason behind a step. Use the review conversation and the incident record to test understanding, not only participation. The count is a prompt for that conversation, not a replacement for it. Ask why a step exists and what would break if it were skipped, then record the answer.
Concentration is also rational in small teams. A pod of three cannot have two experts in every area without slowing all work. The goal is to protect the tasks where an absence blocks the buyer, not to eliminate every dependency. Record the accepted concentration and the mitigation, such as a documented vendor contact or a tested break-glass path. Name the mitigation for each accepted area and review it when the pod composition changes.
From coverage to continuity evidence
Continuity is proven by a drill, not by a coverage chart. Ask the second engineer to perform the critical task while the primary observes and takes notes. Record the time, the questions, and the step where help was needed. The drill result is stronger evidence than a list of names, because it shows what the team can actually do. A successful drill with a written note is the smallest unit of evidence.
Schedule the drill before a known absence, such as a holiday or a planned leave, so the result has a real deadline. Repeat the task after the notes are updated to confirm the improvement. Keep the drill in normal working hours rather than treating it as extra work. A continuity plan that depends on unpaid overtime will not survive a busy quarter.
Reporting the result
Report the coverage of critical tasks, the two rotations completed, and the remaining single-person areas. Name the risk owner for each remaining area and the date of the next review. Keep the task-level detail in an appendix for the engineers who will perform the work. The buyer should see where an absence would hurt and what the team is doing about it.
Review the report beside the continuity section of the maintenance agreement. If the agreement promises a backup for a role, the task evidence should show that the backup has performed the work. A promise without a drill is only an intention. The report is the place to compare the two and record the difference. A named backup who has never performed the task is a plan, not evidence of continuity.
Evidence table
| Signal | What to inspect | Owner |
|---|---|---|
| Outcome | Acceptance evidence for the bounded task | Task reviewer |
| Control | Access, test, and approval boundary | Internal owner |
| Handoff | Open risks and next decision | Next owner |
Good distributed work is observable at the handoff: the result, evidence, limitations, and next owner are explicit.
Frequently asked questions
Is a bus factor of two enough?
It depends on the task. Two performers are the minimum for a critical task, but the second person should have completed it recently and without help. A name on a list is not coverage.
How do we transfer knowledge without slowing delivery?
Choose two high-risk tasks per quarter and make the transfer part of real work. A focused rotation is less disruptive than moving everyone through every service.
What if the buyer owns the only expert?
Record it as a cross-boundary risk and name a mitigation, such as a documented break-glass path or a support contract. The pod cannot rotate a role it does not control.
