Card 1 of 50
Front. Slo-user. What should an SLO measure?
These cards are original study material. They are not questions from a named employer or a live hiring loop.
All 50 cards in this deck
1.Slo-user
What should an SLO measure?
The user action that matters, such as a completed purchase. A homepage that returns 200 can stay green while that action is down.
2.Sli
What is an SLI?
The measurement behind an SLO, such as the share of checkout requests that succeed in under a second.
3.Error-budget
What is an error budget?
The unreliability the SLO allows over the window. Spending it is the signal to slow launches and fix reliability.
4.Burn
What is burn rate?
How fast the error budget is being consumed. A high burn pages now. A slow burn can wait for the workday.
5.Multi-burn
Why page on more than one burn window?
A short window catches a sharp outage. A long window catches a slow leak that never spikes hard enough to page.
6.Retry-sync
What does synchronized retry do?
Every client waits the same delay and calls again together. The recovery traffic arrives as one wave.
7.Backoff
What does a backoff with jitter do?
Each retry waits longer, plus a random spread, so clients do not line up. The dependency gets a chance to recover.
8.Budget-retry
When should a client stop retrying?
When the deadline the user is waiting on has passed, or a budget of attempts is spent. Retrying past that only adds load.
9.Split-brain
What is a split brain on a last seat?
Two regions both accept the write because they cannot see each other. Without a conflict rule, both succeed and the seat is sold twice.
10.Nat
Why is one NAT a reliability problem?
Every zone's traffic shares it. When it fails, zones that are otherwise healthy fail together.
11.Dependency
What is a hard dependency?
The request cannot succeed without it. A soft dependency can fail and the request still returns something useful.
12.Circuit
What should a circuit breaker do after it opens?
Allow a trial call later. A breaker that never tries again stays open after the dependency has recovered.
13.Half-open
What is the half-open state?
A few trial requests are let through. Success closes the circuit. Failure opens it again.
14.Stall
CPU is idle and reads look fine, but writes stall. Where do you look?
At the disk or the log the commit waits on. A saturated commit disk holds transactions while the app looks bored.
15.Low-traffic
Why is an error ratio misleading when traffic stops?
The few remaining health checks succeed, so the ratio looks fine, while real users have given up calling.
16.Absolute
What do you alert on besides a ratio?
The count of failures, or the absence of expected successes. A ratio hides an outage that also removes the traffic.
17.Stale-dash
What if the metrics pipeline freezes on green?
The dashboard is no longer a signal. A check that the metrics themselves are fresh has to sit beside the SLO.
18.Symptom
What is a symptom-based page?
A page that the user-facing SLI is burning, not that a CPU crossed a line. CPU is a clue after someone is already looking.
19.Cause-page
Why not page on every cause?
Causes are many and often harmless. Paging on them trains people to ignore pages. Page on the user symptom, then debug the cause.
20.Saturation
What is saturation?
How full a resource is: connections, disk, queue depth. It predicts the next latency climb, before the errors start.
21.Queue-delay
What does queue time tell you?
Requests are waiting for a worker. The work itself may be fast. Adding workers, or shedding load, is the lever, not tuning the handler.
22.Timeout
Where do timeouts belong?
On every call a request waits for, shorter than the caller's own deadline, so a stuck hop fails and frees the worker.
23.Deadline
What is a propagated deadline?
The time the original request must finish, passed down so a hop does not start work the user has already abandoned.
24.Shed
What is load shedding?
Rejecting work you cannot finish, early, so the work you accept still meets the SLO. A queue that grows without bound sheds by timing out everyone.
25.Brownout
What is a brownout?
Turning off an expensive optional feature so the core action survives. The page is worse and still useful.
26.Redundancy
What does redundancy require?
The spare must not share the failure. Two processes on one disk, or two zones behind one gateway, fail together.
27.Failover
What do you learn from a failover test?
Whether the spare actually takes traffic, and how long that takes. A design that has never failed over is a hope.
28.Dr
What is a disaster-recovery test?
A restore or a region switch you actually run. A backup that has never been restored is an untested file.
29.Rpo
What is a recovery point objective?
How much data you can afford to lose, which sets how often you must replicate or back up.
30.Rto
What is a recovery time objective?
How long you can be down, which sets how fast the spare must be ready. A cold spare misses a short RTO.
31.Multi-az
What does multi-AZ buy you?
Survival of one zone's failure, if the other zone has the capacity and does not share a hidden dependency.
32.Capacity-fail
Why size for the failure?
When one zone dies, the other must hold the traffic. Splitting the fleet in half with no headroom means the failover is the outage.
33.Toil
What is toil in operations?
Repeating manual work tied to running the service. If it grows with users, it will outrun the team.
34.Auto-restart
What does a quiet restart loop hide?
A crash that users barely see because the process comes back. The crash rate is still a bug, and one day the restart will not be fast enough.
35.Error-budget-policy
What is an error-budget policy?
The agreement on what the team does when the budget is gone: feature freeze, reliability work, or a conscious decision to spend more.
36.Synthetic
What is a synthetic check?
A scripted user action from outside, on a schedule. It sees a failure even when no real user is calling, and it can miss failures only real data triggers.
37.Probe-gap
A probe passes and users fail. Why?
The probe does not do what the user does: it skips auth, uses a cached edge, or never buys anything.
38.Region-drain
What is draining a region?
Shifting traffic away before the region is empty, so in-flight work finishes and the other region is proven under load.
39.Dns
Why is DNS a slow failover?
Caches keep the old address for the TTL. A failover plan that ignores TTL waits on resolvers you do not control.
40.Health-deep
What is a deep health check?
One that touches the real dependency. Used as a load-balancer check, it can remove every pod when that dependency blips.
41.Health-shallow
What is a shallow health check?
The process is up and can take traffic. It stays green during a dependency blip, and the request path handles that blip itself.
42.Graceful
What is graceful degradation?
A reduced answer that still completes the user's main job. Failing the whole request because a recommendation service is down is the opposite.
43.Timeout-default
What is wrong with no timeout?
The call waits forever, holding a worker. Under load those workers run out, and the service fails for everyone.
44.Connection-age
Why recycle old connections?
A connection to a dead backend can sit in a pool looking idle. A maximum age, and a check on borrow, retires it.
45.Deploy-corr
How do you tie errors to a deploy?
The metric carries the version. Without it, a rise in errors is a guess about which change caused it.
46.Rollback-ready
What makes a rollback real?
The previous artifact is still there, the data is still readable by it, and someone can trigger it without a new build.
47.Oncall-load
What is an actionable page?
The receiver can do something now that the user will feel. A page that says 'look tomorrow' should be a ticket.
48.Silence
What is an alert silence for?
A known window, such as a test failover, with an end time. A silence with no end is how a real page gets lost.
49.Dependency-slo
Your SLO cannot outrun what?
The SLOs of the hard dependencies, combined. Five hops at 99.95 percent are not a 99.95 percent path.
50.User-journey
What is a user-journey SLI?
One measurement of a multi-step action, such as search-to-pay. Per-service green lights can all be true while the journey fails.