Reliability language: SLIs, SLOs, SLAs
Senior DevOps roles are increasingly "DevOps/SRE" in practice: you're expected to define reliability targets, run on-call, lead incidents, write blameless postmortems, plan capacity and reduce toil. Interviewers ask behavioural questions ("tell me about an outage you handled") alongside concepts like SLOs, error budgets and burn-rate alerting. This topic gives you the vocabulary, the maths and the structure for strong answers.
Definitions
| Term | Meaning | Example |
|---|---|---|
| SLI (indicator) | A measured ratio of good events to valid events | Proportion of checkout requests served with non-5xx status in under 400 ms |
| SLO (objective) | Target for an SLI over a window | 99.9% of checkout requests good over 28 days |
| SLA (agreement) | Contractual promise with consequences (credits) | 99.5% monthly availability or 10% credit |
| Error budget | 1 − SLO: the allowed unreliability | 0.1% of requests, about 40 minutes of total downtime in 28 days |
SLAs should be looser than internal SLOs so you get warned before breaching contracts.
Choosing good SLIs
Measure what users experience, as close to them as possible (load balancer or edge logs, synthetic probes, client RUM — not just pod CPU):
| Service type | Typical SLIs |
|---|---|
| Request/response API | Availability (non-5xx ratio), latency (ratio under threshold) |
| Pipeline/batch | Freshness (data updated within X), correctness, coverage |
| Storage | Durability, availability, latency |
| Streaming/queue | End-to-end lag |
Express SLIs as ratios: good events / valid events. Exclude invalid traffic (health checks, 4xx caused by clients) deliberately and document it.
Availability nines
| SLO | Downtime per 30 days | Per year |
|---|---|---|
| 99% | 7.2 hours | 3.65 days |
| 99.5% | 3.6 hours | 1.83 days |
| 99.9% | 43.2 minutes | 8.76 hours |
| 99.95% | 21.6 minutes | 4.38 hours |
| 99.99% | 4.3 minutes | 52.6 minutes |
Dependencies multiply: a service calling three serial dependencies each at 99.9% can be at most about 99.7% unless it handles their failures (retries, fallbacks, caching). Each extra nine costs roughly 10× more in engineering and redundancy — choose targets by user need and business value, not "as high as possible".