5 exercises on the language of performance — reading p99/p95 metrics, expressing uptime in "nines", describing improvements, and calculating error rates.
three nines = 99.9% uptime = ~8h 46m downtime/year
RPS / TPS / QPS — throughput units (requests/transactions/queries per second)
error rate = failed requests ÷ total requests × 100%
Describe changes: "decreased/dropped from X to Y — an N% reduction"
0 / 5 completed
1 / 5
A monitoring dashboard shows: "p99 latency: 420ms". What does this metric mean to a backend engineer?
Percentile latency (p50 / p95 / p99) — how to read and say them:
Percentile metrics tell you the distribution of request latency, not just an average.
How to read it: "p99 = 420ms" means that 99% of all requests completed in 420ms or less. Only 1% of requests were slower than 420ms.
How to say it aloud:
"p99 latency" = "the ninety-ninth percentile latency"
"p95" = "the ninety-fifth percentile"
"p50" = "the fiftieth percentile" (= the median)
Why percentile metrics matter more than averages: An average can hide severe outliers. If p50 is 20ms but p99 is 8 seconds, most users are happy — but 1 in 100 requests is terrible. Averages smooth that out. Percentiles expose it.
Common values and their meaning:
p50 — the median. Half of requests are faster.
p95 — 95% of requests are within this threshold. Gold standard for SLA targets.
p99 — 99% of requests pass. Identifies tail latency issues.
p99.9 — the "three nines" percentile. Critical for revenue-impacting transactions.
Common in SLOs: "p95 latency must remain below 200ms under normal load."
2 / 5
An SRE writes in a post-mortem: "We maintain three-nines availability. This incident reduced us to two-nines for approximately 90 minutes." What does "three-nines" mean and how much downtime does it allow per year?
The "nines" vocabulary of availability:
Availability is expressed as a percentage, and the informal shorthand counts the 9s:
Name
Availability
Downtime/year
Two nines
99%
~87.6 hours
Three nines
99.9%
~8.76 hours
Four nines
99.99%
~52.6 minutes
Five nines
99.999%
~5.26 minutes
How to say the percentages:
"ninety-nine point nine percent" = 99.9%
"ninety-nine point nine nine percent" = 99.99%
"three-nines" — the informal collocation (hyphenated as adjective: "a three-nines SLA")
SLA / SLO language:
"We have a three-nines SLA with the client."
"Our error budget is 8 hours 46 minutes per year."
"We burned through our error budget in Q3."
3 / 5
A developer presents optimisation results. Which sentence describes a performance improvement most precisely?
Avoid vague language in technical reporting: "much faster", "significantly improved", "a lot quicker" — all lack the numbers needed for credibility.
4 / 5
An incident report states: "The error rate averaged 0.5% over the 2-hour incident window, during which the service processed approximately 3 million requests." How many requests failed?
"0.5% error rate" = "zero point five percent" = 5 errors per 1,000 requests
"0.01% error rate" = "zero point zero one percent" = 1 error per 10,000 requests
"5 nines" availability corresponds to 0.001% error rate
Common expressions in incident reports:
"The error rate spiked to 12% at peak."
"Error rate remained elevated throughout the window."
"We saw a sustained error rate of 0.5% for approximately 2 hours."
"That translates to approximately 15,000 affected requests."
Units for request volume:
1,000 = one thousand
1,000,000 = one million (1M)
1,000,000,000 = one billion (1B)
Real-world context: A 0.5% error rate sounds small, but on a high-traffic system (3M requests in 2 hours = 25,000 requests/minute), it means 15,000 failed user interactions — significant impact.
5 / 5
A load test result shows: "The service handles 50,000 requests per second at peak load with p95 latency under 120ms." How would an SRE describe this verbally in a meeting?
Saying performance metrics aloud — register and accuracy:
All three versions are technically acceptable and frequently heard. The register differs:
Option A — clear, accessible, fine for mixed audiences: "fifty thousand per second", "95% of requests are done in 120ms or less"
Option B — formal, full pronunciation, appropriate for clients, reports, or presentations: "fifty thousand requests per second", "p-ninety-five latency", "one hundred and twenty milliseconds"
Option C — technical shorthand, natural among engineers: "fifty-K RPS" (requests per second), "ninety-five percentile"
Common throughput vocabulary:
RPS — requests per second (most common in web/API contexts)
TPS — transactions per second (databases, financial systems)
QPS — queries per second (search, databases)
throughput — general term for processing rate
Saying large numbers:
50,000 → "fifty thousand" / "50K" / "fifty-K"
1,200,000 → "one point two million" / "1.2M"
3,500,000,000 → "three point five billion" / "3.5B"