Percentiles lie when the sample is small
A p99 computed over 40 events is not a p99. What to report instead when traffic is thin.
A p99 computed over 40 events is not a p99. What to report instead when traffic is thin.
HyperLogLog gets you within two percent and keeps the log free of identifiers.
Peeking at results daily inflates false positives. A short note on sequential testing.
The most common reporting bug we see, and a one-line fix.