Baselines
X-ray's memory of what normal looks like on master. For every service it remembers what a
typical Tuesday 14:00 — or any other hour of the week — looks like: how much traffic, how many errors, how slow, how
much Neo4j load. The Anomalies system alert compares live
numbers with this memory and raises a finding when
something is far outside it.
X-ray has not learnt this hour yet.
27 hours compared with their normal; 4 could not be — X-ray has not learnt those hours yet; 4 known from too few samples to judge.
- Mon 21:00 2.0 min, normal ≈ 275 ms 436.4× normal see it →
- Tue 03:00 2.0 min, normal ≈ 368 ms 326.5× normal see it →
- Mon 22:00 2.0 min, normal ≈ 445 ms 269.9× normal see it →
- Mon 20:00 2.0 min, normal ≈ 938 ms 128.0× normal see it →
- Mon 19:00 2.0 min, normal ≈ 1.10 s 109.1× normal see it →
- Mon 08:00 31.04 s, normal ≈ 1.32 s 23.4× normal see it →
- Mon 03:00 15.50 s, normal ≈ 858 ms 18.1× normal see it →
- Tue 09:00 12.15 s, normal ≈ 919 ms 13.2× normal see it →
- Mon 16:00 12.02 s, normal ≈ 945 ms 12.7× normal see it →
- Tue 00:00 9.13 s, normal ≈ 712 ms 12.8× normal see it →
- Mon 07:00 5.32 s, normal ≈ 867 ms 6.1× normal see it →
- Tue 02:00 3.09 s, normal ≈ 461 ms 6.7× normal see it →
Not learnt yet: 30 scattered hours. An hour counts once two different weeks stand behind it, so the rest fills in as weeks pass.
whole cluster · Latency p95 response time
Rows are days, columns are hours of the day, in UTC — baselines are learnt per UTC hour, so the grid cannot shift with your timezone; the cards above give the times in UTC as well. Hover a cell for its numbers. ?How a finding is decided: three of the last four complete five-minute readings must be far outside their hour's normal — at least 4 scaled MADs and 35% away from the median, and above a small absolute floor so a quiet service cannot alarm on noise. An hour needs 20 readings behind it before it may raise anything. The thresholds are editable on the Anomalies system alert.