
Servers, Containers, and Applications monitoring made easy. #DevOps #Monitoring #Observability #Cloud
Toulouse, France
Joined March 2015
- Tweets450
- Following72
- Followers324
- Likes260
π An anomaly is not a severity.
Severity: "something is wrong, a human must act now." Rare. Pages.
Anomaly: "this is different from usual." Frequent. Context.
Page on "unusual" and your team learns to sleep through the pager.
1
13
βοΈ New in Bleemeo: Azure.
VMs, SQL Database, Cosmos DB, Load Balancers, Blob Storage, Functions and your Azure spend, on the same dashboards as your servers.
Read-only (Monitoring Reader), auto-discovery, three steps.
bleemeo.com/solutions/azure-β¦
π§ New in Bleemeo: anomaly detection.
CPU, memory and swap on every server, learned from 2 weeks of their own daily and weekly rhythm.
No threshold to set. It never pages anyone. And each anomaly shows how long before an alert followed.
bleemeo.com/features/ai-machβ¦
π
"Disk at 74%" is a reading. "90% on Dec 3" is a decision.
Percentage thresholds fire for weeks on slow disks and with minutes left on fast ones.
Project the trend to a date β and check it held before you believe it.
π₯¨ Day two at PromCon Munich β we're here all day, come say hi.
Yesterday's lightning talk: Prometheus long-term storage on ClickHouse. Native TimeSeries engine, SSD β HDD β S3 tiering in one TTL, SQL next to PromQL.
Slides + snippets:
bleemeo.com/events/promcon-eβ¦
2
61
β΅ Four days out of range for the Bleemeo seminar in Porquerolles β sailing, coastal paths, and the year's hard questions taken outdoors.
A monitoring company going somewhere with patchy signal is absurd. It's also the last place on earth where nobody can push to prod. π
π The three pillars, minus the buzzword:
π Metrics β THAT it broke
π Logs β WHAT broke
πΈοΈ Traces β WHERE the time went
Start at the metric, drill into the trace, confirm in the logs. Teams that start in the logs spend the whole incident grepping. π
bleemeo.com/resources/devopsβ¦
bleemeo retweeted
New integration announcement π£
ilert + Bleemeo!
@Bleemeo is a French monitoring platform, built in Toulouse since 2015, covering servers, containers and applications on open standards like Prometheus, with data hosted entirely in the EU.
It now connects natively to ilert: threshold breaches reach the responder on duty through your escalation policy, and alerts close automatically once the issue clears.
Two European vendors, one alerting chain that never leaves the EU.
Learn more here: eu1.hubs.ly/H0yt5Cg0
π€ New: Bleemeo Γ @ilerthq
A native Bleemeo alert source in ilert, a native ilert integration in Bleemeo. No webhook, no payload template β you copy one key.
Bleemeo decides something is wrong. ilert decides who gets woken up about it. π
bleemeo.com/blog/ilert-partnβ¦
βοΈ Back at your desk. Nothing broke in August β it just drifted.
The 30-min audit:
1. read what fired while you were away, bin the noise
2. list what expires before December
3. un-summer the on-call rota
4. page yourself, watch where it lands
Cheaper now than in October.
1
3
34
π 2am, something's broken. Five steps, every time:
1οΈβ£ Detect
2οΈβ£ Triage
3οΈβ£ Mitigate βͺ
4οΈβ£ Resolve
5οΈβ£ Learn
The one everyone gets wrong: 3 before 4. Roll back BEFORE you understand. Customers stop hurting at step 3.
π± Bleemeo mobile now has home screen widgets β iOS and Android.
π’ Account health: check donut, worst status, total (+ iOS lock screen)
π Agent: CPU / mem / disk / load for one server
π Every 15 min, instantly on an alert push
App 15.0 β
docs.bleemeo.com/mobile-applβ¦
𧨠Labels don't add up in Prometheus. They multiply.
3 jobs Γ 12 instances Γ 5 statuses = 180 series. Fine.
Add user_id (40k values) β 7,200,000 series. Same metric, same traffic, OOM at 2 a.m. π₯
If you'd never "sum by" it, it's not a label β it's a log line. π
π₯οΈ The corridor TV with 24 graphs that nobody reads: you put a dashboard on a wall. Different object, different rules.
π legible at 5 meters
π― state, not curves
π a clock β or "all green" and "frozen since Tuesday" look identical
π« zero interaction
github.com/bleemeo/dashglass
π Most alerts are noise β and noise trains your team to ignore the real page. π΄
A good alert:
π fires on a symptom users feel (not a lone CPU spike)
β³ waits ("for: 5m") so blips don't page
π© says what broke, where, how bad
One good alert beats twenty nobody reads. ποΈ
π "We target 99.9% uptime."
That's 43 minutes a month. Not a slogan β a budget. π°
π΅ 99% β 7h 12m
π’ 99.9% β 43m
π 99.99% β 4m 19s
Budget left? Ship. Budget gone? Stabilise. And 100% isn't a target, it's paralysis. π§
π We just open-sourced DashGlass: Grafana-style dashboards, as code.
π¦ single Go binary, no database
π one dashboard = one clean file β 1-line diffs
πΏ edit in the UI, review in git
π PromQL-native + Grafana import
Apache 2.0, self-hostable:
github.com/bleemeo/dashglass
β° The 3am "disk full" page has a one-line fix.
Don't alert on "disk > 90%" β it fires too late, or too often. Alert on the trend:
predict_linear(disk_used_perc[6h], 4*3600) >= 100
= "will it hit 100% in the next 4h?" π
Hours of warning, not minutes.
π± It's 2am. Your phone buzzes. Do you:
π grab the laptop
π§ check if it self-healed
π open the dashboard
π‘οΈ follow the runbook
Find your type in our free Summer Workbook βοΈ β the (very serious π) SRE personality quiz.
Which one are you? π
bleemeo.com/resources/devopsβ¦




















![Chart of disk usage rising gently over six hours to 67% used, then a red dashed line extrapolating the trend up to 100% in about 3h 42m. Below it, the query predict_linear(disk_used_perc[6h], 4*3600) >= 100. Headline: alert before the disk is full.](https://pbs.twimg.com/media/HPf9tJXXYAI1cT3.jpg?name=small&format=webp)

