The configuration advisor
Monitoring rots quietly. A channel gets detached, a grace window set once for a job that has since got slower produces a false alarm a week, a monitor paused during an incident six weeks ago is still paused. None of it looks like failure: it looks like silence, and silence is what a healthy monitor looks like too.
The advisor tells those two silences apart. It runs about every two hours and writes its findings to the advisor page.
Every check
| Finding | Severity | Weight | What it means |
|---|---|---|---|
| No alert channel attached | critical | −25 | This monitor can go down and nothing will tell you. It records state and history as normal, so it looks healthy from the dashboard right up until you need it. |
| Channel has never delivered anything | critical | −20 | A channel that has never succeeded is not a channel. The first time it is needed will be the first time it is tested. |
| Last delivery to this channel failed | high | −15 | A webhook that started returning 404, an email address that now bounces. The alerts are being generated and thrown away. |
| Paused for a long time | high | −12 | Paused monitors never alert. Most were paused during an incident and never resumed, so the job it watches has been unmonitored ever since. |
| Certificate expiring soon, with no SSL alert enabled | high | −10 | The certificate expires inside 30 days. When it does, this monitor will report the site down and the cause will not be obvious from the alert. |
| Never received a ping | high | −8 | Created but never used. Either the ping URL never made it into the job, or the job it was made for never shipped. |
| Flapping between up and down | medium | −8 | Repeated short outages that recover on their own. Usually a threshold set slightly too tight rather than a service that is genuinely failing every few hours. |
| Grace window is tighter than this job actually varies | medium | −6 | The job is arriving late often enough to trip the window without having failed. Every one of those is a false alarm, and false alarms are how people learn to ignore the real ones. |
| Grace window is wider than the job’s own period | medium | −6 | The job could miss a full run and nothing would fire until well into the next one. The monitor is real but slower to notice than the schedule it watches. |
| Metric rule stuck learning | medium | −6 | The bucket it compares against is not accumulating samples, so the rule has never been able to evaluate. Usually the wrong bucket for how often the job runs. |
| Repeat interval is aggressive enough to be self-defeating | medium | −5 | Re-alerting this often during a long outage produces a wall of identical messages. People mute the channel, and then the next unrelated alert is muted too. |
| Approaching a plan limit | medium | −4 | Says what happens at the limit rather than only that it is near, so the choice can be made before it is made for you. |
| Checking http:// on a host that redirects to https | low | −3 | You are measuring the redirect, not the site. A TLS failure or an expired certificate would not show up here at all. |
| The same error, over and over | low | −3 | This fingerprint has fired many times without the underlying error changing. If it is known and not actionable, muting it makes the rest of the alerts worth reading. |
| Rule has never fired in 90 days | low | −2 | Either the thing it watches has never gone wrong, or the rule does not match what it was meant to match. Worth a look to know which. |
| Still on its inferred schedule | low | −2 | The schedule was guessed from the first few pings and has never been confirmed. A guess that was close enough then may not be now. |
| Fleet instance missing long enough to retire | low | −2 | A machine that has not reported in weeks is usually a machine that no longer exists. Until it is retired it counts against the fleet and keeps the breach alive. |
| A feature on your plan that would catch something | low | −1 | Described by what it would have caught, not as an upsell — it is already included in what you pay for. |
The health score
100 minus the weight of every open finding, floored at zero. That is the entire formula, and it is printed next to the number on the page: a score nobody can explain is a score nobody trusts.
Two properties matter more than the exact weights. It cannot only go up. Adding a monitor with no channel lowers it immediately: a score that improves whenever you use the product more would reassure you while the thing it measures gets worse. It is monotonic. Fixing something never lowers it, and each finding costs its own weight regardless of what else is open, so “why did it move” always has a one-line answer.
There is no badge on the dashboard when everything is fine. A permanent 100 becomes furniture within a week, and furniture is what people look past on the night it reads 62.
Dismissing a finding
Every finding can be dismissed with a reason, and the dismissal is remembered per finding per monitor. Without that the page becomes wallpaper after a fortnight: a list nobody reads because it always says the same thing.
A dismissed finding comes back only when it gets materially worse: half again the figure it was dismissed at, not one day more. A monitor dismissed at 30 days paused does not return at 31; that is the same fact with a bigger number, and re-raising it would teach you that dismissing does nothing. At 45 days it is a different fact.
The weekly email
Opt-in, off by default, and it is a section added to the Monday digest rather than a second email: two weekly emails from one product is how both stop being opened. Only critical and high findings reach the inbox; everything else waits on the page, where you went looking for it. A week with no findings sends nothing extra at all, because an “all clear” every Monday is how a reader learns to skip the whole thing.
What it does not do
It never changes your configuration. Every fix on the page is a link to the screen that owns that setting: one place that can change a monitor is worth more than a second one that has to stay in step with it. It is on every plan, including free.