Basis Lab

Watching a Model After Launch

What Deserves an Alarm

Last timeThe Model Changing Its Own Inputs

An alarm nobody can act on is noise with a pager attached, so every threshold has to be computed from the ordinary variation and tied to something a person can actually do.

An alarm is a request for an action

The previous lessons produced a long list of things that can be measured. This

lesson is about the much shorter list of things worth waking somebody for, and

the distinction is not about how important the signal is. It is about whether

anybody can do anything.

Suppose the input drift number on a column doubles. Somebody is paged at two in

the morning. What do they do? They cannot retrain a model at two in the

morning, and they should not. They look at a chart, note that a column moved,

and go back to bed. The alarm was correct and useless, and its real effect was

to make the next alarm slightly less likely to be taken seriously.

FIG 1Where each signal belongs
can somebody act within is the action obviouspageticketweekly report
a required input is sudd11100
scoring requests are fai11100
volume has collapsed11100
a feature lookup is retu11100
an input column has drif00011
labelled accuracy has fa00011
the model is refusing mo00001
Coded 1 for yes and 0 for no. The split is clean and it is not a split by importance: the last three rows are the signals that matter most for whether the model is any good, and none of them belongs on a pager, because none of them can be acted on at speed. The first four are comparatively dull and all of them are worth being woken for, because each one has a specific repair that works immediately.

The arithmetic of false alarms

The reason to be strict about this is that false alarms are not a nuisance,

they are the mechanism by which monitoring dies. And their rate is computable

before anything is built.

FIG 2How long between false alarms
how many checks run in a day
the chance that one check fires when nothing is wrong
days between false alarms
Forty checks a day at a one in a hundred threshold gives a false alarm every two and a half days. That is roughly two a week, every week, for ever, and two a week is enough for a team to stop reading them inside a month. The arithmetic is worth doing before building the checks rather than after, because the usual response to the noise is to raise the thresholds until the system is silent, which is worse than not having built it.

The same arithmetic in the other direction gives the threshold a set of checks

can afford. If sixty checks run hourly and one false alarm a week is the

budget, each check has to be set so it fires about once in ten thousand runs

when nothing is wrong, which is a much stricter threshold than anybody chooses

by eye.

The lesson stops here

4 more paragraphs to go

You have read the opening. The rest of the argument, the problems that check whether it landed, and the lines worth keeping at the end all come with a plan.

The first lesson of every course in the library reads the whole way through, free, so you can see exactly what the rest of them are.

See the planThe contents

This is the reading half

Starting the course gives you your own copy of it. Every idea on every page has problems standing under it, marked with a reason rather than a tick, and any sentence you do not believe can be opened and argued with. None of that can happen on a page nobody owns.

The contents

The rest of this course

  1. 01What You Can Measure Without Labels
  2. 02When the Inputs Moveopening only
  3. 03When the Right Answer Movesopening only
  4. 04Labels That Arrive Weeks Lateropening only
  5. 05The Model Changing Its Own Inputsopening only
  6. 06What Deserves an Alarmyou are here
  7. 07Deciding to Retrainopening only
  8. 08The Failures That Make No Noiseopening only

Read alongside