There is a job that did not exist a decade ago and now quietly sits on a lot of teams. Someone whose week goes on making the observability bill smaller. They tune retention windows, argue with a vendor about ingest tiers, write scripts to drop the logs nobody reads, model the saving from a different pricing plan. It is real, careful work, and it is rewarded, because the number it moves is a number the business can see.

I started turning this over after listening to Yuri Massar of Better Stack pull at the topic on the Code Red podcast. His framing, that the industry has talked itself into financial engineering on the monitoring invoice, stuck with me. The price figures in that conversation are his and his company's to defend, not mine, and I am not going to repeat them as fact. The shape of the argument is what matters, and the shape is right. When a whole discipline grows up around optimising a bill, the bill is not the problem. It is the symptom.

Financial engineering is what you do when you cannot change the underlying thing, so you rearrange the costs of it instead. It is moving money around to make a number look better without making the thing underneath any better.

That is exactly what most observability cost work is. The data was never designed, so it arrives in a shape nobody chose, in a volume nobody decided, and the only lever left by the time it reaches the invoice is to negotiate, compress, sample blindly, or delete. All of that is effort spent downstream of the actual decision, which was never made. You are not engineering your observability. You are engineering the receipt.

Ask why the volume is what it is and you get the same three answers, every time.

We instrument for fear. A bad incident leaves a mark, and the response is to add signals so it never surprises us again. Each one feels prudent on the day. Nobody adds up the running total.

We keep everything, because deleting feels risky. Retention creeps up because no one can say out loud which data informs a decision and which is just sitting there accruing charges in case a future self wants it. Default retention is a decision nobody made, paid for monthly.

And we never decide what a signal is for. This is the one underneath the other two. A metric, a log line, a span, each one exists to answer a question or support a decision. Most estates cannot tell you the question. The signal is there because it was easy to emit and frightening to remove. Cost is downstream of that absent decision, and no pricing plan fixes an absent decision.

The real work is less glamorous than a renegotiation and far more durable. It is deciding, signal by signal, what each one is for.

Every signal gets a purpose: the question it answers or the decision it informs. If it has neither, it is not cheap data, it is waste you are paying to store. Retention then attaches to that purpose rather than to a default. Data that supports a regulatory decision lives as long as the obligation; data that only helps the first thirty minutes of an incident does not need to live for a year. Sampling and aggregation become deliberate choices about fidelity, made where you can still see the trade, not panic settings reached for when the bill lands.

Do this and the cost review stops being a cost review. It turns into a design review wearing a different name, which is what it always should have been. You are no longer asking what this costs. You are asking what decision it serves, and the cost takes care of itself because the volume was chosen rather than inherited.

There is a reason this stops being a tidy-your-spend story and becomes urgent. The data layer you own is about to become the thing your automation stands on.

As AI assisted triage arrives, it works the same data you have been paying to hoard. An automated investigator is only as good as the layer beneath it, and a layer nobody designed is a layer nobody can reason about. If you cannot say what a signal is for, neither can the system you are asking to act on it. A bill you cannot explain is a data estate you cannot automate on, because automation needs the one thing the estate never had: a decision about what each piece is for. The teams that did the design work are not just spending less. They are the only ones with a foundation worth putting an agent on top of.

So before you hire another person to shave the invoice, or move to the vendor with the friendlier per-gigabyte line, ask the question that sits one level up. What is each signal for. Most of the bill is the cost of never having answered it.

The cheapest gigabyte is not the one you negotiated down. It is the one you decided not to collect.

The cost-as-design argument runs all the way through the book. The first chapter of Metrics & Mayhem is free, no signup wall on the first read. If you would rather get this thinking weekly, the Observability Digest is here.