Cost models and indexing rules

Two things you configure decide more of your revenue than your hardware does. Both are usually left at defaults, and leaving them at defaults is a decision with a price.

Indexing rules: what you agree to index.

Rules tell indexer-agent which subgraphs to allocate to and how much. They can be global, per network, or per deployment, and the modes are the interesting part:

• always: index this, regardless of the economics.
• never: do not, regardless.
• rules: decide by thresholds you set, on signal, on stake, on fee expectations.
• offchain: index it without allocating stake, so you can serve it without earning indexing rewards on it.

Most operators run threshold-based rules with a handful of explicit overrides, and that is a sensible shape. The overrides are where your judgement lives.

The competition problem with following signal.

If your rule is “allocate to the highest-signal subgraphs”, you will be allocated alongside everyone else running that rule. Indexing rewards on a subgraph are shared across the stake allocated to it, so crowding reduces what each allocation earns.

The counter-position is not automatically better. Allocating to low-signal subgraphs because they are uncrowded is allocating to subgraphs the market thinks are not worth much, and it may be right.

The genuine edge is in disagreeing with the market for a reason you can state: a subgraph you know is about to carry an application’s traffic, a network you have specific competence in, a deployment you can serve at lower cost than others. Everything else is following.

Cost models: what you charge.

You publish a cost model written in Agora. It maps query shapes to prices, and it is how the gateway knows what serving your answer costs.

The structure is a set of predicates and prices. Match a query shape, apply a price, otherwise fall through to a default.

Two ways to lose money here, and they are symmetric:

Priced too high. The gateway routes elsewhere. You have capacity, you index correctly, and you serve nothing. From your side this is indistinguishable from there being no demand, which is why it can persist for months.

Priced too low. You serve expensive queries at a loss. A query that walks a large result set costs you real resources, and a flat cheap price on everything means the worst queries are the ones you attract most.

The second is why a flat model is a poor idea even though it is the easy one. Query cost is not uniform, and a single price is either too high for the cheap ones or too low for the expensive ones.

How the two interact.

Rules decide where your stake goes. Cost models decide whether queries come to it. They fail together in a specific and common way:

You allocate to a high-signal subgraph, so you have stake committed and you are earning indexing rewards. Your cost model prices you above the field for that subgraph’s typical query shape, so you serve none of its queries. You are earning issuance and no revenue, on a subgraph you chose because it looked valuable.

Nothing is broken. Half the point of the allocation is simply not happening, and no metric you would naturally watch says so.

Compare query volume against your allocated stake, per subgraph. Where the ratio is poor, the cost model is the first place to look.