Hook
The logs show a stage boundary at 03:14 UTC. The request completed. Status: 200. Payload: a JSON object with twelve fields, every key present, every key typed correctly — and every value a placeholder string explaining what the field was supposed to contain.
The schema validated. The parser passed. The dashboard rendered. No alert fired, because no error occurred. Nobody writes an alert for the absence of an alert.
That is the whole problem in one sentence: a system that cannot distinguish "I found nothing" from "I failed to look" will report silence as fact. The ledger never lies, it only waits to be read — and a stage that never wrote anything is not a ledger at all.
I have seen this exact shape on-chain more times than I can count. The Graph returns {"data": {...}, "errors": [...]}. A naive client destructures data, discards errors, and renders a partial picture as a complete one. eth_getLogs on a block range your provider declines to serve returns an empty array. Zero is a number. Zero looks like data.
Context
The stack is always the same four hops: RPC → indexer → transform → dashboard. Each hop has exactly three failure modes. Error, timeout, and empty. Industry-standard instrumentation covers the first two comprehensively. The third arrives wearing the uniform of success.
JSON Schema is where the seam opens. required guarantees a key exists. It says nothing about whether the key contains anything. An empty array satisfies required. An empty string satisfies type: string. So a schema written to enforce structure will happily certify a document with the structural integrity of a skeleton and none of the marrow.
The specific failure I examined carried a diagnostic tell I want every analyst to memorize: the field values contained downstream instructions. Not data. Instructions — sentences addressed to the next stage, sitting where a payload belonged. That is not a data-loss problem. That is an interface written as prose and parsed as though it were a contract. If your schema's field contents read like a task description, you don't have a pipeline. You have a chain letter.
In 2025 I built a compliance dashboard tracking stablecoin reserves across 10 million transaction records. The final audit reported a 0% error rate. That number was only achievable because we refused to accept a clean zero. Every address with no transfers in a 24-hour window had to be classified: dormant treasury, or broken join. Two entirely different facts, indistinguishable at the presentation layer.
Core
Exhibit A — partial data on a 200. The subgraph response carries an error array beside the data object. Clients that destructure data and ignore errors will under-report every time a resolver times out. The dashboard doesn't flicker. It quietly loses a subset of reality. A checksum on the row count would catch this. Almost nobody computes one.
Exhibit B — the range cap. Most RPC providers cap eth_getLogs at roughly 10,000 blocks. Request more and you receive an error — or, depending on the provider, an empty array. Retry logic that treats empty as terminal records the gap permanently. The provenance of a missing event and the provenance of a nonexistent event become identical in your database. That is an evidence-integrity failure, not a UX bug.
Exhibit C — stale oracle rounds. latestRoundData() returns an answer and an updatedAt. Read only the answer and a feed that stopped updating three hours ago is indistinguishable from a live one. When I audited the first release of MakerDAO's contracts in 2018 — 450 lines of Solidity, traced by hand against the collateralization ratio logic — the two edge-case liquidation bugs I filed on GitHub both lived in branches where a zero-valued input satisfied a guard written for a different invariant. The code didn't lie. It answered a question I hadn't asked.
Exhibit D — the null-versus-zero join. A left join that finds no matching row yields null, and every aggregation downstream treats null as zero or as exclusion depending on the SQL dialect. Two different lies, one query.
Exhibit E — the empty batch. A rollup batches its transactions, the sequencer posts, the indexer reads, the transform flattens, the dashboard counts. When the posted batch carries nothing — which happens more often than the marketing suggests — the chart shows a flat line for the epoch. Not a bug. An accurate count of nothing. Dedicated data-availability capacity gets credited with demand it never carried.
In 2020 I mapped 50 whale addresses across early Uniswap V2 pools and found 30% of initial liquidity traceable to a single IP cluster. That finding existed only because I refused to accept "no attribution returned" as "no relationship present." Forensics is just history written in hexadecimal. The distinction between an absence and an unasked question is the entire discipline.
Contrarian
The intuitive fix — retry on empty — makes this worse.
A valid-but-empty response is deterministic. Retrying produces another valid-but-empty response, and now you hold three of them, and the third carries a timestamp that makes it look corroborated. Retries convert a single false negative into a pattern of false negatives, which is far harder to disbelieve.
The second blind spot is seasonal. In a bull market throughput is high and every chart slopes up, so a flat reading gets attributed to regime rather than to instrumentation. Everyone assumes the volume series is correct because it agrees with the price series. Correlation is doing the persuading; nobody has verified whether the underlying query returned rows at all.
The third blind spot is organizational. Errors get pagers. Timeouts get dashboards. Emptiness gets a null and a shrug, because emptiness has no owner and no severity code. The ledger never lies, it only waits to be read — but only if someone is still reading it.
Takeaway
Add minItems: 1 to your arrays. Compute a daily row-count baseline and page at ±40%. Run a canary query you already know the answer to, hourly, and fail loudly when it comes back empty. That last control is the whole fix, and it costs about twenty lines.
The forward-looking question is liveness, not accuracy. When a rollup's data-availability posting volume drops to zero for six consecutive batches — is that a quiet migration, an empty batch, or a broken indexer?
You can only answer that if you know whether the zero you are reading is a fact or a template.
