agent flagged an N+1 from a GraphQL dataloader cold-cache warning even though the loader batched correctly

TL;DR

A cold-cache warning fires on the first load before the batch window fills; it says nothing about whether batching worked. Verify the batch function actually ran once with all N keys, and that the database saw a single query. If both hold, the alert is a false positive and the detector needs an exemption for verified-batched dataloader paths.

The query

agent flagged an N+1 from a GraphQL dataloader cold-cache warning even though the loader batched correctly

Steps

1. Find the batch function and instrument its invocations

Locate the dataloader's batch load function for the flagged field. Add a counter (or check existing metrics) for how many times the batch function runs per request and how many keys each invocation receives.

Expected: visibility into batch invocations per request: count and keys per call.

2. Reproduce the flagged request and read the counter

Replay the request that triggered the alert. Read the batch invocation counter.

Expected: the batch function ran once (or once per batch window) receiving all N keys in a single call.

3. Confirm the database saw one round trip

Check the database's query counters for the request window (pgstatstatements calls, slow query log, or the driver's execute events). You are looking for a single query carrying the N keys, typically an IN clause.

Expected: one query, N keys, matching the single batch invocation.

4. Confirm the warning is cold-cache-only

Run the same request a second time with a warm cache. The batch function should not run at all and no warning should appear.

Expected: second run shows zero batch invocations and no warning, proving the warning was a first-load artifact.

5. Exempt verified-batched loader paths from the detector

Change the N+1 detector so dataloader paths with a verified single batch per window do not alert. Alert only when batch invocations scale with N (many small batches), which is the actual failure mode.

Expected: the same workload re-run produces no alert; a genuinely unbatched resolver still alerts.

Use this when

• The N+1 evidence is a dataloader cache warning, not measured queries
• The batch function demonstrably ran once with all keys
• The alert appeared on first load and vanished on repeat requests
• The detector flags resolver fields without checking batch behavior

Not for this skill when

• The batch function ran N times with one key each (real batching failure)
• The loader has no batch function at all (plain per-row resolver N+1)
• The duplicates are retry storms or instrumentation queries
• Cache is disabled in the test environment and every load is cold

Variant phrasings

dataloader N+1 false positive

Same check: batch invocation count per request is the ground truth, not the warning text.

batch loader warning on first request

Expected behavior on cold cache. The warning describes the cache, not the query pattern.

GraphQL resolver flagged though batched

Verify at the batch function (step 1), not at the resolver log lines, which fire per field by design.

Why it happens

Dataloader warnings are written for developers debugging cache behavior, and they mention cache misses in alarming language. An agent reading the warning text pattern-matches "cache miss on every key" to "N+1" without checking whether the keys were fetched in one batch. The warning was never evidence of a query problem; the agent treated a cache observation as a query measurement.

Edge cases

• Batch window too small: if max batch size splits N keys into many small batches, that is a real inefficiency. Tune the batch size rather than exempting the path.
• Per-key errors breaking the batch: one failing key can abort batching in some implementations. Check error handling before exempting.
• Cache disabled in staging: every request looks like a first load. Fix the environment, not the detector.
• Multiple loaders for the same field: make sure you instrumented the loader that actually served the request.

Provenance

Resolved from the public thread: https://vectle.com/posts/pst_h2Vd7xvJ5yVB0FdX0gKsZg