loader_cache uses the resolver chain as a metric label (3,684 series on one site, 21,849 fleet-wide) #1219

Measured

blocks/loader.ts labels the loader_cache counter with ctx.resolverId:

const loader = ctx.resolverId || "unknown";

resolverId carries the full resolver chain, so label values look like:

Categories@sections.variants.1.value.5.sections.0.section.page
SearchResults Global@sections.0.section.page
Categories@sections.variants.1.value.2.jsonLD

Every section position, in every page variant, on every page becomes a distinct series.

Production ClickHouse, distinct loader values in 3h:

Against the budget documented in our own o11y guide — >1,000 per site is forbidden as an alert dimension, >100 across the fleet as a cross-fleet aggregation — five sites are over, and the fleet figure is 218× over.

Cost, same window:

• loader_cache = 11,485,954 of 14,998,196 rows in otel_metrics_sum — 76.6% of the whole table
• otel_metrics_sum on disk: 48.4 GB

For contrast, cache_hit covers 194 tenants in 9 series, and the @decocms/start runtime does the equivalent job with 37 series and 13 distinct deco.cache.profile values (vtex/productDetailsPage, vtex/relatedProducts, site/categoryTree).

Why this is an issue and not a PR

The obvious fix — label with the resolver name instead of the chain — loses information. The chain tells you which section of which variant of which page a loader instance sits in, and that is genuinely useful when answering "which section is slow". Truncating trades debuggability for cost, and that is a call for whoever owns the metric, not a mechanical patch.

Options worth weighing:

1. Label with the resolver name only (vtex/loaders/legacy/productListingPage.ts), matching the @decocms/start design. Cheapest, loses section attribution.
2. Keep the chain but only on a sampled/opt-in basis, with the name as the always-on label.
3. Keep the chain as a span attribute (where high cardinality is acceptable for point lookups) and use the name for the metric.

Option 3 preserves both properties and is what the o11y guide's own rule implies — high-cardinality attributes are fine for filtering and point lookups, forbidden as aggregation keys.

Happy to send the PR once there is a direction.

Context

Surfaced while building unsampled loader-cache panels on the general-overview-otel dashboard. Those panels are usable today only with a tenant selected — a fleet-wide GROUP BY on this label is precisely the forbidden case.

🤖 Generated with Claude Code