Default Tracing Needs an Explicit Collection Policy

Summary

Vercel made Tracing generally available on November 6, 2026, across all plans. The release adds project and team trace search, shared KQL queries across the dashboard and CLI, and Live mode. Eligible new projects receive automatic collection, while retention expands to three days on Hobby and 14 days on Pro and Enterprise. Pricing is unchanged from beta. Vercel announcement

The operational question is what evidence a team can expect to have when something fails. Collection defaults need an owner, a budget and a recovery procedure.

Key Updates

Default enablement is conditional. Projects created on or after November 5 qualify when built on Vercel with Next.js 16.5 or later, at least one Node.js function and tracing not disabled. The first successful build adds an all-environment 100% sampling rule. Prebuilt deployments, static exports and Edge-only projects do not qualify. Existing instrumentation can suppress automatic OpenTelemetry injection. Other projects need an explicit sampling rule. Configuration

Budget indexed span units. Hobby includes one million monthly; Pro costs $0.50 per million, and Enterprise starts there. An indexed span uses one unit per 2 KiB, rounded up. Exhausting Hobby's team quota pauses collection across every project. Resetting the billing period does not resume it automatically; an authorized team member must resume collection. Configuration

Rule order determines coverage. The first matching sampling rule wins; unmatched traces are not collected. Sampling keeps or drops a whole trace. Trace Drains use separate rules: no Tracing rules means no collection, whereas no Trace Drain rules means forwarding all traces. Teams should inspect collection and export policies separately. Tracing documentation

Instrumentation remains part of the contract. Adding your own instrumentation file takes over from Vercel's injected setup. Listed OpenTelemetry dependencies can also transfer responsibility for fetch spans to the application; without SDK registration, those spans can be absent. Context-propagation controls distinguish destinations that receive trace context from URLs excluded from tracing. Neither distinction establishes a comprehensive sensitive-data policy. Instrumentation

Why It Matters

Our earlier workflow-and-trace coverage focused on reading parent-child execution, latency and cost to diagnose agent behavior. This follow-up concerns whether that evidence was collected and remains available. A convincing waterfall cannot explain a request that never entered the retained dataset.

Vercel documents a 10 MB compressed trace-data limit per request. Oversized attributes may be truncated; spans still exceeding 1 MB are dropped. Edge-runtime custom spans do not appear. These limits further constrain what a 100% sampling rule means. Tracing limitations

Our interpretation: absence from a trace search should trigger a coverage check before it becomes evidence that an operation never happened. This article is based on vendor documentation reviewed November 8, not an independent reliability benchmark or a measured cost-saving result.

Builder Takeaway

Run a representative request through each environment and compare the expected operations with the resulting trace. Record the sampling policy, instrumentation owner, retention window and quota-recovery owner alongside that check. Estimate costs from observed span sizes and volume, then test exclusions and export destinations separately.

Review the fields your application records before broad collection. Inspect actual emitted data and destination settings; do not assume tracing availability proves sensitive values are removed. Repeat the coverage check after instrumentation or deployment changes.