Course contentsShow
← System Design
Lesson 1258 of 1,91931. Observability: TracingFree lesson

Cardinality Explosion

How high-cardinality tags (user IDs, request IDs) multiply span volume; strategies to limit tag cardinality.

Cardinality Explosion

What you'll learn: How high-cardinality tags multiply trace span volume and strategies to limit cardinality in distributed tracing.

The Problem

Cardinality is the number of unique values a tag or attribute can have. In tracing, adding tags like user_id, request_id, or session_id to spans creates cardinality explosion—each unique combination of tag values produces a distinct trace or span entry, multiplying your storage requirements exponentially.

Imagine a library catalog. Organizing by genre (low cardinality: ~20 values) is manageable. But indexing by every reader's name who borrowed each book (high cardinality: thousands/millions) creates an unmanageable explosion of catalog entries.

Why It Matters

High-cardinality tags can:

  • Overwhelm storage backends with millions of unique span combinations
  • Slow queries as indexes grow massive
  • Increase costs dramatically (you're storing and querying far more data)
  • Degrade performance of trace collection pipelines

A single tag like user_id with 1M users on a service processing 100M requests/day creates astronomical storage needs.

Strategies to Limit Cardinality

  1. Avoid user-specific identifiers as span tags—use them as searchable attributes only when essential
  2. Sample more aggressively for high-cardinality scenarios (as covered in your sampling lessons)
  3. Hash or bucket values (e.g., user_bucket: hash(user_id) % 100 reduces 1M users to 100 buckets)
  4. Use span events for contextual data instead of creating new tag dimensions
  5. Limit tag sets to meaningful dimensions for debugging (service, endpoint, status_code)

Think dimensions, not identifiers—trace systems excel at showing patterns across services, not tracking individual users.

Key Takeaway:
High-cardinality tags like user IDs or request IDs cause exponential span volume growth; control cardinality by avoiding unique identifiers as dimensions, sampling strategically, and bucketing values when necessary.