Building and distributing outside this repository

• I understand I can use and distribute my component outside of this repository.

Existing component implementation

https://github.com/grafana/opentelemetry-collector-extras

Components covering similar use cases

No direct equivalent exists for aggregating similar leaf spans within a trace while preserving statistical information.

The purpose and use-cases of the component

Problem Statement

Modern distributed applications often generate traces with many repetitive leaf spans—N+1 database queries, batch HTTP calls, fan-out operations. These patterns create:

• High storage costs from redundant span data
• Noisy trace views that obscure the actual request flow
• Cardinality pressure on backends

Solution

The spanpruningprocessor identifies similar leaf spans within a trace and replaces each group with a single summary span containing aggregated statistics. This reduces trace volume while preserving meaningful operational information.

Key Features

Use Cases

1. Database query optimization - Aggregate N+1 query patterns into a single summary showing query count and latency distribution
2. Batch operations - Consolidate fan-out calls (parallel HTTP requests, message publishes) into representative summaries
3. Cost reduction - Reduce trace storage costs for repetitive workloads while retaining operational insight
4. Debugging support - Preserve outlier spans with correlated attributes to identify root causes of latency issues

Example

Before (10 similar SELECT spans):

handler
├── SELECT - 5ms
├── SELECT - 6ms
├── SELECT - 7ms
├── SELECT - 500ms (outlier, cache_hit=false)
└── ... (6 more)

After (with outlier preservation):

handler
├── SELECT (summary, span_count=9)
│   ├── aggregation.duration_avg_ns: 8000000
│   ├── aggregation.duration_median_ns: 7000000
│   ├── aggregation.outlier_correlated_attributes: "cache_hit=false(100%/0%)"
│   └── aggregation.histogram_bucket_counts: [3, 6, 8, 9]
└── SELECT - 500ms (preserved outlier)
    ├── aggregation.is_preserved_outlier: true
    └── cache_hit: false

Example configuration for the component

processors:
  spanpruning:
    # Attribute patterns for grouping (glob syntax)
    group_by_attributes:
      - "db.*"
      - "http.method"
    
    # Minimum spans required before aggregation (default: 5)
    min_spans_to_aggregate: 5
    
    # Parent aggregation depth: 0=none, -1=unlimited (default: 1)
    max_parent_depth: 1
    
    # Prefix for aggregation attributes (default: "aggregation.")
    aggregation_attribute_prefix: "aggregation."
    
    # Histogram bucket bounds (default: 5ms to 10s)
    aggregation_histogram_buckets: [5ms, 10ms, 25ms, 50ms, 100ms, 250ms, 500ms, 1s, 2.5s, 5s, 10s]
    
    # Enable attribute loss analysis (default: false)
    enable_attribute_loss_analysis: false
    
    # Enable outlier detection (default: false)
    enable_outlier_analysis: false
    
    # Outlier analysis settings
    outlier_analysis:
      method: iqr                    # or "mad"
      iqr_multiplier: 1.5
      mad_multiplier: 3.0
      min_group_size: 7
      correlation_min_occurrence: 0.75
      correlation_max_normal_occurrence: 0.25
      max_correlated_attributes: 5
      preserve_outliers: false
      max_preserved_outliers: 2
      preserve_only_with_correlation: false

Telemetry data types supported

• Traces (alpha stability)

Code Owners

@portertech, @csmarchbanks, @sherinabr

Sponsor

@jmacd

Additional context

• Pull Request: Add Trace Span Pruning Processor #45617
• Test Coverage: >80% with unit tests, benchmarks, and integration tests
• Documentation: Comprehensive README with configuration reference and examples

Tip

_{React with 👍 to help prioritize this issue. Please use comments to provide useful context, avoiding +1 or me too, to help us triage it. Learn more here.}