Observability

Built-In Telemetry. Every Layer.

Metrics, tracing, and logging across your entire data stack. Export to Datadog or Prometheus with no agents to deploy and no code to change.

Built for the Hot Path

Traditional metrics libraries add microseconds of latency per operation. Eden's fast-telemetry engine adds nanoseconds.

~2ns
Increment Latency

80x faster than standard tooling via thread-sharded counters

0
Allocations on Hot Path

All metric primitives are pre-allocated and cache-line padded

13
Histogram Buckets

10μs to 10s latency tiers for per-command tracking

500
Ring Buffer Samples

Per-thread bounded distribution buffers for DogStatsD export

Metric Primitives

Four core metric types purpose-built for proxy workloads. Thread-sharded, lock-free, and cache-padded — each designed to never contend on the critical path.

Counter

~2ns increment

Thread-sharded CachePadded<AtomicIsize> with power-of-two bitmask indexing. Supports inc(), add(), sum(), and swap() for delta export.

Gauge

Atomic set/get

CachePadded<AtomicI64> for integer values and AtomicU64 bit-representation for f64. Set/get only — no inc/dec to prevent misuse.

Histogram

O(1) bucket lookup

Fixed-bucket with thread-sharded Counter per bucket. Default latency buckets from 10μs to 10s across 13 tiers.

Distribution

Lock-free record

Thread-local bounded ring buffers (500 samples default) for DogStatsD distribution export. Lock-free recording with drain(), count(), sum(), min(), max(), mean().

Label Support

Static Labels (Enum-Indexed)

O(1) array-based lookup via the LabelEnum trait. Counters, gauges, and histograms all support labeled variants with zero overhead per metric access.

Dynamic Labels (Runtime)

Runtime-labeled metrics with canonicalized BTreeMap label sets. Automatic stale series eviction with configurable cycle thresholds to prevent unbounded cardinality.

Export & Distributed Tracing

Plug into your existing observability stack. No proprietary formats, no vendor lock-in.

DogStatsD Export

UDP-based export at configurable intervals (default 10s). Max packet size of 8000 bytes with multi-value distribution batching and service tag support.

  • Delta temporality with swap()-based counter export
  • Multi-value distribution batching for efficient transfer
  • Service name and custom tag injection
  • FastMetricsExportState for stale series eviction (default 6 cycles)

Prometheus Export

Full Prometheus text exposition format with HELP and TYPE headers. Cumulative temporality for seamless scraping by Prometheus, Grafana Agent, or any compatible collector.

  • Standard text exposition with HELP/TYPE metadata
  • Cumulative temporality for standard Prometheus scraping
  • Histogram bucket boundaries with _le labels
  • Label-aware output with proper escaping and ordering

OpenTelemetry Tracing

Distributed tracing with automatic context propagation across every proxy hop. The #[with_telemetry] proc macro instruments functions with zero boilerplate.

  • Automatic span creation and context propagation
  • MetricsMiddleware for Actix-web HTTP routes
  • OTLP-compatible collector export
  • Trace correlation with proxy metrics

Structured Logging

Two-tier filtering system with compile-time feature flags and runtime level control via EDEN_LOG_LEVEL. Audience classification routes logs to the right destination.

  • Compile-time feature flags to eliminate dead code paths
  • Runtime EDEN_LOG_LEVEL for dynamic verbosity control
  • Audience classification: Internal, Client, Both
  • Structured key-value fields for machine parsing

Eight Metrics Categories

Complete visibility across every layer of the Eden stack. From gateway lifecycle events to per-command proxy latency to migration dual-write health.

Eden Core

Gateway lifecycle, config reloads, uptime

LLM

Token usage, provider latency, streaming metrics

IAM

Auth attempts, RBAC evaluations, JWT validation

Endpoint

Connection state, health checks, pool utilization

Proxy

Per-command latency, throughput, error rates

Wire / Polling

Protocol-level bytes, connection counts, metadata collection

Workload

Request patterns, hot key tracking, fanout metrics

Migration

Dual-write latency, replay queue depth, divergence counts

Proxy Target
p99 < 100μs

Added latency per proxied operation, inclusive of all telemetry collection

Full Observability, Zero Overhead

See every query, command, and request across your entire data stack, without slowing it down.