K-reflex Logo

See What Your Kernel Sees.

Next-gen eBPF Observability with < 1% overhead.

Production Ready 100%v1.2.0eBPF 5.8+Go 1.23

Why K-Reflex?

The Problem

Your application says "Network Slow", but your network team says "I'm fine". Standard observability tools like TCPDump and APM miss the invisible culprit: Kernel Write Latency.

When your application calls send(), the data doesn't magically appear on the wire. It first sits in the kernel's TCP send buffer. If that buffer is full or the kernel is under pressure, your application blocks—even though the network itself is healthy.

The Solution

K-Reflex uses eBPF to monitor the kernel's network stack directly, detecting these "invisible stalls" that happen between your application and the network. It bridges the observability gap by exposing what happens inside the kernel buffer.

Performance Metrics

< 5%
CPU Usage
Even at 1Gbps+
~13MB
Memory
Bounded & Predictable
5K/sec
Events
Rate Limited
Connections
Millions/sec

Zero Copy

Perf event arrays enable direct kernel-to-userspace data transfer without memory copies. Events flow seamlessly via ring buffers.

Safe Execution

Production-grade safety: LRU maps prevent memory leaks, rate limiting caps CPU usage at 5,000 events/sec. Never crashes your node.

Real-time Metrics

Prometheus-native metrics with intelligent label sanitization. Low-cardinality by default, high-cardinality optional.

Architecture

User Space

Go Controller
Event processing, label sanitization, Prometheus export
Ring Buffer

Kernel Space

eBPF Hooks
kprobe/tcp_v4_connect
kretprobe/tcp_sendmsg
Maps
LRU_HASH (connect_reqs)
ARRAY (rate_limit)

Live Demo

kreflex@kernel

Prometheus Metrics

# Low-Cardinality Metrics (Always Enabled)
kreflex_connection_latency_seconds_low_card
Labels: comm, dst_ip_net, dst_port_range
kreflex_write_stall_seconds_low_card
Labels: comm
# Observability Metrics
kreflex_lost_samples_total
kreflex_rate_limit_drops_total
kreflex_map_update_failures_total
kreflex_panic_recoveries_total

Deep Kernel Tracing

Kprobes hook directly into kernel functions like tcp_v4_connect() and tcp_sendmsg(), providing function-level observability.

Efficient Data Sharing

LRU_HASH and ARRAY maps enable efficient kernel-user space communication with automatic memory eviction and bounded growth.

Stall Detection

Detects kernel write stalls >100µs, exposing invisible bottlenecks that occur between your application and the network stack.

Production Ready

Tested at millions of connections/second. CPU usage stays below 5% even under extreme load. Memory bounded to predictable limits.