Offline Cross Profiler May 2026

"CPU usage is 50%. I don't see the problem."

type Event struct Ts uint64 PID uint32 Domain string // "cpu", "mem", "io" Key string // "cycles", "page_fault", "write_bytes" Value uint64 Stack []uintptr

type RingRecorder struct buffer []Event head atomic.Uint64 fd *os.File compressor *snappy.Writer offline cross profiler

func (r *RingRecorder) Record(domain string, key string, value uint64) idx := r.head.Add(1) % uint64(len(r.buffer)) e := &r.buffer[idx] e.Ts = uint64(time.Now().UnixNano()) e.Domain = domain e.Key = key e.Value = value // Capture stack (omit for brevity)

"timestamp_ns": 1732145678123456789, "core_id": 3, "pid_tid": "1337/1402", "domain": "CPU", "metric": "cycles_stalled_backend", "value": 42000, "stack_id": "0x7fffdeadbeef" "CPU usage is 50%

Running on a developer’s workstation (not the target), the analyzer loads the 10GB trace file, builds indexes on timestamp_ns and stack_id , and executes cross-domain queries. 3. The "Cross" Advantage: Correlating Domains The magic of the offline cross-profiler is not in a single flame graph, but in the intersection of graphs. Consider a real-world bug: High latency in a database.

if idx%1024 == 0 // Flush every 1024 events r.flush() The "Cross" Advantage: Correlating Domains The magic of

func (r *RingRecorder) flush() // Snappy compress and write to offline file // No analysis happens here.