The Missing Pillar in APAC Observability
APAC engineering teams that have implemented metrics (Prometheus), logs (Loki/ELK), and traces (Jaeger/Tempo) still have a blind spot when diagnosing performance problems: they know that an APAC service is consuming 80% CPU but not which function is responsible. Continuous profiling fills this gap — the fourth observability pillar that connects high-level resource signals to specific code paths.
Without continuous profiling, APAC performance investigations require manually triggering profilers during incidents — often too late, as the condition has already resolved — or running profiling in staging environments that do not replicate APAC production load patterns.
Three tools cover the APAC continuous profiling spectrum:
Grafana Pyroscope — always-on profiling integrated with Grafana LGTM stack for correlated APAC observability.
Parca — self-hosted eBPF continuous profiling with Prometheus-native label model and diff flamegraphs.
Speedscope — browser-based flamegraph visualizer for on-demand APAC profile investigation and team sharing.
APAC Continuous Profiling Fundamentals
The four APAC observability pillars
APAC Observability Stack (complete):
Pillar 1: METRICS (Prometheus + Grafana)
Question answered: "Is my APAC service unhealthy?"
Signal: CPU 84%, memory 6.2GB, HTTP error rate 0.3%
Limitation: tells you WHAT is high, not WHY
Pillar 2: LOGS (Loki / ELK)
Question answered: "What happened in my APAC service?"
Signal: error stacktrace at 14:32:41 SGT
Limitation: captures events, not resource consumption
Pillar 3: TRACES (Tempo / Jaeger)
Question answered: "Where is my APAC request slow?"
Signal: /api/orders spans: DB 847ms, cache 3ms, auth 12ms
Limitation: attributes latency to service, not code function
Pillar 4: PROFILES (Pyroscope / Parca) ← the gap
Question answered: "WHICH FUNCTION is consuming APAC CPU?"
Signal: flamegraph shows apacOrderService.buildReport() = 67% CPU
Closes the loop: metrics identify WHAT, profiles explain WHY
APAC flamegraph interpretation basics
APAC CPU Flamegraph (reading bottom to top):
main() ──────────────────────────────────────────── 100%
└── http.HandleFunc() ──────────────────────────── 98%
└── apacOrdersHandler() ──────────────────────── 96%
├── apacDBQuery() ────────────────────────── 62% ← 62% of CPU here
│ ├── buildApacSQL() ───────────────────── 8%
│ └── json.Marshal() ──────────────────── 54% ← WHY is SQL slow?
│ └── reflect.ValueOf() ────────────── 54% ← reflection in hot path
└── apacCacheCheck() ─────────────────────── 34%
└── redis.Get() ──────────────────────── 34%
APAC Analysis:
→ json.Marshal using reflection = slow for high-volume APAC DB rows
→ Fix: pre-generate JSON or use struct tags with faster APAC JSON library
→ Expected improvement: 54% CPU reduction for apacOrdersHandler()
Grafana Pyroscope: APAC Always-On Flamegraphs
Pyroscope APAC Go service integration
// APAC: Instrument Go service with Pyroscope SDK
import (
"github.com/grafana/pyroscope-go"
)
func apacInitProfiling() {
pyroscope.Start(pyroscope.Config{
ApplicationName: "apac-orders-service",
// APAC: Pyroscope server (self-hosted or Grafana Cloud)
ServerAddress: "http://apac-pyroscope:4040",
// APAC: Labels for Prometheus-compatible querying
Tags: map[string]string{
"apac_env": "production",
"apac_region": "sg",
"apac_version": os.Getenv("APP_VERSION"),
"apac_pod": os.Getenv("POD_NAME"),
},
// APAC: Profile types to collect
ProfileTypes: []pyroscope.ProfileType{
pyroscope.ProfileCPU,
pyroscope.ProfileAllocObjects,
pyroscope.ProfileAllocSpace,
pyroscope.ProfileInuseObjects,
pyroscope.ProfileInuseSpace,
pyroscope.ProfileGoroutines,
},
})
}
Pyroscope APAC Kubernetes deployment
# APAC: values.yaml for Pyroscope Helm chart
# helm install pyroscope grafana/pyroscope -f values.yaml
pyroscope:
replicaCount: 2
# APAC: Storage for profile data
persistence:
enabled: true
size: 50Gi
storageClass: "apac-standard"
# APAC: Retention for profile data
config: |
storage:
backend: filesystem
filesystem:
dir: /data/pyroscope
# APAC: Grafana data source integration
grafana-agent:
enabled: true
# APAC: eBPF profiler (profiles all pods without SDK instrumentation)
pyroscope-ebpf:
enabled: true
# Profiles every process on every APAC K8s node automatically
# No application code changes required
Pyroscope APAC Grafana correlation
APAC Grafana dashboard: correlate CPU spike with profile
14:32 SGT: Prometheus alert: apac-orders-service CPU > 80% for 5 min
APAC Investigation in Grafana:
1. Open Grafana Explore → switch to Pyroscope data source
2. Query: {apac_env="production", apac_region="sg"}
→ Renders flamegraph for the 14:32-14:37 SGT window
3. Switch to "Explore Metrics" split view:
→ Left pane: Pyroscope flamegraph (WHICH function)
→ Right pane: Prometheus CPU metric (HOW HIGH)
4. Flamegraph reveals: apacReportBuilder.generatePDF() = 73% CPU
→ APAC root cause: new report feature released 14:28 SGT
→ APAC action: add async queue for report generation
5. Time to root cause: 4 minutes (vs 45 min without profiles)
Parca: APAC Self-Hosted eBPF Profiling
Parca APAC Kubernetes deployment
# APAC: Deploy Parca server and agent via kubectl
# Parca server (stores and serves APAC profiles)
kubectl apply -f https://github.com/parca-dev/parca/releases/latest/download/kubernetes-manifest.yaml
# Parca Agent (eBPF profiler on each APAC node)
kubectl apply -f https://github.com/parca-dev/parca-agent/releases/latest/download/kubernetes-manifest.yaml
# APAC: Parca Agent automatically profiles ALL processes on each node
# No application changes needed for APAC services
# Profiles annotated with Kubernetes labels (namespace, pod, container)
Parca APAC diff flamegraph — deployment regression
APAC Parca Diff Flamegraph — Before vs After Release 4.2.1
Selection A: 14:00-14:28 SGT (before release 4.2.1)
Selection B: 14:30-15:00 SGT (after release 4.2.1)
APAC Diff visualization (red = more CPU, green = less):
Red (worse after release):
apacProductService.listProducts() ──── +34% CPU
└── apacPriceEngine.calculate() ──── +34% CPU
└── apacTaxRule.applyRules() ─── +34% CPU ← new tax rules in release
Green (better after release):
apacCacheService.get() ──── -12% CPU
(APAC cache hit rate improved in release 4.2.1)
APAC Conclusion:
→ New APAC tax calculation rules (release 4.2.1) added 34% CPU per product list
→ Unintended APAC performance regression from tax rule expansion
→ Fix: memoize APAC tax calculations per product category (not per product)
Speedscope: APAC Collaborative Profile Visualization
Speedscope APAC profile sharing workflow
# APAC: Capture pprof from Go service and share via Speedscope
# Step 1: APAC — Capture CPU profile from production service
# (safe for production — 30-second CPU profile at low overhead)
curl -s "http://apac-orders-service:6060/debug/pprof/profile?seconds=30" \
-o /tmp/apac-orders-cpu-$(date +%Y%m%d-%H%M).prof
# Step 2: APAC — Open in browser (local visualization, no upload)
# drag-and-drop .prof file to https://speedscope.app
# OR use self-hosted speedscope for sensitive APAC profiles
# Alternative: open local file directly
open https://speedscope.app # then drag .prof file
# Step 3: APAC — Share with team
# speedscope.app shows URL-encoded profile in URL fragment
# APAC team members open same URL — see identical flamegraph
# No server access or profiling infrastructure needed
# APAC: Speedscope view modes
# ① Time Order: left-to-right = chronological APAC call order
# ② Left Heavy: tallest left bars = most APAC CPU time cumulative
# ③ Sandwich: shows callers+callees of selected APAC function
Speedscope APAC Python profile investigation
# APAC: Profile Python service and visualize in Speedscope
import cProfile
import pstats
import io
# APAC: Profile specific function during incident
apac_profiler = cProfile.Profile()
apac_profiler.enable()
# APAC: Run the slow function
apac_generate_report(apac_customer_id, apac_date_range)
apac_profiler.disable()
# APAC: Export in pstats format for Speedscope
apac_stats = pstats.Stats(apac_profiler)
apac_stats.dump_stats('/tmp/apac-report-profile.prof')
# Open /tmp/apac-report-profile.prof in speedscope.app
# → Reveals: apacPDFRenderer.renderTable() = 78% of total time
# → APAC fix: pre-paginate table data before PDF rendering
APAC Profiling Tool Selection
APAC Profiling Need → Tool → Why
APAC always-on production profiling → Pyroscope Grafana LGTM native;
(Grafana stack, continuous) → eBPF + SDK support;
APAC correlated view
APAC self-hosted Prometheus native → Parca eBPF no-instrument;
(Prometheus labels, diff flamegraphs) → CNCF governed;
APAC diff view
APAC incident investigation → Speedscope No infra required;
(on-demand, share with APAC team) → URL sharing;
zero APAC cost
APAC Java / JVM profiling → async-profiler Low overhead;
(OpenJDK, GraalVM APAC services) → JVM-specific; APAC
allocation profiling
APAC Python profiling → py-spy Zero-overhead; APAC
(Django, FastAPI, ML inference) → no code changes;
speedscope compatible
Related APAC Observability Resources
For the AIOps and observability platforms (Grafana, Datadog, Prometheus, Dynatrace) that continuous profiling data integrates with for correlated APAC root cause analysis, see the APAC AIOps and observability guide.
For the tracing tools (Jaeger, Tempo, OpenTelemetry) that capture APAC request latency at the span level — complementary to profiling's function-level CPU analysis, see the APAC GitOps and IaC observability guide.
For the SLO management tools (Pyrra, Sloth, OpenSLO) that define the APAC service-level objectives that continuous profiling helps maintain by catching regressions before SLO burn rates spike, see the APAC SLO management guide.
Beyond this insight
Cross-reference our practice depth.
If this article matches your stage of thinking, the underlying capabilities ship across all six pillars, ten verticals, and nine Asian markets.
Other service pillars
All industries
Asian markets