Chapter 17: Monitoring & Observability

Monitoring tells you when something is wrong. Observability tells you why. The difference is whether your system was built to be questioned.

---

Mind Map

---

The Observability Signals

Historically framed as "three pillars," observability in 2026 is better understood as four signals: metrics, logs, traces, and the emerging fourth signal — continuous profiling. All four are standardized under OpenTelemetry (OTel), which is now the de facto instrumentation layer across the industry.

Dimension	Metrics	Logs	Traces	Profiling
What is it?	Numeric measurements over time	Timestamped text/structured records	Request journey across services	Continuous CPU/memory/allocation flame graphs
Granularity	Aggregated (not per-request)	Per-event	Per-request, cross-service	Per-process, sampled continuously
Best for	Dashboards, alerting, trends	Debugging specific events	Latency attribution, dependency mapping	Finding hot code paths; unexplained CPU/memory cost
When to use	"Is something wrong?"	"What happened at 14:03:22?"	"Which service added 800ms?"	"Which function is burning CPU at 4 AM?"
Cost	Low (aggregates)	Medium (storage scales with volume)	High (sampling required at scale)	Low–medium (eBPF-based has <1% CPU overhead)
Common tools	Prometheus, Datadog, Atlas	ELK Stack, Loki, CloudWatch	Jaeger, Zipkin, Grafana Tempo	Parca, Pyroscope, Grafana Profiles
OTel status	GA (stable)	GA (stable)	GA (stable)	RC (targeting GA Q3 2026)

You need all four. Metrics tell you latency spiked; logs tell you which user was affected; traces tell you which service in the call chain caused it; profiling tells you which function inside that service is burning CPU or allocating memory unexpectedly.

Profiling as the 4th Signal (as of 2026)

OpenTelemetry's profiling signal entered Release Candidate in Q1 2026, targeting GA by Q3 2026. It uses eBPF-based continuous CPU/memory sampling and Linux perf format, exportable via OTLP. Production tools like Grafana Pyroscope and Parca can already receive OTel profiling data. Frame it as "emerging GA" — not experimental, not yet fully standardized.

---

Metrics Types

Prometheus and most metrics systems define three fundamental types:

Type	Definition	Example Use Cases	Example Values
Counter	Monotonically increasing integer, only goes up (resets on restart)	Total HTTP requests, total errors, total bytes sent	http_requests_total{method="GET", status="200"} 10482
Gauge	Arbitrary value that can go up or down	CPU %, active connections, queue depth, memory usage	queue_depth{queue="payments"} 42
Histogram	Samples observations into configurable buckets, exposes _count, _sum, _bucket	Request latency distribution, request size	http_duration_seconds_bucket{le="0.1"} 8234

Why Histograms Matter: Percentiles vs. Averages

Averages hide tail latency. A p99 latency of 2s means 1% of your users wait 2 seconds — that can be thousands of users. Always alert on percentiles (p95, p99, p999) for latency metrics.

p50 = 50ms  ← Typical user
p95 = 200ms ← Most users
p99 = 800ms ← Tail (worst 1%)

Prometheus computes percentiles from histograms server-side using histogram_quantile(0.99, ...).

---

Distributed Tracing

In a microservices system, a single user request may fan out to 10+ services. Traditional logging — per-service — cannot answer "which service is slow." Distributed tracing reconstructs the full call tree.

Core Concepts

• Trace: The complete journey of one request, from entry point to all leaf calls
• Span: One unit of work within a trace (e.g., one service call, one DB query)
• Correlation ID / Trace ID: A unique ID injected at the entry point and propagated via HTTP headers (X-Trace-ID) to every downstream call
• Parent-child spans: Each span records its parent span ID, enabling tree reconstruction

Trace Sequence Example

Span Tree Visualization

abc123 [120ms total]
├── span_001: Auth (User Service)  [15ms]
└── span_002: Order Service        [98ms]
    └── span_003: DB Query         [45ms]

This immediately surfaces that the DB query and Order Service overhead account for 98ms, making optimization target obvious.

Sampling Strategies

At 10,000 req/sec, storing every trace is prohibitively expensive. Two approaches:

Strategy	How It Works	Pros	Cons
Head-based	Decide at trace start (random %)	Low overhead, simple	Misses rare errors
Tail-based	Buffer full trace, decide after completion based on outcome	Captures all errors/slow requests	Higher memory/processing cost

Production recommendation: sample 1% normally, 100% of errors and traces > p99 latency threshold.

W3C TraceContext: Trace Propagation Standard

Distributed tracing requires trace and span IDs to flow across service boundaries via HTTP headers. W3C TraceContext (RFC 2019) defines the standard:

traceparent: 00-4bf92f3577b34da6a3ce929d0e0e4736-00f067aa0ba902b7-01
             ^^  ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ ^^^^^^^^^^^^^^^^ ^^
          version       trace-id (128-bit)         parent-span-id   flags

Every service reads this header, creates a child span with the parent-span-id, and propagates the same trace-id to downstream calls. No proprietary vendor format needed.

Tracing Tools Comparison

	Jaeger	Zipkin	Grafana Tempo
Storage backends	Cassandra, Elasticsearch, in-memory	MySQL, Cassandra, Elasticsearch	Object storage (S3, GCS, Azure Blob)
UI	Full-featured: trace timeline, service dependency graph	Simpler: trace list + timeline	Minimal native UI; relies on Grafana
Sampling	Head-based and adaptive (remote controlled)	Head-based	Delegated to OTel Collector
OpenTelemetry	Native OTLP support	Via adapter	Native OTLP (primary protocol)
Integration	CNCF project, Kubernetes-native	Widely adopted, many client libs	Grafana stack (pairs with Prometheus + Loki)
Scale	Medium-large	Medium	Very large (object storage = cheap at scale)
Deployment	Moderate complexity	Simple	Simple (no search index needed)
Cost	Free (Elasticsearch costs extra)	Free	Free (storage cost only)
Best for	Kubernetes microservices, CNCF stack	Existing Zipkin instrumentation	Grafana-centric observability stacks

---

OpenTelemetry

OpenTelemetry (OTel) is a CNCF project that provides a vendor-neutral standard for instrumentation, collection, and export of telemetry data — traces, metrics, and logs — under a single unified API and SDK.

Before OTel, each observability vendor required its own SDK. Switching from Jaeger to Datadog meant re-instrumenting every service. OpenTelemetry solves this with a single instrumentation layer and pluggable exporters.

Three Pillars Unified

OTLP (OpenTelemetry Protocol) is the wire format. It runs over gRPC (port 4317) or HTTP/protobuf (port 4318). Any backend that speaks OTLP can receive OTel data.

OTel Collector Architecture

The Collector is an optional but recommended middle tier. It decouples instrumented apps from backend destinations:

Collector benefits:

• Tail-based sampling at the Collector level (buffer spans, sample on outcome)
• Data transformation: filter PII from logs, rename labels, add resource attributes
• Multi-backend export: send same data to two backends simultaneously (migration, redundancy)
• Decoupled upgrades: swap backends without touching application code

Auto-Instrumentation vs Manual Instrumentation

Dimension	Auto-Instrumentation	Manual Instrumentation
How	Agent/bytecode injection at startup, no code changes	Developer adds tracer.start_span() calls in code
Effort	Zero code changes	Per-operation instrumentation required
Coverage	HTTP clients, DB drivers, frameworks (Spring, Express, Django)	Any custom business logic, critical paths
Span quality	Generic (framework-level, missing business context)	Rich (custom attributes: user_id, order_id, feature flags)
Latency	Slight overhead from agent	Minimal (only instrumented operations)
Maintenance	Agent version updates	Code changes when logic changes
Best for	Getting started, standardizing infrastructure calls	High-value business operations, SLO-critical paths

Recommendation: Use auto-instrumentation as the baseline (catches all HTTP and DB calls), then add manual spans around critical business operations (payment processing, fraud check, recommendation engine) where custom attributes are needed for debugging.

OTel Language Support

OTel provides SDKs for all major languages: Java, Python, Go, JavaScript/Node, .NET, Ruby, Rust, C++, PHP. Auto-instrumentation is most mature for Java (via Java agent) and Node.js (via @opentelemetry/auto-instrumentations-node).

Cross-reference: Chapter 16: Reliability Patterns for SLO-driven reliability engineering. Chapter 23: Cloud-Native Architecture for OTel in Kubernetes.

---

eBPF-Powered Observability (2026)

Extended Berkeley Packet Filter (eBPF) allows safe, sandboxed programs to run inside the Linux kernel without modifying kernel source or loading kernel modules. In observability, this unlocks something previously impossible: full application telemetry — function calls, network flows, file I/O, system calls — with near-zero overhead and zero application code changes.

As of 2026, CNCF survey data shows 67% of teams running Kubernetes at scale have adopted at least one eBPF observability tool, with 300% year-on-year growth. eBPF is no longer an advanced topic — it is the practical answer to the "how do I instrument a polyglot service mesh without touching every service's code?" problem.

Why eBPF Changes the Observability Model

Traditional observability requires either:

• SDK instrumentation: developers add tracing/metrics code to each service (coupled, language-specific)
• Sidecar proxies: Envoy/Linkerd sidecars intercept traffic (works for L4/L7 but adds 50–100ms cold overhead and memory per pod)

eBPF operates at the kernel level, intercepting events from any process on the node — regardless of language, framework, or whether a sidecar is present. Key advantages:

Property	Traditional Agent	eBPF-Based
CPU overhead	5–15%	< 1%
Code changes required	Yes (SDK)	None
Language-agnostic	No	Yes
Kernel/syscall visibility	No	Yes (security events, file I/O)
Precision	Millisecond	Microsecond
Works with sidecars	Complements	Can replace for L4 visibility

The 2026 eBPF Observability Stack

Key tools in the stack (as of 2026):

• Cilium — eBPF-native Container Network Interface (CNI). Provides L3–L7 network policy, service mesh connectivity, and full network flow visibility via its companion UI Hubble. When combined with Istio ambient mode (see Ch13), Cilium handles the data-plane while Istio handles control-plane policy — replacing per-pod sidecar proxies entirely.
• Tetragon — security-focused eBPF tool from Isovalent (Cilium's creators). Captures process execution, file access, and network connections at the syscall level. Useful for runtime threat detection without a dedicated security agent.
• Pixie — auto-instrumented APM (Application Performance Monitoring) from New Relic. Captures HTTP/1.1, HTTP/2, gRPC, MySQL, PostgreSQL, Redis traffic — with request/response bodies — without any SDK. Run px run px/http_data and get per-endpoint latency histograms in seconds.
• Grafana Beyla — generates OTel-compatible spans from eBPF-intercepted HTTP, gRPC, and SQL calls. Designed to feed the OTel Collector directly, plugging eBPF-sourced traces into a standard Grafana/Tempo/Jaeger stack with zero application modification.

When to Adopt eBPF Observability

Situation	Recommendation
Polyglot microservices (Go, Java, Python, Rust mix)	Strong fit — no per-language SDK needed
Kubernetes-native deployments	Strong fit — node-level eBPF agents deploy as DaemonSets
Security compliance requiring syscall audit	Use Tetragon for runtime enforcement
Replacing sidecar service mesh overhead	Cilium eBPF + ambient Istio reduces memory 90%, latency 25%
Legacy monolith on bare metal (non-Kubernetes)	Limited — eBPF tools are Kubernetes-optimized
Teams already invested in OTel SDK instrumentation	Complement with Beyla/Pixie; don't replace rich manual spans

eBPF Observability and OTel are Complementary

eBPF tools like Beyla and Pixie export OTel-compatible telemetry — they feed into the same OTel Collector pipeline as SDK-based instrumentation. The pattern for 2026: use eBPF for infrastructure-level and network-level visibility (zero effort), use OTel SDKs for business-logic-level spans (high-value custom attributes like order_id, user_tier). Both flow into the same Grafana/Jaeger/Tempo backend.

---

Log Aggregation Pipeline

Individual service logs are useless if you cannot search them across all instances. A log aggregation pipeline centralizes logs from every container, VM, and serverless function:

Structured logging (JSON over plaintext) is essential for the search step. A structured log line:

{
  "timestamp": "2026-03-12T00:40:00Z",
  "level": "ERROR",
  "service": "order-service",
  "trace_id": "abc123",
  "user_id": "user_789",
  "message": "Payment timeout after 5000ms",
  "duration_ms": 5000
}

This allows queries like: level:ERROR AND service:order-service AND duration_ms:>3000.

---

SLI / SLO / SLA

Google's Site Reliability Engineering introduced the SLI → SLO → SLA hierarchy as a way to make reliability quantitative and contractual:

Definitions

Concept	Owner	Consequence of breach	Example
SLI	Engineering	None — it is a measurement	99.92% of requests succeeded this month
SLO	Engineering	Internal alert, error budget consumed	Target: 99.9% success rate
SLA	Business/Legal	Financial penalty, contract clause	Guarantee: 99.5% or credit issued

SLO is always stricter than SLA. The gap between SLO (internal target) and SLA (contractual guarantee) is the safety buffer — if engineering hits 99.8% and the SLO was 99.9%, the team is alerted and investigates before breaching the 99.5% SLA.

Common SLI Examples by Metric Type

Metric Type	Example SLI	Example SLO	Measurement Method
Availability	Fraction of successful HTTP requests (2xx/3xx)	99.9% success rate over 30 days	Synthetic probes + real traffic
Latency	p99 response time	p99 < 500ms over 1-hour window	Histogram (Prometheus histogram_quantile)
Error rate	Fraction of 5xx responses	< 0.1% error rate	Error counter / total request counter
Throughput	Requests processed per second	> 1,000 RPS sustained	Gauge metric on queue consumer
Freshness	Age of most recent data ingested	Data lag < 5 minutes	Timestamp comparison metric
Durability	Fraction of written objects successfully retrieved	99.999999% (11 nines)	Periodic read-back verification

Error Budgets

An error budget is the allowable unreliability within an SLO period:

Error budget = 1 − SLO target
99.9% SLO → 0.1% budget → 43.8 minutes/month of allowed downtime
99.99% SLO → 0.01% budget → 4.38 minutes/month

The Nines: Downtime Allowance per SLO Target

SLO Target	Downtime per Month	Downtime per Year	Downtime per Week	Common Name
99%	7.3 hours	3.65 days	1.68 hours	Two nines
99.5%	3.65 hours	1.83 days	50.4 minutes	—
99.9%	43.8 minutes	8.77 hours	10.1 minutes	Three nines
99.95%	21.9 minutes	4.38 hours	5.04 minutes	—
99.99%	4.38 minutes	52.6 minutes	60.5 seconds	Four nines
99.999%	26.3 seconds	5.26 minutes	6.05 seconds	Five nines
99.9999%	2.63 seconds	31.6 seconds	0.605 seconds	Six nines

Cost of an additional nine: Each additional nine of availability roughly doubles infrastructure cost and operational complexity. Going from 99.9% to 99.99% is not a 10× improvement — it requires eliminating every planned maintenance window, active-active multi-region deployment, and sub-minute failover. Most SaaS products target 99.9% (43 min/month), which is achievable with a single region + good health checks. Five nines (26s/month) requires active-active multi-region with automated failover in under 10 seconds.

Error budget policy: When the budget is exhausted, new feature deployments halt and reliability work takes priority. This creates a natural feedback loop: engineering teams that want to ship features are incentivized to keep the service reliable.

Real-World — Google SRE: Google's SRE teams hold joint ownership of error budgets with product teams. If a service exhausts its error budget, the SRE team can unilaterally halt launches. This removes the "reliability vs. velocity" organizational conflict by making reliability a shared engineering metric.

---

Connecting SLOs to Observability Signals

SLOs require SLIs, which require the right observability signals:

Burn rate alerts: Rather than alerting when the budget is fully exhausted, alert when the consumption rate predicts exhaustion. A burn rate of 14.4× means you will exhaust a 30-day budget in 2 days. Alert at 2× burn rate (14-day warning) and 14.4× burn rate (2-hour warning). This is the multi-window, multi-burn-rate alerting pattern from Google's SRE Workbook.

---

Alerting Strategies

Alert Fatigue

The biggest failure mode in alerting is alert fatigue — too many low-signal alerts cause engineers to ignore them, including critical ones. Symptoms:

• On-call engineers acknowledge without investigating
• Alert volume exceeds 10/day on average
• Many alerts resolve without human action

Solution: Every alert must be actionable. If an alert fires and no action is required, delete or demote it.

Severity Levels

Severity	Definition	Response SLA	Example
P1 / Critical	Service down, revenue impact	Wake on-call immediately, < 5 min	Payment API returning 500
P2 / High	Degraded, SLO at risk	Alert on-call during business hours, < 30 min	p99 latency > 2s
P3 / Medium	Anomaly, no immediate user impact	Ticket, fix in sprint	Disk > 80% on non-critical host
P4 / Low	Informational	Review weekly	Dependency approaching end of support

On-Call Best Practices

• Rotate on-call weekly to distribute burden and knowledge
• Keep runbooks for every P1/P2 alert — reduce MTTR with documented steps
• Conduct blameless post-mortems within 48 hours of incidents
• Track Mean Time To Detect (MTTD) and Mean Time To Resolve (MTTR) as team metrics
• Alert on symptoms (user-visible impact) not causes (CPU high) where possible

---

Health Checks

Health checks allow orchestrators (Kubernetes, load balancers) to route traffic away from unhealthy instances automatically. Three probe types:

Probe	Question	Failure Action	Example Endpoint
Startup	Has the app finished initializing?	Wait (don't kill yet)	/health/startup — checks migrations complete
Liveness	Is the process alive and not deadlocked?	Restart the container	/health/live — returns 200 if process is responsive
Readiness	Can the app serve traffic right now?	Remove from LB pool	/health/ready — checks DB connection, cache connection

Cross-reference: Chapter 6 covers how load balancers use health checks to remove unhealthy backends from rotation.

Readiness check design: Be conservative. If your app cannot reach its database, it should fail readiness — sending traffic that will fail is worse than not sending traffic at all. However, a slow downstream service should not fail readiness if the app can degrade gracefully.

---

Tools Comparison

Tool	Category	What It Does	Deployment	Strengths	Weaknesses	Cost
Prometheus	Metrics	Pull-based metrics collection, storage, PromQL	Self-hosted	CNCF standard, powerful query language	No long-term storage built-in, cardinality limits	Free
Grafana	Visualization	Dashboards for metrics, logs, traces from many sources	Self-hosted / Cloud	Universal frontend, supports 50+ data sources	Requires data source backends	Free / Paid
Elasticsearch	Log storage	Distributed search and analytics engine	Self-hosted / Cloud	Full-text search, flexible schema	Resource-intensive, complex to operate	Free / Paid
Logstash	Log processing	ETL pipeline for logs — parse, filter, enrich	Self-hosted	Powerful filter plugins	Heavy JVM resource usage	Free
Jaeger	Tracing	Distributed trace collection, storage, UI	Self-hosted	CNCF, OpenTelemetry compatible	No metrics, no logs	Free
Datadog	All-in-one APM	Metrics + logs + traces + APM + alerting	SaaS	Low operational overhead, fast setup	Expensive at scale	Per-host pricing
New Relic	All-in-one APM	Full-stack observability, error tracking	SaaS	Good out-of-box instrumentation	Cost scales with data ingest	Per-GB ingest
AWS CloudWatch	Cloud-native	Metrics + logs for AWS resources	SaaS (AWS)	Zero setup for AWS services	Vendor lock-in, limited query capability	Per metric/log

Practical guidance:

• Startups: Datadog or New Relic for speed of setup
• Mid-size, cost-conscious: Prometheus + Grafana + ELK + Jaeger (more ops burden, much cheaper)
• AWS-native: CloudWatch + X-Ray + managed Prometheus/Grafana
• OpenTelemetry: Use the vendor-neutral OTLP standard for instrumentation — swap backends without re-instrumenting code

Real-World — Netflix Atlas: Netflix built Atlas, their internal metrics platform, to handle billions of time series from thousands of services. Atlas uses in-memory storage optimized for real-time dashboards and pattern-matching queries across tag dimensions. Netflix open-sourced Atlas; its design influenced Prometheus's label model.

---

Trade-offs & Comparisons

Decision	Option A	Option B	Recommendation
Metrics storage	Prometheus (self-hosted)	Datadog (SaaS)	SaaS if <$5K/month matters less than ops cost
Log sampling	Store all logs	Sample + retain errors	Sample at high volume (>10GB/day)
Trace sampling	Head-based (simple)	Tail-based (smart)	Tail-based if budget allows — captures all errors
SLO target	99.9% (43 min/month budget)	99.99% (4 min/month budget)	Higher SLO = higher infra cost, diminishing returns
Alert strategy	Alert on causes (high CPU)	Alert on symptoms (error rate)	Symptom-based reduces noise

---

Key Takeaway: Observability is the foundation of reliability. You cannot improve what you cannot measure, and you cannot debug what you cannot trace. Instrument before you need it — adding tracing during an incident is too late. The four signals (metrics, logs, traces, and the emerging profiling signal) are complements, not substitutes. In 2026, OpenTelemetry is the default instrumentation layer — use it from day one so backends are swappable. For Kubernetes deployments, pair OTel SDKs with eBPF tools (Pixie, Beyla, Cilium) to achieve full-stack visibility without per-service agent sprawl.

---

Incident Management Lifecycle

The Five Phases

Detection

• Automated alerts, not human discovery — if an engineer finds the issue before an alert fires, alerting has failed
• Alert on symptoms, not causes — alert on error rate or latency degradation, not "CPU > 80%" (which may be harmless)
• Multi-signal detection — a latency spike confirmed by both metrics and traces is higher confidence than a single signal; reduce false positives by requiring two or more signals to agree

Triage

Classify severity within minutes of detection to determine escalation path and response urgency:

Severity	Impact	Response Time	Example
P0 / SEV1	Total outage, data loss	Immediate, all hands	Payment system down
P1 / SEV2	Major feature broken	15 min	Login failing for 50% of users
P2 / SEV3	Minor feature degraded	1 hour	Search results slow
P3 / SEV4	Cosmetic / low impact	Next business day	Dashboard chart incorrect

Incident Example: Latency Spike Debugging

---

Designing Effective Dashboards

The Four Golden Signals (Google SRE)

Every service dashboard should lead with these four panels — they cover the majority of user-visible failure modes:

• Latency — how long requests take (show p50, p95, p99 — never just average)
• Traffic — how much demand the system is under (requests/sec, events/sec)
• Errors — rate of failed requests (5xx, timeouts, explicit errors)
• Saturation — how "full" the service is (CPU %, queue depth, connection pool usage)

If a service is degraded, at least one of these four will deviate from baseline. Start your dashboard design here; add domain-specific panels only as supplements.

Dashboard Anti-patterns

Anti-pattern	Problem	Fix
Too many panels	Information overload slows incident response	Max 8–10 panels per dashboard; link to drill-down dashboards
Only averages	Hides tail latency affecting real users	Always show p50, p95, p99 side by side
No baseline	Cannot tell if a value is normal or alarming	Add SLO threshold lines and historical comparison overlays
Wall of text	Slow to scan under pressure	Use time-series graphs and stat panels, not tables of raw numbers

---

Cost-Efficient Observability

The Cardinality Problem

High-cardinality labels cause metric storage to explode. Each unique label combination creates a separate time series:

1,000 unique user_id values
  × 100 metrics per user
  × 60-second resolution
  × 30-day retention
= millions of time series → storage and query cost balloons

Never use user_id, request_id, or session_id as Prometheus label values. Reserve labels for low-cardinality dimensions: service, method, status_code, region.

Strategies

• Sampling: Collect 1% of traces in normal production traffic; use tail-based sampling to capture 100% of error traces and traces exceeding the p99 latency threshold — you get full coverage where it matters at a fraction of the cost
• Aggregation: Pre-aggregate metrics at collection time in the OTel Collector (e.g., sum request counts by service) rather than storing raw per-request data
• Retention tiers: Hot storage (7 days, full resolution) → Warm storage (30 days, downsampled to 1-minute intervals) → Cold storage (1 year, aggregated to hourly) — most incidents are investigated within 7 days; cold data is for trend analysis only
• Log levels: Emit DEBUG logs only in development environments; INFO, WARN, and ERROR in production — a single verbose service can generate gigabytes of low-value log data per day

---

Related Chapters

Chapter	Relevance
Ch16 — Security & Reliability	Reliability SLOs and incident response complement observability
Ch13 — Microservices	Distributed tracing across microservice boundaries
Ch23 — Cloud-Native	Cloud-native monitoring: Prometheus, Grafana, CloudWatch

---

Practice Questions

Beginner

1. Distributed Tracing: A microservices request takes 3 seconds end-to-end, but each individual service logs less than 100ms of processing time. How would you use distributed tracing (spans, trace IDs) to locate the missing ~2.7 seconds? What are the most common hidden latency sources in microservice chains?

   Hint

   Spans capture wall-clock time including network hops and queue wait time that individual service logs don't measure — look for gaps between the end of one span and the start of the next child span.

Intermediate

2. Error Budget: Your team's SLO is 99.9% availability (43.8 min/month error budget). After a 2-hour outage, the SRE lead says you have "used 2.7× your monthly error budget in one incident." What does this mean operationally — what features or deployments must now be frozen, and for how long?

   Hint

   Burning the error budget triggers a freeze on non-critical feature releases until the budget resets (typically monthly); the team must focus entirely on reliability improvements before new features ship.

3. Readiness Probe Design: You are designing a readiness probe for a service that depends on PostgreSQL, Redis, and a third-party payment API. The payment API is sometimes slow (2–5s). How do you design the probe so a slow payment API does not remove your service from the load balancer rotation?

   Hint

   Separate critical dependencies (PostgreSQL, Redis — required for the service to function) from non-critical ones (payment API); probe only critical deps for readiness, and use a separate circuit breaker for the payment API.

4. Observability Stack Decision: Compare Prometheus + Grafana (self-hosted) vs Datadog (SaaS) for a team of 5 engineers running 50 microservices. What hidden costs on each side are rarely surfaced in vendor comparisons?

   Hint

   Prometheus hidden costs: storage sizing, alert manager maintenance, and engineering time managing the stack; Datadog hidden costs: per-host + per-custom-metric pricing that scales steeply with microservice count and cardinality.

Advanced

5. Alert Noise Reduction: Your on-call engineer receives 200 alerts per week: 180 auto-resolve in 10 minutes, 15 require investigation but no action, and 5 require actual fixes. Design an alert restructuring plan (severity tiers, grouping, inhibition rules) to reduce noise while ensuring no critical alert is missed.

   Hint

   Demote self-resolving alerts to warnings or eliminate them; add alert inhibition (suppress child alerts when a parent alert fires); use Alertmanager grouping to collapse 50 pod-restart alerts into one service-level alert — target a ratio where >80% of pages require action.

---

References & Further Reading

• "Site Reliability Engineering" (Google SRE Book) — Chapters on Monitoring Distributed Systems and Alerting: https://sre.google/sre-book/table-of-contents/
• "Observability Engineering" — Charity Majors, Liz Fong-Jones, George Miranda (O'Reilly, 2022) — the definitive guide to high-cardinality observability and the shift from monitoring to observability
• OpenTelemetry documentation — vendor-neutral instrumentation standard for traces, metrics, and logs: https://opentelemetry.io/docs/
• "The Art of Monitoring" — James Turnbull — practical guide to modern monitoring pipelines with Prometheus and the ELK stack
• Datadog blog: "The Four Golden Signals" — https://www.datadoghq.com/blog/monitoring-101-collecting-data/
• Google SRE Workbook — Chapter on Incident Response — covers structured incident management, severity classification, and postmortem culture: https://sre.google/workbook/incident-response/