Prometheus Up Running (Julien Pivotto, Brian Brazil) (Z-Library)

# Prometheus: Up & Running ## 【One-Line Pitch】 A practical, hands-on guide to monitoring with Prometheus—covering everything from basic setup and query language to service discovery, Kubernetes integration, and alerting—written for SREs, Kubernetes administrators, and developers who need reliable operational metrics. ## 【Book Arc】 - **Opening (~0%–10%)**: Introduces monitoring concepts, explains why Prometheus uses a metrics-based approach over logs, and walks through a first working setup with scrape jobs, rules, and alerting configuration. - **Early (~10%–23%)**: Dives into code instrumentation across Python, Java, and Go—covering counters, gauges, histograms, exposition formats, and the Pushgateway pattern for batch jobs. - **Early (~23%–32%)**: Explores PromQL query language fundamentals, including rate calculations, label matching, many-to-one joins with group_left, and building dashboards in Grafana. - **Middle (~32%–48%)**: Covers service discovery mechanisms (Consul, EC2, Kubernetes), target labeling hierarchies, and relabeling techniques to map metadata to useful labels. - **Middle (~48%–end)**: Addresses advanced topics including cAdvisor for container metrics, Kubernetes monitoring patterns, metric relabeling for label management, and production considerations. ## 【Key Takeaways】 - **Metrics over logs for operational monitoring** (Opening): Prometheus tracks aggregations over time rather than individual events, keeping resource usage sane with a ~10,000 metrics-per-process guideline. This makes it unsuitable for event logs or high-cardinality data like email addresses. - **Prometheus prioritizes availability over perfect accuracy** (Opening): The system accepts small inaccuracies from kernel scheduling and failed scrapes, preferring 99.9% correct data over breaking monitoring while waiting for perfect data—use caution in billing or money-related applications. - **Instrumentation is decoupled from exposition** (Early): You define metrics with client libraries and expose them on /metrics (or any path) independently—transient dependencies with Prometheus instrumentation appear automatically on your metrics page. - **Histograms enable average latency calculation** (Early): The _count and _sum time series from histogram metrics let you compute average latency with rate(sum[1m]) / rate(count[1m])—the per-second rates cancel out the time window. - **Label changes are breaking changes** (Early): Adding or removing labels from instrumentation breaks existing queries and aggregations. Info metrics are the exception—they're designed so extra labels don't cause problems. - **Service discovery replaces manual configuration** (Middle): Prometheus supports Consul, EC2, and Kubernetes out of the box, plus file-based and HTTP-based mechanisms for custom sources—essential for dynamic environments where instances change frequently. - **Target labels should form a hierarchy** (Middle): Structure labels from broad to specific (region → datacenter → environment → service → job → instance) to enable meaningful aggregation and organization. - **Relabeling serves two purposes** (Middle): Originally for mapping service discovery metadata to target labels, it's also used for metric relabeling—like dropping arbitrary container_label_* labels from cAdvisor that could break monitoring. ## 【Reading Tips】 - **Skim the opening monitoring theory** (~0%–10%) if you're already familiar with observability concepts—the practical setup example in Chapter 2 is where the real value begins. - **Deep-read the instrumentation chapters** (~10%–23%) if you're a developer—the Python, Java, and Go examples show concrete patterns, but you can focus on your primary language. - **Pay special attention to the PromQL examples** (~23%–32%)—the rate calculations and group_left joins are the most commonly misunderstood concepts; work through the arithmetic yourself. - **The service discovery and relabeling sections** (~39%–48%) are dense but crucial for production deployments—the label hierarchy advice is worth internalizing even if you skip the specific cloud provider examples. - **Watch for the "breaking changes" warnings** about labels—these are hard-won lessons that will save you from painful migrations later. ## 【Coverage Limits】 The excerpts cover roughly the first half of the book in detail (through service discovery and relabeling), with lighter coverage of later chapters on Kubernetes monitoring, alerting, and advanced production topics. Some specific exporter examples and advanced PromQL features may not be fully represented. ##