Callaghan argues the phrase 'three pillars of observability' is misleading — traces, logs, and metrics are three data types, not pillars. His ranking: traces are the most powerful because they capture hierarchical context (nested trees of events queryable at any level), metrics are the least useful. Logs sit in the middle and are the pragmatic starting point: more familiar to developers, cheaper and faster than traces, and easier to sell to a team. Tracing is slower and more expensive because the code captures full context, and brings in head-sampling / tail-sampling cost control problems worse than logging-cost issues. Recommendation: start with structured logs, add incrementally, graduate to traces once teams have momentum.