Configure the OpenTelemetry Collector groupbytrace processor to buffer and group spans by trace ID before a downstream tail-sampling processor

Steps

1. Add a groupbytrace processor block under processors with wait_duration (default 1s) set long enough that spans for a given trace have typically all arrived before release.
2. Set num_traces (default 1,000,000) to the maximum number of in-flight traces buffered in memory; watch the otelcol_processor_groupbytrace_traces_evicted metric and raise num_traces if it stays above zero.
3. Set num_workers (default 1) to control concurrent processing of completed trace groups; GOMAXPROCS is a reasonable starting point for tuning.
4. In service.pipelines.traces.processors, place groupbytrace before tailsamplingprocessor (or any other processor that needs a complete trace) so downstream stages see whole, released traces rather than partial spans.
5. Avoid placing the batch processor before groupbytrace, since batching can interleave spans from different traces before grouping ever sees them; a batch processor placed after groupbytrace is fine and will typically batch one full trace per entry.
6. If the Collector fleet is horizontally scaled, ensure a routing layer (e.g. the loadbalancing exporter with trace-ID-based routing) sends all spans of a given trace to the same Collector instance, since groupbytrace only groups what it receives locally.