analyzing-tdigest-metrics

To compute latency percentiles from tdigest metrics, use the double-combine pattern in OPAL: align combined:tdigest_combine(m_tdigest("metric_name")), then apply tdigest_quantile(tdigest_combine(combined), 0.50/0.95/0.99) to derive p50, p95, and p99 values for your services.

TDigest pre-aggregated percentile metrics store compact distribution data for efficient latency analysis and SLO validation. You need them when tracking tail latency trends and comparing p50, p95, or p99 values against service level targets across production environments.

Converting tdigest latency output from nanoseconds to milliseconds requires applying a unit conversion calculation after extracting percentile values with tdigest_quantile, ensuring the final latency metrics are readable and standardized for reporting.

Yes, this tdigest analysis approach requires Observe MCP access to discover tdigest metric types using discover_context("duration tdigest", result_type="metric") before constructing the OPAL query to compute latency percentiles.

Grouping tdigest latency percentiles by service and environment involves using the group_by(service_name, environment) function within your OPAL query, allowing you to track percentile trends over time and identify specific services showing tail latency degradation.