Skip to content

Tuning evidence (studies)

Comparative case studies from the performance harness: operator questions, controlled axes, and same-host comparisons against baselines on a single reference host. These are not service-level objectives. Reproduce before sizing locally.

Prefer this hub for decisions. Start from Performance findings for a short synthesis, then open the matching study below. The reference results page is the dense chart warehouse; methodology explains how cells were measured.

If you are deciding

Runtime & workers

If you are deciding… Start here
sync vs split_io under fast / slow upstreams Sync vs split_io
More UDP ingress threads on sync Ingress concurrency (sync)
More I/O workers on split_io (ingress fixed) I/O vs ingress (split_io)
Whether raising all worker pools together helps Split_io bulk topology
Warm answer-cache vs forwarding throughput Cache hit vs forward
Warm memory vs LMDB answer-cache throughput Memory vs LMDB warm cache_hit
High-churn memory vs LMDB (QPS and hit/miss) Memory vs LMDB high-churn cache

Observability tax

If you are deciding… Start here
metrics.base minimal vs standard scrape cost Metrics scrape tax
Collect vs emit (recording vs export) Metrics collect vs emit
OTLP metrics push cost vs off / scrape OTLP tax under load
Scrape tax when already on split_io Metrics scrape (split_io)
Frequent Prometheus scrape under load Aggressive scrape cadence
Dnstap off / sampled / fuller emit cost Dnstap emit tax
Standard scrape and fuller dnstap together Combined metrics + dnstap
Process log level warn vs debug Logging verbosity tax
Pipeline tracing on vs off Pipeline tracing tax

Lifecycle

If you are deciding… Start here
Shutdown drain complete / budgeted / minimal Drain policy under slow

After the study: follow its Related guides into config and concepts, then reproduce the same study (--study …) on your binary if the decision is load-sensitive.