Performance scenarios
Each curated performance scenario is a named lab setup: what Conduit is doing, what kind of load it faces, and which knobs stay fixed. Table rows on reference results and study evidence deep-link here.
This page is a glossary for those links — not the primary decision surface. Start from Performance findings or Tuning evidence (studies).
Recurring terms such as sync /
split_io and load shapes
forward_fast /
forward_slow /
cache_hit are defined in
methodology and
runtime and concurrency.
feature_tax
feature-tax-dnstap-off-forward-fast
Baseline for dnstap cost: forwarding to a fast stub upstream with dnstap and event sinks disabled.
Notes: Compare with the sampled and full dnstap cells in the same study.
What varies: runtime=sync, load_shape=forward_fast, obs_posture=dnstap_off, ingress_workers=2
How it was run: config feature-tax-dnstap-off.yml; upstream fast (fast stub upstream); loadgen dnsperf; clients=16, threads=8, max_outstanding=2000.
feature-tax-dnstap-sampled-forward-fast
Measures dnstap cost when only about 10% of response frames are emitted to a lab receiver, still under a fast stub upstream.
Notes: Middle cell between dnstap-off and full dnstap emit. Requires the lab dnstap tracer.
What varies: runtime=sync, load_shape=forward_fast, obs_posture=dnstap_sampled, ingress_workers=2
How it was run: config feature-tax-dnstap-sampled.yml; upstream fast (fast stub upstream); loadgen dnsperf; clients=16, threads=8, max_outstanding=2000.
feature-tax-dnstap-full-forward-fast
Measures how much throughput changes when Conduit emits full dnstap event frames (query, response, and retry) to a lab receiver while forwarding to a fast stub upstream.
Notes: Part of the dnstap cost series with dnstap-off and dnstap-sampled. Requires the lab dnstap tracer.
What varies: runtime=sync, load_shape=forward_fast, obs_posture=dnstap_full, ingress_workers=2
How it was run: config feature-tax-dnstap-full.yml; upstream fast (fast stub upstream); loadgen dnsperf; clients=16, threads=8, max_outstanding=2000.
feature-tax-logging-warn-forward-fast
Measures quieter logging: metrics are off and logging.level is warn while forwarding to a fast stub upstream.
Notes: Paired with the logging-debug cell. Named for the logging axis — it is not the shared metrics-off baseline used by other feature_tax studies.
What varies: runtime=sync, load_shape=forward_fast, obs_posture=logging_warn, ingress_workers=2
How it was run: config feature-tax-logging-warn.yml; upstream fast (fast stub upstream); loadgen dnsperf; clients=16, threads=8, max_outstanding=2000.
feature-tax-logging-debug-forward-fast
Measures the cost of verbose logging: metrics are off and logging.level is debug while forwarding to a fast stub upstream.
Notes: Paired with feature-tax-logging-warn-forward-fast for the logging verbosity comparison.
What varies: runtime=sync, load_shape=forward_fast, obs_posture=logging_debug, ingress_workers=2
How it was run: config feature-tax-logging-debug.yml; upstream fast (fast stub upstream); loadgen dnsperf; clients=16, threads=8, max_outstanding=2000.
feature-tax-metrics-no-collect-forward-fast
Hot-path metric categories neither record nor export (no collect) under a fast stub upstream.
Notes: Paired with collect-only and collect+emit.
What varies: runtime=sync, load_shape=forward_fast, obs_posture=metrics_no_collect, ingress_workers=2
How it was run: config feature-tax-metrics-no-collect.yml; upstream fast (fast stub upstream); loadgen dnsperf; clients=16, threads=8, max_outstanding=2000.
feature-tax-metrics-collect-only-forward-fast
Records hot-path metrics in process but does not scrape or push them (collect on, emit off) under a fast stub upstream.
Notes: Paired with collect+emit and no-collect.
What varies: runtime=sync, load_shape=forward_fast, obs_posture=metrics_collect_only, ingress_workers=2
How it was run: config feature-tax-metrics-collect-only.yml; upstream fast (fast stub upstream); loadgen dnsperf; clients=16, threads=8, max_outstanding=2000.
feature-tax-metrics-collect-emit-forward-fast
Records hot-path metrics and exports them (collect and emit) under a fast stub upstream, using the standard metrics base.
Notes: Paired with collect-only and no-collect for the collect-versus-emit comparison.
What varies: runtime=sync, load_shape=forward_fast, obs_posture=metrics_collect_emit, ingress_workers=2
How it was run: config feature-tax-metrics-collect-emit.yml; upstream fast (fast stub upstream); loadgen dnsperf; clients=16, threads=8, max_outstanding=2000.
feature-tax-metrics-off-forward-fast
Observability-off baseline under a fast stub upstream: metrics disabled and no dnstap sinks.
Notes: Shared baseline for several feature_tax comparisons (scrape cost, combined surfaces, tracing, and similar).
What varies: runtime=sync, load_shape=forward_fast, obs_posture=metrics_off, ingress_workers=2
How it was run: config feature-tax-metrics-off.yml; upstream fast (fast stub upstream); loadgen dnsperf; clients=16, threads=8, max_outstanding=2000.
feature-tax-metrics-standard-scrape-forward-fast
Prometheus scrape with the standard metrics base (collect and emit) while forwarding to a fast stub upstream.
Notes: High end of the scrape cost series; also the collect+emit pole in collect comparisons that reuse this cell.
What varies: runtime=sync, load_shape=forward_fast, obs_posture=metrics_standard_scrape, ingress_workers=2
How it was run: config feature-tax-metrics-standard-scrape.yml; upstream fast (fast stub upstream); loadgen dnsperf; clients=16, threads=8, max_outstanding=2000.
feature-tax-metrics-standard-dnstap-full-forward-fast
Combined observability cost: standard Prometheus scrape plus full dnstap emit under a fast stub upstream.
Notes: Used in studies that stack scrape and dnstap together.
What varies: runtime=sync, load_shape=forward_fast, obs_posture=metrics_standard_dnstap_full, ingress_workers=2
How it was run: config feature-tax-metrics-standard-dnstap-full.yml; upstream fast (fast stub upstream); loadgen dnsperf; clients=16, threads=8, max_outstanding=2000.
feature-tax-metrics-standard-scrape-hammer-forward-fast
Same standard scrape posture as the listener-only scrape cell, but a lab client hammers /metrics about every 100 ms during the run.
Notes: Shows aggressive scrape cadence versus ordinary scrape traffic on the listener.
What varies: runtime=sync, load_shape=forward_fast, obs_posture=metrics_standard_scrape_hammer, ingress_workers=2
How it was run: config feature-tax-metrics-standard-scrape.yml; upstream fast (fast stub upstream); loadgen dnsperf; clients=16, threads=8, max_outstanding=2000.
feature-tax-metrics-off-scrape-ladder-forward-fast
Observability-off baseline for the metrics scrape cost series, using the same elevated dnsperf in-flight window as the other cells in that series.
Notes: Not the shared metrics-off cell used by other feature_tax studies — this one matches the scrape-series loadgen recipe.
What varies: runtime=sync, load_shape=forward_fast, obs_posture=metrics_off, ingress_workers=2, loadgen_recipe=elevated_outstanding
How it was run: config feature-tax-metrics-off.yml; upstream fast (fast stub upstream); loadgen dnsperf; clients=16, threads=8, max_outstanding=2000.
feature-tax-metrics-minimal-scrape-ladder-forward-fast
Same minimal metrics scrape posture as the standard minimal-scrape cell, run with a larger dnsperf in-flight window so the scrape cost series is not limited by the load generator.
Notes: Published recipe for the metrics scrape cost series; pair with the off and standard cells that share this elevated outstanding window.
What varies: runtime=sync, load_shape=forward_fast, obs_posture=metrics_minimal_scrape, ingress_workers=2, loadgen_recipe=elevated_outstanding
How it was run: config feature-tax-metrics-minimal-scrape.yml; upstream fast (fast stub upstream); loadgen dnsperf; clients=16, threads=8, max_outstanding=2000.
feature-tax-metrics-standard-scrape-ladder-forward-fast
Standard metrics scrape under the elevated dnsperf in-flight window used by the scrape cost series.
Notes: Pair with the off and minimal cells that share this loadgen recipe.
What varies: runtime=sync, load_shape=forward_fast, obs_posture=metrics_standard_scrape, ingress_workers=2, loadgen_recipe=elevated_outstanding
How it was run: config feature-tax-metrics-standard-scrape.yml; upstream fast (fast stub upstream); loadgen dnsperf; clients=16, threads=8, max_outstanding=2000.
feature-tax-metrics-off-split-io-forward-fast
Observability-off baseline on the split_io runtime while forwarding to a fast stub upstream.
Notes: Paired with feature-tax-metrics-standard-scrape-split-io-forward-fast for scrape cost on split_io.
What varies: runtime=split_io, load_shape=forward_fast, obs_posture=metrics_off, ingress_workers=2, policy_workers=2, io_workers=2
How it was run: config feature-tax-metrics-off-split-io.yml; upstream fast (fast stub upstream); loadgen dnsperf; clients=16, threads=8, max_outstanding=2000.
feature-tax-metrics-standard-scrape-split-io-forward-fast
Standard Prometheus scrape on the split_io runtime while forwarding to a fast stub upstream.
Notes: Paired with feature-tax-metrics-off-split-io-forward-fast for scrape cost on split_io.
What varies: runtime=split_io, load_shape=forward_fast, obs_posture=metrics_standard_scrape, ingress_workers=2, policy_workers=2, io_workers=2
How it was run: config feature-tax-metrics-standard-scrape-split-io.yml; upstream fast (fast stub upstream); loadgen dnsperf; clients=16, threads=8, max_outstanding=2000.
feature-tax-metrics-otlp-push-forward-fast
Measures the cost of pushing metrics over OTLP HTTP while forwarding to a fast stub upstream.
Notes: Requires conduit-otlp-metrics-tracer. The lab receiver also records otlp_accepts / otlp_failures as secondary signals.
What varies: runtime=sync, load_shape=forward_fast, obs_posture=metrics_otlp_push, ingress_workers=2
How it was run: config feature-tax-metrics-otlp-push.yml; upstream fast (fast stub upstream); loadgen dnsperf; clients=16, threads=8, max_outstanding=2000.
feature-tax-tracing-on-forward-fast
Metrics stay off while pipeline tracing is enabled (query type A at 100% sample) under a fast stub upstream.
Notes: Paired with feature-tax-metrics-off-forward-fast.
What varies: runtime=sync, load_shape=forward_fast, obs_posture=tracing_on, ingress_workers=2
How it was run: config feature-tax-tracing-on.yml; upstream fast (fast stub upstream); loadgen dnsperf; clients=16, threads=8, max_outstanding=2000.
feature-tax-metrics-minimal-scrape-forward-fast
Prometheus scrape with the minimal metrics base (collect and emit) while forwarding to a fast stub upstream.
Notes: Middle cell between metrics-off and standard scrape.
What varies: runtime=sync, load_shape=forward_fast, obs_posture=metrics_minimal_scrape, ingress_workers=2
How it was run: config feature-tax-metrics-minimal-scrape.yml; upstream fast (fast stub upstream); loadgen dnsperf; clients=16, threads=8, max_outstanding=2000.
lifecycle
lifecycle-cold-start
Wall-clock time from process start until Conduit returns its first successful DNS answer.
What varies: runtime=sync, load_shape=forward_fast, obs_posture=metrics_off, ingress_workers=2
How it was run: config lifecycle-cold-start.yml; upstream fast (fast stub upstream); loadgen none.
lifecycle-config-apply
How long a conduitctl apply takes for a sparse overlay that only changes the logging level.
Notes: Small lifecycle cell for local runs; not required in the published reference set.
What varies: runtime=sync, load_shape=forward_fast, obs_posture=metrics_off, ingress_workers=2
How it was run: config lifecycle-config-apply-base.yml; upstream fast (fast stub upstream); loadgen none.
scale
scale-sync-forward-fast
Throughput for the sync runtime while forwarding to a fast stub upstream.
Notes: Thin published pair with scale-split-io-forward-fast.
What varies: runtime=sync, load_shape=forward_fast, ingress_workers=2
How it was run: config scale-sync-obs-off.yml; upstream fast (fast stub upstream); loadgen dnsperf; clients=16, threads=8, max_outstanding=2000.
scale-sync-cache-hit
Warm memory-cache hits on the sync runtime: the harness fills the cache before dnsperf, then measures near-100% hits against a small answer set.
Notes: Read-mostly after warm — not a high-churn fill/evict shape. Optional for the published reference set; compare with scale-sync-lmdb-cache-hit via the memory-vs-lmdb-cache-hit study.
What varies: runtime=sync, load_shape=cache_hit, ingress_workers=2, cache_backend=memory
How it was run: config scale-sync-cache-hit.yml; upstream fast (fast stub upstream); loadgen dnsperf; clients=16, threads=8, max_outstanding=2000.
scale-sync-ingress-1-forward-fast
Sync runtime with one ingress worker under a fast stub upstream.
Notes: Uses an elevated dnsperf outstanding window so achieved QPS reflects Conduit ingress capacity rather than the load generator. Part of the ingress concurrency series (1/2/4/8).
What varies: runtime=sync, load_shape=forward_fast, ingress_workers=1
How it was run: config scale-sync-ingress-1-obs-off.yml; upstream fast (fast stub upstream); loadgen dnsperf; clients=16, threads=8, max_outstanding=2000.
scale-sync-ingress-4-forward-fast
Sync runtime with four ingress workers under a fast stub upstream.
Notes: Elevated dnsperf outstanding window so QPS reflects ingress capacity. Part of the ingress concurrency series (1/2/4/8).
What varies: runtime=sync, load_shape=forward_fast, ingress_workers=4
How it was run: config scale-sync-ingress-4-obs-off.yml; upstream fast (fast stub upstream); loadgen dnsperf; clients=16, threads=8, max_outstanding=2000.
scale-sync-ingress-8-forward-fast
Sync runtime with eight ingress workers under a fast stub upstream.
Notes: Elevated dnsperf outstanding window so QPS reflects ingress capacity. Part of the ingress concurrency series (1/2/4/8).
What varies: runtime=sync, load_shape=forward_fast, ingress_workers=8
How it was run: config scale-sync-ingress-8-obs-off.yml; upstream fast (fast stub upstream); loadgen dnsperf; clients=16, threads=8, max_outstanding=2000.
scale-sync-ingress-1-forward-slow
Sync runtime with one ingress worker under a slow stub upstream.
Notes: Paired with ingress 2/4/8 cells for the ingress-concurrency-sync study.
What varies: runtime=sync, load_shape=forward_slow, ingress_workers=1
How it was run: config scale-sync-ingress-1-obs-off.yml; upstream slow (slow stub upstream (50 ms hold)); loadgen dnsperf; clients=16, threads=8, max_outstanding=2000.
scale-sync-forward-slow
Throughput for the sync runtime while every upstream answer is held for 50 ms (forward_slow).
Notes: Paired with scale-split-io-forward-slow for relative comparison.
What varies: runtime=sync, load_shape=forward_slow, ingress_workers=2
How it was run: config scale-sync-obs-off.yml; upstream slow (slow stub upstream (50 ms hold)); loadgen dnsperf; clients=16, threads=8, max_outstanding=2000.
scale-sync-ingress-4-forward-slow
Sync runtime with four ingress workers under a slow stub upstream.
Notes: Paired with ingress 1/2/8 cells for the ingress-concurrency-sync study.
What varies: runtime=sync, load_shape=forward_slow, ingress_workers=4
How it was run: config scale-sync-ingress-4-obs-off.yml; upstream slow (slow stub upstream (50 ms hold)); loadgen dnsperf; clients=16, threads=8, max_outstanding=2000.
scale-sync-ingress-8-forward-slow
Sync runtime with eight ingress workers under a slow stub upstream.
Notes: Optional noisy rung in the ingress series — keep in the catalog; publish may omit it if the lab is unstable.
What varies: runtime=sync, load_shape=forward_slow, ingress_workers=8
How it was run: config scale-sync-ingress-8-obs-off.yml; upstream slow (slow stub upstream (50 ms hold)); loadgen dnsperf; clients=16, threads=8, max_outstanding=2000.
scale-split-io-io-1-forward-slow
split_io under a slow stub upstream with one I/O worker and ingress/policy fixed at two each.
Notes: Part of the I/O-versus-ingress sizing series (io_workers 1/2/4/8).
What varies: runtime=split_io, load_shape=forward_slow, ingress_workers=2, policy_workers=2, io_workers=1
How it was run: config scale-split-io-io-1-obs-off.yml; upstream slow (slow stub upstream (50 ms hold)); loadgen dnsperf; clients=16, threads=8, max_outstanding=2000.
scale-split-io-forward-slow
Throughput for the split_io runtime while every upstream answer is held for 50 ms (forward_slow).
Notes: Paired with scale-sync-forward-slow; slow upstream is where split_io should show its advantage.
What varies: runtime=split_io, load_shape=forward_slow, ingress_workers=2, policy_workers=2, io_workers=2
How it was run: config scale-split-io-obs-off.yml; upstream slow (slow stub upstream (50 ms hold)); loadgen dnsperf; clients=16, threads=8, max_outstanding=2000.
scale-split-io-io-4-forward-slow
split_io under a slow stub upstream with four I/O workers and ingress/policy fixed at two each.
Notes: Part of the I/O-versus-ingress sizing series (io_workers 1/2/4/8).
What varies: runtime=split_io, load_shape=forward_slow, ingress_workers=2, policy_workers=2, io_workers=4
How it was run: config scale-split-io-io-4-obs-off.yml; upstream slow (slow stub upstream (50 ms hold)); loadgen dnsperf; clients=16, threads=8, max_outstanding=2000.
scale-split-io-io-8-forward-slow
split_io under a slow stub upstream with eight I/O workers and ingress/policy fixed at two each.
Notes: Optional noisy rung in the I/O series — keep in the catalog; publish may omit it if the lab is unstable.
What varies: runtime=split_io, load_shape=forward_slow, ingress_workers=2, policy_workers=2, io_workers=8
How it was run: config scale-split-io-io-8-obs-off.yml; upstream slow (slow stub upstream (50 ms hold)); loadgen dnsperf; clients=16, threads=8, max_outstanding=2000.
scale-split-io-forward-fast
Throughput for the split_io runtime while forwarding to a fast stub upstream.
Notes: Thin published pair with scale-sync-forward-fast.
What varies: runtime=split_io, load_shape=forward_fast, topology=thin, ingress_workers=2, policy_workers=2, io_workers=2
How it was run: config scale-split-io-obs-off.yml; upstream fast (fast stub upstream); loadgen dnsperf; clients=16, threads=8, max_outstanding=2000.
scale-split-io-topology-heavy
Higher worker counts on split_io (ingress, policy, and I/O at four each) under a fast stub upstream — a threads-versus-cores topology check.
Notes: Runnable locally; not required in the published reference set.
What varies: runtime=split_io, load_shape=forward_fast, topology=heavy, ingress_workers=4, policy_workers=4, io_workers=4
How it was run: config scale-split-io-topology-heavy.yml; upstream fast (fast stub upstream); loadgen dnsperf; clients=16, threads=8, max_outstanding=2000.
scale-sync-lmdb-cache-hit
Warm LMDB cache hits on the sync runtime using a real-disk environment path. After warm-up the load is read-mostly (near-100% hits).
Notes: Uses the fixed safe LMDB sync durability default. Not a high-churn fill/evict shape. Compare with scale-sync-cache-hit via the memory-vs-lmdb-cache-hit study.
What varies: runtime=sync, load_shape=cache_hit, ingress_workers=2, cache_backend=lmdb
How it was run: config scale-sync-lmdb-cache-hit.yml; upstream fast (fast stub upstream); loadgen dnsperf; clients=16, threads=8, max_outstanding=2000.
scale-sync-ingress-8-memory-cache-churn
High-churn memory-cache load on sync with eight ingress workers: query set larger than max_entries, stub TTL long enough that entry-cap turnover drives misses, no warm plateau.
Notes: Primary memory member of study memory-vs-lmdb-cache-churn. Matched with the four first-class LMDB sync-mode cells (full / no_meta / periodic / none) at eight ingress workers. Recipe: 4096 unique names, max_entries 2048, stub TTL 60s. The thin-ingress pair (scale-sync-memory-cache-churn) remains as a companion cell, not the study primary.
What varies: runtime=sync, load_shape=cache_churn, ingress_workers=8, cache_backend=memory, lmdb_sync=memory
How it was run: config scale-sync-ingress-8-memory-cache-churn.yml; upstream fast (fast stub upstream); loadgen dnsperf; clients=16, threads=8, max_outstanding=2000.
scale-sync-ingress-8-lmdb-full-cache-churn
High-churn LMDB-cache load on sync with eight ingress workers and explicit multi-env sharding on a real-disk path. Matched entry cap, query diversity, and stub TTL keep continuous fill/evict pressure.
Notes: First-class LMDB sync-mode member of study memory-vs-lmdb-cache-churn. lmdb.sync=full; when_full=evict_one; shard_count=16 (2× ingress). Path /var/tmp/conduit-perf/lmdb-cache-churn-i8-full. Recipe matches the ingress-8 memory cell. Average latency under the elevated outstanding window still tracks roughly outstanding/QPS (Little's Law) — lead takeaways with relative QPS and hit/miss. Thin-ingress scale-sync-lmdb-cache-churn is a companion, not the study primary.
What varies: runtime=sync, load_shape=cache_churn, ingress_workers=8, cache_backend=lmdb, lmdb_sync=full
How it was run: config scale-sync-ingress-8-lmdb-full-cache-churn.yml; upstream fast (fast stub upstream); loadgen dnsperf; clients=16, threads=8, max_outstanding=2000.
scale-sync-ingress-8-lmdb-no_meta-cache-churn
High-churn LMDB-cache load on sync with eight ingress workers and explicit multi-env sharding on a real-disk path. Matched entry cap, query diversity, and stub TTL keep continuous fill/evict pressure.
Notes: First-class LMDB sync-mode member of study memory-vs-lmdb-cache-churn. lmdb.sync=no_meta; when_full=evict_one; shard_count=16 (2× ingress). Path /var/tmp/conduit-perf/lmdb-cache-churn-i8-no_meta. Recipe matches the ingress-8 memory cell. Average latency under the elevated outstanding window still tracks roughly outstanding/QPS (Little's Law) — lead takeaways with relative QPS and hit/miss. Thin-ingress scale-sync-lmdb-cache-churn is a companion, not the study primary.
What varies: runtime=sync, load_shape=cache_churn, ingress_workers=8, cache_backend=lmdb, lmdb_sync=no_meta
How it was run: config scale-sync-ingress-8-lmdb-no_meta-cache-churn.yml; upstream fast (fast stub upstream); loadgen dnsperf; clients=16, threads=8, max_outstanding=2000.
scale-sync-ingress-8-lmdb-none-cache-churn
High-churn LMDB-cache load on sync with eight ingress workers and explicit multi-env sharding on a real-disk path. Matched entry cap, query diversity, and stub TTL keep continuous fill/evict pressure.
Notes: First-class LMDB sync-mode member of study memory-vs-lmdb-cache-churn. lmdb.sync=none; when_full=evict_one; shard_count=16 (2× ingress). Path /var/tmp/conduit-perf/lmdb-cache-churn-i8-none. Recipe matches the ingress-8 memory cell. Average latency under the elevated outstanding window still tracks roughly outstanding/QPS (Little's Law) — lead takeaways with relative QPS and hit/miss. Thin-ingress scale-sync-lmdb-cache-churn is a companion, not the study primary.
What varies: runtime=sync, load_shape=cache_churn, ingress_workers=8, cache_backend=lmdb, lmdb_sync=none
How it was run: config scale-sync-ingress-8-lmdb-none-cache-churn.yml; upstream fast (fast stub upstream); loadgen dnsperf; clients=16, threads=8, max_outstanding=2000.
scale-sync-ingress-8-lmdb-periodic-cache-churn
High-churn LMDB-cache load on sync with eight ingress workers and explicit multi-env sharding on a real-disk path. Matched entry cap, query diversity, and stub TTL keep continuous fill/evict pressure.
Notes: First-class LMDB sync-mode member of study memory-vs-lmdb-cache-churn. lmdb.sync=periodic with sync_interval=1s; when_full=evict_one; shard_count=16 (2× ingress). Path /var/tmp/conduit-perf/lmdb-cache-churn-i8-periodic. Recipe matches the ingress-8 memory cell. Average latency under the elevated outstanding window still tracks roughly outstanding/QPS (Little's Law) — lead takeaways with relative QPS and hit/miss. Thin-ingress scale-sync-lmdb-cache-churn is a companion, not the study primary.
What varies: runtime=sync, load_shape=cache_churn, ingress_workers=8, cache_backend=lmdb, lmdb_sync=periodic
How it was run: config scale-sync-ingress-8-lmdb-periodic-cache-churn.yml; upstream fast (fast stub upstream); loadgen dnsperf; clients=16, threads=8, max_outstanding=2000.
shutdown_drain
shutdown-drain-complete-forward-slow
Stop under load with a long drain timeout budget (drain-complete) while forwarding to a slow stub upstream. Records drain duration and client loss during the stop window.
Notes: Paired with drain-budgeted and drain-minimal for relative comparison.
What varies: runtime=sync, load_shape=forward_slow, drain_policy=drain_complete, ingress_workers=2
How it was run: config shutdown-drain-complete-forward-slow.yml; upstream slow (slow stub upstream (50 ms hold)); loadgen dnsperf; clients=4, threads=2 (dnsperf default outstanding ≈ 100).
shutdown-drain-budgeted-forward-slow
Stop under load with a short drain timeout (drain-budgeted) while forwarding to a slow stub upstream. Records how long drain takes and how many client queries are lost in the stop window.
Notes: Paired with drain-complete and drain-minimal for relative comparison.
What varies: runtime=sync, load_shape=forward_slow, drain_policy=drain_budgeted, ingress_workers=2
How it was run: config shutdown-drain-budgeted-forward-slow.yml; upstream slow (slow stub upstream (50 ms hold)); loadgen dnsperf; clients=4, threads=2 (dnsperf default outstanding ≈ 100).
shutdown-drain-minimal-forward-slow
Stop under load with drain disabled (shutdown.drain false — no wait) while forwarding to a slow stub upstream. Records drain duration and client loss during the stop window.
Notes: Paired with drain-complete and drain-budgeted for relative comparison.
What varies: runtime=sync, load_shape=forward_slow, drain_policy=drain_minimal, ingress_workers=2
How it was run: config shutdown-drain-minimal-forward-slow.yml; upstream slow (slow stub upstream (50 ms hold)); loadgen dnsperf; clients=4, threads=2 (dnsperf default outstanding ≈ 100).