Homepage Performance Evidence

The Eden homepage leads with each directly measured Eden result and then shows the matched baseline or capacity context. Every claim is deliberately scoped to its measured workload; none should be read as a universal capacity or cost guarantee.

Gateway capacity by workload

The homepage capacity graph shows the highest clean result at each 1, 2, 4, and 8 vCPU container budget. "Clean" means the measured phase completed with zero benchmark errors, Eden shedding, and status failures; it is not an SLO or a universal sizing guarantee. Each graph lane uses its own scale, so compare the trend within a workload rather than bar height across different protocols.

• Workloads: Redis real-backend mixed GET/SET; PostgreSQL pgbench

TPC-B; PostgreSQL synthetic prepared select-only; MongoDB synthetic OP_MSG find; and synthetic OpenAI-compatible chat with a 128-byte response.

• Environment: Adam 32-CPU Linux benchmark host. Eden was constrained to

the displayed 1, 2, 4, or 8 vCPU budget, with Tokio workers and proxy shards matched to that budget; observability exporters were disabled for the gateway hot path.

• Run date: June 21, 2026.
• Source: Gateway capacity planning

and benchmark/GATEWAY_VCPU_SCALING_REPORT_2026-06-21.md.

Gateway efficiency

• Claim: 32% less CPU per completed request.
• Workload: mixed Redis GET/SET at a matched 400,000 requests per

second, with zero shedding for both gateway paths.

• Baseline: Envoy Redis proxy.
• Hardware: 32-CPU Linux benchmark host.
• Run date: June 5, 2026; profile

profile_adam-redis-profile-400k_20260605_082446.

• Result: Eden Gateway used 10.83 CPU-seconds per million completed

requests; Envoy used 15.95.

• Sources: docs/PERFORMANCE.md,

benchmark/EDEN_EVE_PERFORMANCE_BRIEF.md, and the public Gateway benchmarks.

Database forwarding

• Claim: 1.24× throughput on the audited comparison.
• Workload: select-only pgbench over the simple-query protocol, 256

clients, 32 jobs, and 10-second target phases. This uses a fixed-response PostgreSQL wire backend, so it measures connection handling, multiplexing, and forwarding—not storage, WAL, locks, or query execution.

• Baseline: direct synthetic PostgreSQL wire backend.
• Hardware: Adam benchmark host, 32 logical CPUs. The Eden process was

pinned to 16 logical CPUs and used eight Tokio workers and eight proxy shards.

• Run date: July 12, 2026.
• Result: Eden completed 90,011 TPS; the direct baseline completed

72,551 TPS.

• Source: benchmark/PGBOUNCER_PGDOG_COMPARISON_2026-07-12.md.

Telemetry throughput

• Claim: 2,364× throughput.
• Workload: labeled-counter churn.
• Baseline: OpenTelemetry SDK.
• Hardware: 10-core Apple M1 Pro.
• Run date: April 6, 2026.
• Result: fast telemetry measured 8,798,611,227 operations per second;

the OpenTelemetry SDK measured 3,721,890 operations per second.

• Sources: eden-website/lib/fast-telemetry-benchmark-data.ts, the

benchmark result, and the raw harness artifact.

AI Gateway throughput

• Claim: Gateway's clean ceiling was 12,000 requests per second for the

buffered-response scenario. Envoy completed 2,530 requests per second when offered 8,000 requests per second, while recording 42,779 errors.

• Workload: buffered 64 KiB responses from the same synthetic upstream.
• Baseline: generic Envoy HTTP proxy.
• Scope: this is a sustained-load comparison, not the small-response

1/2/4/8-vCPU capacity sweep displayed in the separate OpenAI chat row.

• Source: Gateway benchmarks, AI and LLM

Gateway sustained-load results.

ShardKV throughput

• Published comparisons: ShardKV completed 3,255,162 operations per second

in the single-core TCP run and 17,504,612 operations per second in the clustering run.

• Workloads: the single-core scenario uses TCP, 64-byte values, an 80/20

GET/SET mix, and pipeline depth 64. The clustering scenario uses small-value GET requests through the ShardKV direct path.

• Baseline: the documented market baseline completed 1,204,399 operations

per second in the single-core scenario and 9,596,424 operations per second in the clustering scenario.

• DashMap follow-up: the 134,596,497 operations-per-second embedded

reference-lookup result is deliberately excluded from the homepage until a like-for-like DashMap run is recorded with the same workload, hardware, and source artifact.

• Hardware: 16 vCPU. The underlying report does not retain a host model or

run date, so these results are shown as documented throughput reports rather than capacity-planning guarantees.

• Result: ShardKV measured 2.7× the reported single-core baseline and 1.8×

the reported cluster baseline.

• Source: sales/reports/shardcache-throughput.html.

Review rule

Homepage performance claims must be removed or updated when their workload, baseline, hardware, run date, or source is no longer available for review.