Performance

This page summarizes the current local cost evidence for the v3-preview gateway. It is a planning input for a controlled preview, not a provider-price, dedicated-runner, or external-provider performance claim. The dated reports link to every sample, revision binding and method detail.

Current measured costs

Bounded compaction

At be30107, compaction can select 256 small source runs instead of 128 while keeping the 16 MiB encoded-source and 131,072-mutation limits. The wire format, recovery promise, publication verification and reader limits do not change. Three alternating pairs passed for each workload. Stored growth subtracts each initial inventory; retained setup objects still occupy storage.

Median metric	512 B serial overwrites, baseline → candidate	64 Kopia snapshots, baseline → candidate
Stored growth, bytes	2,010,316 → 1,659,487	1,662,906,181 → 1,662,596,134
Uploaded body bytes	2,010,775 → 1,659,792	1,662,906,486 → 1,662,596,287
Downloaded body bytes	7,484,177 → 6,000,267	286,443,243 → 285,291,645
Backend HTTP requests	14,610 → 14,669	12,965 → 13,286

For tiny serial writes, stored growth falls 17.45%, uploaded body data 17.46%, and downloaded body data 19.83%; requests rise 0.40%. Storage amplification falls from 7.669x to 6.330x submitted plaintext. For Kopia, stored and uploaded bytes fall 0.019%, downloads 0.36% at the paired median, and requests rise 2.48%. These serial writes cannot share a publication batch. The bounded-compaction report explains the endpoint checkpoint tradeoff and preserves all 12 results.

Gateway versus direct S3

At 2d284c9, three alternating direct/gateway pairs passed for each workload. Ratios are median gateway/direct values. Stored bytes include all surviving versions and gateway setup data; operating traffic excludes bootstrap and includes workload, restore, maintenance, and observed background activity.

Workload	Stored bytes	Uploaded body bytes	Downloaded body bytes	Backend requests
Kubernetes-shaped Kopia, two snapshots	1.035x	0.999x	0.567x	0.105x
PostgreSQL-shaped Kopia, two snapshots	1.013x	0.999x	0.966x	3.379x
Two 100 MiB multipart versions	1.025x	1.000x	2.001x	24.385x
512 serial overwrites of 512 B	27.693x	7.670x	6,658.494x	28.259x

Gateway setup retains about 5.006 MiB and uses 319 backend requests, about 5.007 MiB uploaded and 0.068 MiB downloaded. It remains part of retained storage even though it is excluded from operating traffic. Tiny-write ratios have very small direct denominators and do not describe concurrent batching.

What the ratios include

Standalone and multipart writes read completed ciphertext back once before publication. Replacing that readback with provider-reported checksums would change the malicious or eventually inconsistent backend contract. The multipart write body path is therefore about 2x direct S3 before later downloads.

Kopia restores clear the client cache; the running gateway keeps its default 256 MiB decrypted-segment cache. Aggregate restore-download reductions are not cold-gateway-restart results. Direct S3 and rs3 also have different namespace and historical-read behavior, so request ratios are workload-specific.

Both paths request a 30-day protection floor. Direct S3 uses bucket-default COMPLIANCE retention; rs3 uses explicit version locks, a 30-day recovery window, one-day renewal margin and 60-second clock allowance. Whole-day rounding produced about 32 days of observed rs3 protection versus 30 days directly. Current locked byte-day obligations are 1.080x–1.101x direct for the large workloads and 28.205x for tiny overwrites, including setup. They are not a monthly bill or a 30-day elapsed-storage measurement.

Method and limits

The current reports use fresh disposable RustFS, real Kind Leases, generated salts, exact-version retention audits and content recovery checks. The bounded-compaction comparison uses explicit 30-day COMPLIANCE protection, a 30-day recovery window, one-day margin and 60-second allowance. The direct comparison uses its distinct direct bucket-default policy and rs3 explicit locks; the policies provide the same fresh-data floor but are not long-lived retention equivalents.

Both reports count decoded HTTP bodies and requests with a common streaming meter. Headers, framing, TCP/TLS, Kubernetes traffic, provider internal storage, prices, and compute are outside the byte totals. Provisioning and inventory audits bypass the meter; observed background backend traffic remains included. Lanes run serially on a shared host without concurrent builds. Elapsed time and writer CPU/RSS are descriptive, not isolated capacity data.

The bounded-compaction report and gateway/direct report contain complete workload shapes, restore checks, phase rules and source/binary/helper hashes. The corresponding machine-readable records are bounded compaction and gateway/direct storage.

Earlier measured comparisons

The September 8 performance report retains all sample tables, revision bindings and reproduction recipes for the previous v03 comparison and history experiments. Its two open candidate budget failures remain evidence, not waived limits:

• 453c2ff sequential 512 B writes measured 3.918274x backend-write amplification against a 3x ceiling.
• 453c2ff large PostgreSQL-shaped Kopia measured 1.904731x gateway/direct elapsed time against a 1.75x ceiling.

It also retains the source-bound history comparison: cdedf39 baseline and 3e879cb candidate used 256 MiB metadata at 10,000 writes and explicit 1 GiB at 100,000, with a 64 MiB pending buffer. The baseline 100,000-append full history walk refused those same budgets; the candidate completed its lifecycle. Those controlled-provider results are not HTTP, Kubernetes, external-provider, or process-RSS qualification.

Use a clean committed checkout and build outside timed intervals when repeating a report. Recreate each named revision, input, compiler profile, provider image and resource envelope; the commands and failure gates are preserved in the historical report. Correctness and retention acceptance are described in Testing and Production Preview. Neither changes a measured cost or failed budget.