This page renders the pilot report verbatim from the data pack, with two mechanical link fixes noted in scripts/generate-site-data.mjs. Download the verbatim file.

Tier 1 default-transport validation (2026-10-18)

Socket-option erratum: historical tcp_nodelay=off / override 0 labels did not prove TCP_NODELAY was disabled. The old client skipped its override while NIO enabled TCP_NODELAY by default; its explicit 1 used the wrong option level. See the correction and validation (companion report, not included in this site’s data pack). Raw measurements are preserved; rebaseline with verified socket values.

Tier 1 is now the preferred HTTP/1 transport. Applications registering any request-body streaming route use the stock SwiftNIO pipeline for the entire application. Explicit useTier1HTTP1Experimental: false also selects stock. The historical configuration name remains source-compatible. This change does not establish a new Express/Hono leaderboard result.

Correctness and scope

The promotion includes the previously unfinished parser, response-stream, and file-region work. The chunk-extension grammar and trailer-value checks are tested against the pinned SwiftNIO decoder. File regions stay zero-copy IOData.fileRegion; response streaming shares length accounting and event-loop checks with the stock writer. HEAD and bodyless responses, explicit flushes, declared-length failures, producer errors, HTTP/1.0 close framing, and pipelined ordering have regression coverage.

An additional audit found that the old Tier 1 handler dispatched pipelined followers and forwarded socket reads while an async response was pending. The new handler retains already received bytes and withholds reads until the response or open response stream ends. It still reads incomplete request bodies. Disconnect drops pending state; a reentrancy guard keeps synchronous completion from parsing a follower before request state resets. The new async/read regression fails on 4acc77c with both unwanted actions observed, and passes after the fix. This correctness fix is retained independently of throughput.

Request-streaming selection tests exercise the same pipeline installer as production: ordinary keep-alive requests, stock override, incremental uploads, consumer backpressure, body limits, middleware bypass, and Content-Length/Transfer-Encoding ambiguity. The standalone Tier 1 handler’s 500 guard for request-streaming actions remains covered; the production bootstrap prevents streaming applications from reaching it.

Fail-safe differences from stock remain explicit: CONNECT/bodyless-status streams do not start producers; post-end producer writes close; repeatedly invalid streaming error responses terminate with a fixed 500.

Build and test evidence

In-process diagnostic

Active EmbeddedChannels, identical raw GET bytes and response bytes, 1,000 warm-up requests per case, six alternating rounds of 20,000 requests. These are local pipeline costs including parsing, boundary conversion, dispatch, and framing, without sockets. Timing outliers remain in the data.

Response Stock median ns/op Tier 1 median ns/op
Synchronous ok 3222 1789
Chunked stream ok 4744 3881

Stock sync rounds span 3176–3555 ns; Tier 1 sync spans 1786–1817 ns. Stock stream spans 4716–4779 ns; Tier 1 stream spans 3857–4195 ns. This diagnostic is not a network-throughput claim. See /tmp/wt-linux/promotion-linux-pipeline-post-cleanup.log and the opt-in command in BENCHMARKS.md. CI executes only a smoke-sized version without thresholds.

Network experiment

Baseline: 4acc77c29655c6a6f408359ff46dfafe62f2d91b, temporary checkout outside the repository. Measured production-source diff SHA-256: eca917afcc45b84e81e3cec3beb4e0ea0d600f4ca3e67936da91e39542d994d4.

After this matrix, a cold-path cleanup releases stream-validation error keys after mapped responses emit and hops foreign error futures to the channel loop. It adds no ordinary-response check and changes none of the measured request/response paths. Final production diff SHA-256: e65c180a89bf24078f28c816b4491fed1cd823f382d3acf1bd76a30ac1fff2fe. The full release suite, pipeline diagnostic, product builds, and final binary network smoke are repeated after that cleanup; the matrix below identifies the earlier measured source explicitly.

Apple M1 Pro host; Docker Desktop Linux VM, 8 vCPUs; Linux 7.0.12-linuxkit, aarch64; Swift 6.4 release builds using SwiftPM’s SwiftBuild backend. Server processes have two event-loop threads and CPU affinity 0–3; the same current benchmark client is pinned to CPUs 4–7. The client requests four threads but uses two loops because of the existing double group.next() behavior; this harness limitation is held constant. HTTP/1.1, persistent connections, no TLS/compression, client override flag TCP_NODELAY=0 (effective value unverified).

Each cell has a 20,000-request warm-up followed by separate throughput and latency samples of 250,000 requests. Four rounds alternate server order. Concurrency is 64, 256, and 512. The primary matrix covers GET /, GET /user/42, and POST /user across baseline stock, baseline Tier 1, current default Tier 1, and unchanged raw NIO. Response-stream controls compare baseline stock, current forced stock, and current default on known-length and chunked Hello, World! responses. The baseline Tier 1 cannot serve these streams, so it is not a stream control.

Response status, payload, and relevant headers are checked before timing. The optional response-stream routes are registered on all framework servers. CPU ticks and process peak RSS are recorded from /proc with each sample. RSS is a process-lifetime high-water mark including earlier warm-up/load, not a fresh-process peak for each cell.

All 432 measured samples completed: 108 million measured requests with no client failures. Request counts, body-byte counts, and latency sample counts were checked for every sample. Startup readiness retries are outside the measured runs. The tables use medians of four separate runs; latency samples are not pooled. Full per-round metrics, including p50, p90, p99.999, maximum, CPU ticks, RSS, and body throughput, are in the measurement data.

Throughput, requests/second (latency collection disabled):

Route Connections Old stock Old Tier 1 Current default Raw NIO
GET / 64 203,664 202,248 199,148 196,847
GET / 256 206,680 200,972 195,248 205,403
GET / 512 187,596 198,518 188,394 203,912
GET /user/42 64 172,526 201,150 203,708 195,782
GET /user/42 256 177,696 201,876 195,313 185,049
GET /user/42 512 175,212 189,378 196,387 186,685
POST /user 64 192,266 191,226 188,956 194,428
POST /user 256 190,680 187,680 178,070 192,308
POST /user 512 184,685 183,616 183,238 183,930

p99 cycle latency, microseconds (separate latency-enabled runs):

Route Connections Old stock Old Tier 1 Current default Raw NIO
GET / 64 673 863 764 729
GET / 256 2548 2929 2538 2689
GET / 512 4941 5202 5037 6160
GET /user/42 64 801 695 727 755
GET /user/42 256 2851 2698 2631 2661
GET /user/42 512 6898 5053 5028 5940
POST /user 64 715 709 788 700
POST /user 256 2638 2392 2448 2964
POST /user 512 5734 3628 5175 5328

Stream results below are throughput RPS / p99 microseconds, taken from the separate throughput and latency runs respectively:

Response Connections Old stock Current forced stock Current default
Known length 64 127,650 / 1284 120,894 / 1049 141,240 / 938
Known length 256 130,528 / 3020 128,141 / 3537 141,240 / 3193
Known length 512 127,944 / 5001 128,762 / 5596 138,387 / 8766
Chunked 64 109,574 / 1102 112,207 / 1131 122,856 / 1098
Chunked 256 110,691 / 3931 111,648 / 4685 123,906 / 3673
Chunked 512 109,386 / 7692 110,284 / 9863 128,152 / 5602

For known-length streams at c512, the four current-default p99 values are 10,928 / 15,658 / 6604 / 5545 us, versus old stock 5345 / 16,762 / 4363 / 4657 us. The medians indicate a substantial tail cost despite higher throughput; the overlapping excursions prevent a clean mechanism attribution. For chunked streams, current forced-stock c512 p99 also rises in the median, despite approximately neutral throughput.

CPU examples, median process CPU-seconds per 250,000 throughput requests (100 ticks/second):

Route / connections Old stock Old Tier 1 Current default Raw NIO
GET /, 256 2.395 2.185 2.170 2.380
GET /user/42, 512 2.840 2.515 2.420 2.685

Known-length c512 streams use 3.910 CPU-seconds on old stock, 3.880 on current forced stock, and 3.615 on current default. Maximum observed process RSS is 24,968 KiB (old stock), 24,084 (old Tier 1), 24,192 (current default), 24,996 (current forced stock), and 24,332 (raw). No material memory growth appears in this finite workload; it is not a long-duration leak test.

Interpretation and tradeoff

This is a correctness-backed default selection, not an overall performance win. Ordinary-route throughput versus old Tier 1 ranges from -5.1% to +3.7% in median, with broad overlapping round spreads. The raw NIO gap is not reduced on every route. POST c512 median p99 is worse than old Tier 1; known-length streaming c512 median p99 is worse than both stock controls. Those signals must not be hidden by the better streaming RPS figures.

The explicit tradeoff is to retain bounded pipelining and the Tier 1 default while preserving the stock override and upload fallback. This does not pass the gate for claiming uniformly faster throughput or improved tails. The remaining tail investigation and independent load-generator validation are still open. No cross-framework rankings are updated from these results.

Raw logs and the external runner remain under /tmp/wt-linux/: promotion-network/, promotion-network.sh, and the two production diff snapshots. The tracked JSON preserves all measured per-run metrics without requiring those temporary files. Reproduce the matrix with the server and client commands in BENCHMARKS.md, both baseline modes, current default, and raw, retaining the exact routes, affinities, counts, flags, and order above.

Tail-latency limits

The historical bimodality investigation is diagnostic evidence, not a resolved mechanism. Additional 500,000-request c512 runs showed Tier 1 p99 between 3418 and 7254 microseconds and Hono between 3809 and 7397 microseconds. Changing the client from four requested threads to three did not eliminate excursions. Per-connection means were broadly uniform. These observations do not support a persistent group of slow connections or a Tier-1-only explanation; they do not prove a scheduler or VM cause.

The exact mechanism remains unresolved. This promotion prioritizes audited protocol behavior and bounded pipelining, preserves a stock escape hatch, and makes no claim that the old p99 issue is fixed. Bare-metal Linux with a separate generator and an independent harness remains necessary before generalizing VM results or claiming uniformly better tails.