[michaelstaib]

With the HotChocolate.Fusion.Experimental.SocketsHttpTransport AppContext switch, Fusion sends source schema requests through one SocketsHttpHandler per source schema that it owns, with YARP's forwarding defaults (no cookies, no redirects, no decompression, 15 s connect timeout, pooled connections closed after 2 minutes instead of handler rotation), and copies the gateway request headers named in an AppContext data entry onto source schema requests itself instead of through ASP.NET Core header propagation. A schema reload keeps the handler of every source schema whose settings did not change, and requests without an execution deadline keep HttpClient's 100 s bound. This builds on upstream 4adcdf2 (the diff includes ChilliCream#10711 and ChilliCream#10712 until the fork's main catches up), so it measures only what goes beyond the invoker route of ChilliCream#10712: on the official benchmark with both inline settings, the transport keeps throughput, latency, CPU and RSS within run-to-run noise (about 1% req/s, 2% CPU) and removes about 0.7 KB of allocations per request. The switch keeps the default path unchanged, so this can land before a public configuration API exists.

Metric	Before Optimization	After Optimization
Constant-latency req/s, AOT	3,342	3,354
Constant-latency p50 / p95, AOT	13.91 / 19.43 ms	13.87 / 19.30 ms
Constant-latency req/s, JIT	3,398	3,375
Constant-latency p50 / p95, JIT	13.70 / 19.10 ms	13.77 / 19.32 ms
Constant req/s, AOT	4,775	4,786
Constant req/s, JIT	4,984	4,984
Gateway CPU per request, constant-latency, AOT / JIT	0.790 / 0.679 ms	0.802 / 0.692 ms
Allocated B/request, JIT, constant-latency / constant	26,014 / 26,360	25,318 / 25,677
Idle RSS, AOT / JIT	50.5 / 94.4 MB	50.3 / 94.3 MB
Peak RSS, constant-latency, AOT / JIT	102.5 / 173.8 MB	102.2 / 171.7 MB
AOT binary / startup	42.63 MB / 76 ms	42.65 MB / 74 ms

Official 16.7.0 subgraphs, alias-batching gateway configuration, k6 deep-recursion with 50 VUs, Kestrel UnsafePreferInlineScheduling and DOTNET_SYSTEM_NET_SOCKETS_INLINE_COMPLETIONS=1 on for both sides. Medians of 4 (AOT constant-latency), 3 (AOT constant) and 2 (JIT) rounds; the first AOT variant rounds used a binary that still set AutomaticDecompression explicitly, which differs only in linked code. The benchmark gateway opts in with BENCH_FUSION_SOCKETS_TRANSPORT=1.

Where the remaining transport cost is

JIT profile at constant latency (thread-time samples scaled to the measured 0.68 ms CPU per request, allocations from gc-verbose):

• IHttpClientFactory and HttpClient layers: about 0 µs and 0.7 KB per request in the base (the header propagation middleware and RedirectHandler). The transport removes all of it.
• Inside SocketsHttpHandler: about 28% of CPU samples, including send and receive syscalls, and 6.9 KB per request (27% of all gateway allocations). That splits into header collections 2.4 KB, HttpConnection 2.1 KB, pool and handler continuations 1.4 KB, and request/response messages with content 1.1 KB.
• Fusion request building and response parsing: about 6% of CPU and 8.1 KB per request (32%).
• A microbenchmark (per subgraph call, HttpMessageInvoker with these handler settings vs a raw-socket HTTP/1.1 floor) puts the most a Fusion-owned HTTP/1.1 client could still save at 3.7 to 3.8 µs per call on JIT and 5.5 to 6.7 µs per call on AOT, plus 1.8 KB per call. With 4 subgraph calls per request here, that is about 15 µs (JIT) or 23 µs (AOT) per request, 2 to 3% of gateway CPU, and about 7 KB of allocations per request.

Configuration needs found by the prototype

The prototype ignores every AddHttpClient("fusion") setting, so each of these needs a home in the configuration API:

• Per source schema handler settings: MaxConnectionsPerServer (the benchmark used 256), ConnectTimeout, PooledConnectionLifetime (DNS refresh, replaces handler rotation), PooledConnectionIdleTimeout, EnableMultipleHttp2Connections, HTTP/2 keep-alive pings. These settings also decide whether a schema reload keeps a handler.
• TLS: client certificates for mTLS, a custom CA or certificate validation callback, SslProtocols.
• HTTP version: request Version and VersionPolicy (HTTP/2 or h2c to source schemas). Today these come from HttpClient.DefaultRequestVersion, which needs ConfigureHttpClient and that turns off client sharing.
• Delegating handlers: token injection for source schema auth, AddRequestDeduplication, resilience handlers, logging. The prototype has no handler pipeline.
• Header forwarding: which gateway request headers go to which source schemas, with renames and value factories (what AddHeaderPropagation offers). The prototype reads a comma-separated list from AppContext.
• Timeouts: requests without an execution deadline (deferred fragments, subscriptions) keep HttpClient's 100 s time-to-headers bound, which is not configurable per source schema; a per source schema request timeout is missing.
• Proxy: UseProxy stays at its default, so HTTP_PROXY, HTTPS_PROXY and ALL_PROXY still reroute source schema calls. YARP turns it off and configures a proxy per cluster.
• Decompression: off. Anyone who relies on compressed source schema responses needs AutomaticDecompression as an opt-in. Setting it, even to None, makes NativeAOT link the Brotli and Zstandard decoders (1.1 MB).
• Test and in-process hooks: tests route source schema calls to TestServer by replacing IHttpClientFactory. The transport needs a way to supply the handler (like YARP's IForwarderHttpClientFactory), and ConnectCallback would cover Unix domain sockets to sidecars.
• Telemetry: ActivityHeadersPropagator, and the System.Net.Http.HttpClient.<name> log categories of IHttpClientFactory, which disappear.

Open questions

• UseProxy: keep the default (environment proxies apply) or turn it off like YARP.
• Replaced handlers are not disposed, so in-flight requests finish and idle connections close through the idle timeout (YARP does the same). Is that acceptable for frequent reloads?
• [Fusion] Build subgraph request bodies with fewer allocations #80 (SendJsonAsync) adds a second send path, which would also need to call the header forwarder.