[Default Read/Write Timeout 3/N] Bake per-service exemption tiers and apply the rollout gate

[zoewangg]

Motivation and Context

Increment 3 of N for the default read/write inactivity timeout on the AWS CRT-based HTTP clients. This increment targets the feature branch feature/master/2026-enable-default-read-timeout, on top of increment 1 (the AWS_ENABLE_DEFAULT_READ_TIMEOUT_2026 rollout gate, #7319) and increment 2 (the CRT client applying the resolved SdkHttpConfigurationOption.SDK_INTERNAL_FALLBACK_READ_WRITE_TIMEOUT, #7326).

Not every service should get the same default. Some services stream or hold connections open for long periods, so a single flat inactivity timeout would break them. This increment introduces the per-service exemption tiers and wires the rollout gate to the resolved value:

• Fully-exempt services get no default read/write timeout.
• Partially-exempt (long-hold) services get a longer 15-minute timeout.
• Every other service gets the general 5-minute default.

The per-service tier is independent of the rollout gate and is resolved at code-generation time (service identity does not reach the HTTP client). The gate is then applied once, above the HTTP client, so the value handed is already gate-resolved.

Modifications

Code generation (codegen):

• New DefaultReadWriteTimeoutExemptionProcessor (registered in DefaultCustomizationProcessor, alongside LongPollingOperationProcessor) reads a checked-in exemption artifact (default-read-write-timeout-exemptions.json, keyed by service sdkId) and bakes each listed service's tier onto the intermediate model: -1 marks a fully-exempt service, a positive value is the timeout in milliseconds.
• Metadata carries the baked tier (defaultReadWriteTimeoutMillis); BaseClientBuilderClass emits it into the generated serviceHttpConfig() as SDK_INTERNAL_FALLBACK_READ_WRITE_TIMEOUT (Duration.ZERO for fully-exempt, Duration.ofMillis(...) for a positive tier), composing with any existing service HTTP config (e.g. HTTP/2). Services absent from the artifact bake nothing.
• The processor also exposes a whole-artifact key validation used by the coverage test to fail the build if any artifact key matches no real service sdkId (guards against a stale or mis-cased key silently leaving a service unlisted). Malformed artifact values fail with the offending key and resource named.

aws-core:

• AwsDefaultClientBuilder.resolveHttpClientConfig applies the rollout gate to the codegen-baked value. The gate is resolved via the AWS_ENABLE_DEFAULT_READ_TIMEOUT_2026 environment variable / system property, else the codegen-baked SdkClientOption.DEFAULT_ENABLE_READ_TIMEOUT_2026 default, else off.
  • Gate on: a baked tier is kept as-is (Duration.ZERO fully-exempt, 15 minutes partial); an unlisted service (nothing baked) gets the flat 5-minute default.
  • Gate off: a baked positive tier (partial) is forced to Duration.ZERO so it cannot apply; otherwise the option is left untouched (an unlisted service stays absent, which is equivalent to Duration.ZERO because the HTTP client's option-absent path applies nothing when the gate is off; a fully-exempt tier is already Duration.ZERO).

The CRT client (increment 2) consumes this resolved value and does not re-evaluate the gate. Other HTTP clients ignore the option. No public API is added or changed.

Testing

• codegen module builds and tests pass. New tests: DefaultReadWriteTimeoutExemptionProcessorTest (each artifact key bakes its declared tier; case-sensitivity; every key cross-checked against the real service sdkId set; bogus-key guard fires), a full-class fixture-comparison test for the generated serviceHttpConfig() (partial tier composed with HTTP/2), and programmatic emission cases (fully-exempt Duration.ZERO, partial Duration.ofMillis, unlisted omitted).
• aws-core module builds and tests pass. New AwsDefaultClientBuilderReadWriteTimeoutTest parameterizes the gate/tier matrix (gate off + partial -> ZERO; gate on + unlisted -> 5 min; gate on + fully-exempt -> ZERO; gate on + partial -> 15 min; gate on via codegen default -> 5 min; gate property false overrides the codegen default -> ZERO) plus an explicit case that an unlisted, gated-off service leaves the option absent.
• Verified against real service models by regenerating samples: SQS (partial) bakes Duration.ofMillis(900000L), Polly (fully-exempt, HTTP/2) composes the HTTP/2 config with Duration.ZERO, STS (unlisted) bakes nothing.

Types of changes

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)

Checklist

• I have read the CONTRIBUTING document
• Local run of mvn install succeeds (built and tested the affected modules: codegen and aws-core)
• My code follows the code style of this project
• My change requires a change to the Javadoc documentation
• I have updated the Javadoc documentation accordingly
• I have added tests to cover my changes
• All new and existing tests passed
• I have added a changelog entry (deferred to the feature-branch finalization; individual increments on the feature branch do not each add a changelog)

License

• I confirm that this pull request can be released under the Apache 2 license

[Fred1155]

If the http client is passed in directly, like S3AsyncClient.builder().httpClient(AwsCrtAsyncHttpClient.create())? ServiceHttpConfig() never reaches it in that case, so all service would fall back to the 5 min behavior.

[Fred1155]

Just curious: do we intend to remove this processor and the json artifact at the end of the rolling out?