[azcosmos] tighten default HTTP client timeouts and increase pool size

[tvaron3]

Cosmos accounts are reachable from the preferred region in well under a second and have a server-side request budget of ~60s, but the SDK has been inheriting azcore's general-purpose HTTP defaults (30s dial timeout, unbounded http.Client.Timeout, 100/10 idle pool). That means a dead endpoint blocks every request for 30s, a runaway request can hang a caller indefinitely if the caller forgot to set a context deadline, and any workload with more than ~10 concurrent in-flight requests per host forces needless TCP+TLS reconnects. This PR installs Cosmos-tuned defaults to fail fast on dead endpoints, backstop runaway requests, and sustain reasonable concurrency.

Defaults changed

Setting	Old (azcore default)	New
Connect (dial) timeout	30s	5s
http.Client.Timeout (per HTTP attempt)	unbounded	65s
MaxIdleConns / MaxIdleConnsPerHost	100 / 10	1000 / 100
HTTP/2 ReadIdleTimeout / PingTimeout	10s / 5s	2s / 1s

About the 65s http.Client.Timeout

This is a wall-clock cap on a single HTTP attempt (dial + request write + header read + body read). It is intentionally set as http.Client.Timeout so it survives custom per-call policies and acts as a hard safety backstop. Trade-offs to know:

• A caller-supplied context.WithTimeout shorter than 65s still wins; longer deadlines are truncated by the HTTP client. This is the intended safety property: no Cosmos request should hang for an unbounded amount of time even if the caller forgot to set a deadline.
• The azcore retry policy sits above the transport, so the 65s cap applies per attempt; the policy can still issue additional retries when one attempt exceeds the cap.
• 65s was chosen to slightly exceed the Cosmos gateway's server-side request budget (~60s) so the server has a chance to return a structured error before the client gives up locally.
• Callers that legitimately need to drain very large query/change-feed pages in a single attempt should supply their own Transport via azcore.ClientOptions.

About the idle pool sizing

MaxIdleConnsPerHost = 100 is well below the per-process FD ceiling on typical multi-host deployments while comfortably handling normal Cosmos concurrency. The total cap of 1000 prevents unbounded pool growth across many hosts (e.g., GEM-routed replica endpoints) on multi-region accounts.

Approach

• New file cosmos_default_http_client.go builds a package-level defaultCosmosHTTPClient that mirrors the azcore default transport (TLS 1.2+, ForceAttemptHTTP2, keep-alive 30s, etc.) but overrides the settings above.
• New build-tagged files cosmos_default_http_client_dialer_{other,wasm}.go mirror azcore's WASM/wasip1 handling so the Cosmos defaults remain portable across build targets.
• New helper withDefaultTransport() in cosmos_client.go is invoked exactly once at the top of NewClient and NewClientWithKey (NewClientFromConnectionString delegates to NewClientWithKey). The normalized options are reused for both the user-facing pipeline and the global-endpoint-manager pipeline. The helper installs the default only when the caller has not supplied a Transport, and shallow-clones the caller's ClientOptions before assigning so the top-level Transport field never leaks back to the caller. Slice fields (PreferredRegions, PerCallPolicies, etc.) still share backing arrays with the caller, as is conventional for Go config structs.
• Caller-supplied transports (including the mock.Server used in tests) flow through untouched.

Validation

• go build ./..., go vet ./..., and go test -count=1 -short ./... in sdk/data/azcosmos pass locally and on CI (ubuntu/windows, go 1.25/1.26).
• New unit tests cover: default timeout constants are wired into the client, nil options get the default, non-nil options without a Transport get the default without mutating the caller, and a caller-supplied Transport is preserved.

Checklist

• The purpose of this PR is explained in this or a referenced issue.
• The PR does not update generated files.
• Tests are included and/or updated for code changes.
• Updates to module CHANGELOG.md are included.
• MIT license headers are included in each file.

[Copilot]

Pull request overview

Adds Cosmos-specific HTTP transport defaults (timeouts + connection pool sizing) instead of inheriting azcore’s general-purpose defaults, and wires these defaults into Cosmos client construction only when callers haven’t provided a custom transport.

Changes:

• Introduces a package-level default http.Client configured with Cosmos-specific dial/request timeouts, idle pool sizing, and HTTP/2 ping settings.
• Routes newClient and newInternalPipeline through a helper that applies the default transport without mutating caller-supplied ClientOptions.
• Adds unit tests and a CHANGELOG entry documenting the new defaults.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 2 comments.

File	Description
sdk/data/azcosmos/cosmos_default_http_client.go	Adds Cosmos-specific default HTTP client/transport configuration.
sdk/data/azcosmos/cosmos_default_http_client_test.go	Adds tests for default transport wiring and non-mutation behavior.
sdk/data/azcosmos/cosmos_client.go	Applies default transport logic in client/pipeline creation via withDefaultTransport().
sdk/data/azcosmos/CHANGELOG.md	Documents the default HTTP client behavior changes.

[kushagraThapar]

Thanks for this — really clean implementation and the tuning aligns well with peer SDKs (cross-checked against .NET / Java / Python / Rust and the 65s / 5s / pool numbers all match up). Have one thing I'd love to discuss before merge, dropped inline on the new 65s constant. Curious to hear your thinking!