Bound Cosmos diagnostics under retry storms

sdk/cosmos/azure_data_cosmos_driver/CHANGELOG.md

@@ -9,6 +9,7 @@
 - Added a `.NET CosmosDiagnostics`-style top-level **`summary`** block to `DiagnosticsContext` (`DiagnosticsSummary`, reachable via `DiagnosticsContext::summary()`). It aggregates over the collected requests — `(status, sub-status)` histogram, total request charge, request/retry/throttle counts, regions contacted, final status, top error, total elapsed, and the client **`user_agent`** (SDK + runtime identity, via `DiagnosticsSummary::user_agent()`) — and is **computed lazily on first access** (never on the hot path) and cached via `OnceLock`, so a context whose diagnostics are never inspected pays nothing for it. Serialized as the first section of the diagnostics output; the `user_agent` is omitted when the driver did not supply one, so existing output is unchanged.
 - Added a **`DiagnosticsEncoding`** client option on `DriverOptions` (`DriverOptionsBuilder::with_diagnostics_encoding` / `DriverOptions::diagnostics_encoding`) controlling how diagnostics render to a string via the new `DiagnosticsContext::encode(encoding)` and `CosmosResponse::diagnostics_string(encoding)`: `Json` (pretty, **default** — output unchanged), `Compact` (minified JSON), and `Encoded` (base64 of the compact JSON, decodable, for size-sensitive logging). Additive and non-breaking.
 - Added serializable **wall-clock timestamps** (RFC 3339 / ISO 8601) to the diagnostics output. `DiagnosticsContext` (and its `summary`) now carry an operation `start_time`/`end_time`, and each `RequestDiagnostics` attempt carries a `start_time` plus an optional `end_time`, captured with `azure_core::time::OffsetDateTime` alongside the existing `Instant`-based durations. These absolute timestamps let diagnostics be correlated against server-side and external logs. Additive and non-breaking: the fields exist only when a context is built, durations are unchanged, and an in-flight attempt's `end_time` is omitted (`skip_serializing_if`).
+- Added **retry-storm compaction** to the always-on `DiagnosticsContext`, giving a bounded-size guarantee under 4xx/5xx retry storms. When an operation accumulates more per-attempt records than the new configurable `DiagnosticsOptions::max_request_diagnostics` cap (`DiagnosticsOptionsBuilder::with_max_request_diagnostics`, env `AZURE_COSMOS_DIAGNOSTICS_MAX_REQUESTS`, default `512`, min `16`), runs of near-identical consecutive retries (same region / endpoint / status / sub-status / execution-context) are collapsed at operation finalization into first + last + a count, with an order-robust global key-bucket fallback that keeps the retained count within the cap even under a region ping-pong. The aggregate `summary` is computed from the full attempt list *before* compaction, so the `(status, sub-status)` histogram, retry/throttle counts, total RU and regions stay exact; `DiagnosticsContext::request_count()` still reports the true total, and new accessors `DiagnosticsContext::retained_request_count()` and `DiagnosticsContext::compaction()` (returning the new `CompactionInfo` / `CompactedRun`) expose the bounded retained count and a per-run rollup. Detailed JSON gains a top-level `compaction` object only when compaction fires (`skip_serializing_if`), so output stays byte-identical for normal operations. Additive and non-breaking.
 - Restructured the client / runtime options layering on the driver. Two new nested option groups, a per-client overrides surface on `DriverOptionsBuilder`, and a single canonical `AZURE_COSMOS_PPCB_*` namespace for partition-failover environment variables. The driver now consumes partition-failover configuration once at construction (`CosmosDriver::new` no longer fabricates an `OperationOptionsView` outside any operation context) ([#4588](https://github.com/Azure/azure-sdk-for-rust/pull/4588)):
 - Added new nested `OperationOptions::throughput_control` group (`ThroughputControlOptions` / `…Builder` / `…View`, mirroring the `ThrottlingRetryOptions` pattern). Exposes three layered fields ([#4588](https://github.com/Azure/azure-sdk-for-rust/pull/4588)):
   - `group_name: Option<ThroughputControlGroupName>` — replaces the old top-level `OperationOptions::throughput_control_group`.

sdk/cosmos/azure_data_cosmos_driver/Cargo.toml

@@ -136,6 +136,11 @@ name = "multi_region_failover"
 path = "tests/multi_region_failover.rs"
 required-features = ["fault_injection"]
 
+[[test]]
+name = "live_storm_diagnostics"
+path = "tests/live_storm_diagnostics.rs"
+required-features = ["fault_injection"]
+
 [[test]]
 name = "gateway_query_plan_comparison"
 path = "tests/gateway_query_plan_comparison.rs"

sdk/cosmos/azure_data_cosmos_driver/src/diagnostics/capture/mod.rs

@@ -107,6 +107,7 @@ pub use recorder::{AttemptRecord, DiagnosticsRecorder, HedgeOutcome};
 // `crate::diagnostics` so the public boundary (`diagnostics::DiagnosticsContext`, consumed by the
 // `azure_data_cosmos` SDK) is unchanged.
 pub(crate) use model::DiagnosticsContextBuilder;
+pub use model::{CompactedRun, CompactionInfo};
 pub use model::{
     DiagnosticsContext, DiagnosticsSummary, ExecutionContext, FailedTransportShardDiagnostics,
     PipelineType, RequestDiagnostics, RequestEvent, RequestEventType, RequestHandle,