Status: owner-selected session contract, 2026-10-17; implementation and exact-source Linux qualification remain open. ADR-142 and SessionExecution supersede the previous active per-scenario workers, warmup namespaces and fresh repetition resets. Native one-node/three-node topology and the fairness/correctness requirements of ADR-122 remain applicable. Original historical reports retain their actual settings and are never relabelled.
Each database has two isolated Linux measurement jobs: native one node and native three nodes. Different databases/topologies never share a runner, containers, volumes or measurement session. One Aspire-owned deployment lives for the entire job. Many separate real native TUnit tests share its fixture; only one workload test executes at a time, with the declared concurrent clients inside that test. The common preparation builds the runner and images once; each worker imports and verifies the same-source binaries, native tools and image identities.
flowchart LR
Ready[Aspire readiness and actual membership] --> Load[Measure initial million writes]
Load --> Index[Measure native index construction]
Index --> Operations[Separate read search and query tests]
Operations --> Mixed[Concurrent create delete update read search]
Mixed --> Verify[Full final content and native replica verification]
Verify --> Original[Original per-test reports samples and TRX]
Original --> Aggregate[Authenticate all database and topology jobs]
The initial 1,000,000-record load is a measured native test. Setup contains no hidden corpus insertion. There is no warmup and no per-case reset, cleanup, reseed or restart. The state of the runtime, cache, storage and indexes evolves naturally during the declared operation order; this does not claim cold-cache measurements. Native index construction and settlement have their own measured result where the database exposes an equivalent operation. Intrinsic index maintenance remains in measured writes. Missing native capabilities are explicit unavailable results with no numeric substitute.
Initial load, operation lanes and the wall time of a concurrent mixed test are separate measurements. Readiness and full correctness/replica verification are outside operation latency. Actual caller-native acknowledgement defines write completion. No extra read is silently included in a measured write. A native batch acknowledgement records its real batch size and duration; it cannot be expanded into invented per-record latency samples.
The deterministic corpus, seed, payload, key groups, query predicates, requested counts and client limits are identical across applicable adapters. The initial corpus has 1,000,000 records. The production session uses 16 bounded clients; operation tests perform at least 100,000 original requests per applicable lane. One client has one outstanding original call unless an explicitly equivalent native batching/pipelining contract is part of the shared case.
| Test | Measured work and correctness contract |
|---|---|
| Initial load | 1,000,000 distinct actual writes, followed by independent complete identity/body verification. |
| Index | Actual native build/settlement on the retained corpus; unavailable when no equivalent native lifecycle exists. |
| Sequential read | 100,000 stable-group keys in identity order, validating returned content. |
| Random read | 100,000 seed-bound stable-group keys in deterministic order, validating returned content. |
| Search/query | 100,000 equivalent native predicates with bounded results and exact content oracles. Range/complex-query expansion remains separately tracked. |
| Mixed | Concurrent lanes: 100,000 creates, 100,000 deletes, 100,000 updates, 100,000 reads and 100,000 search requests. Each lane retains its actual outcomes and timestamps. |
| Final verification | Exactly 900,000 retained original keys plus 100,000 new keys; correct replacements and absence of deleted keys, with actual native replica proofs. |
Delete, update and stable-read key groups are disjoint. Creates use new keys. The mixed test starts admitted native clients through a common barrier and proves observed overlap from original timestamps; it does not serialize the lanes and label them concurrent. The final cardinality remains 1,000,000. Count alone is insufficient: complete bounded readback verifies identities, bodies and mutation results. A million payload objects or 100,000 pending tasks must not be retained merely to schedule the work.
The full product comparison still requires native vectors/search accuracy, graphs, queues, event streams, time series, ordered/range and complex queries, plus ingestion at 1/10/500 clients. These run as additional real cases on the same database/topology infrastructure. A seven-case document session does not qualify that complete inventory. Unsupported native functionality is distinguished from an unimplemented adapter or failed prerequisite; implementation gaps must not be reported as native absence.
Every family uses the same canonical scenario, corpus/seed, operation order, counts, metric/filter, accuracy target, resource limits and effective durability contract across applicable databases. Vector indexes retain exact/HNSW/IVFFlat, filtered/concurrent-update and recall requirements where natively supported. Native batching, index build and accuracy are declared explicitly. Closed-loop request timings do not prove scheduled-arrival/open-loop or endurance behavior.
The existing exclusive functional RF3 million-record ingestion gate remains mandatory and separate from numerical benchmark publication. Its actual .NET SDK and official MCP callers operate while ingestion is active. Heavy-load admission still requires explicit enabled settings, the exact native class filter, one-test parallelism and disabled coverage; missing/false/broad/duplicate filters, coverage or wrong parallelism fail before resources are acquired. Normal functional concurrency and all process-recovery/RF3 gates are unchanged.
Retain attempted, completed, successful, failed, cancelled and unfinished counts, actual timestamps, useful completed throughput and per-operation p50/p95/p99. Counts and latency distributions must be reproducible from bounded original full-count samples, including failures; do not average percentiles or choose a best repetition. The mixed wall time is not the sum of concurrent lane times. Client CPU/allocations/working set and server per-node CPU/RAM observations stay separate. Missing instrumentation is explicit and cannot become a zero result.
Record exact source/run/attempt/job/artifact identities, Linux/runtime/CPU, DLL/PDB hashes, native images, observed membership/data copies, resource limits, filesystem, transport, read consistency, index settings and effective ACK/durability. A node count alone does not establish equivalent guarantees. RF3 majority remains two acknowledgements. The one-node KeyLoad arm is explicitly RF1; production RF3 is unchanged. Unequal contracts prohibit a performance-winner claim.
Cancellation stops admission, cancels and joins original calls, then disposes native clients and the Aspire deployment. Preserve the primary and shutdown failures and completed original reports. A failed load poisons dependent tests; they retain real failed prerequisite outcomes rather than invented measurements. No automatic measurement retry repairs the statistics. No skipped, missing, corrupt or mixed-cohort original is accepted as a complete session.
Aggregation authenticates all 22 database/topology jobs, their exact planned native cases, original reports and TRX, same-source images, compiled-runtime identity/import receipts and statistics. Actual failed cases retain null metrics and original provenance. Website independently qualifies its source, archives, real browser and coverage before rendering these results; content-only publication contains no fabricated measurements when no eligible cohort exists.
REQ/AC-BSESSION-001..008 in SessionExecution own matrix, fixture, measured load, mixed work, shared build, no warmup/reset, authenticated accounting and exact-source Linux completion. Authored source, controlled rejection inputs, an upstream benchmark or a local build closes none of the performance gates. All benchmark-stage builds, checks, tests and workloads run in Linux GitHub Actions. Full model-family completion, actual native execution, fault/endurance, resource fairness and website qualification remain open until their original evidence exists. No power-loss or production-readiness claim follows from process-kill tests or this redesign.