Skip to content

AdditionalDatabases.md

Latest commit

 

History

History
121 lines (104 loc) · 8.6 KB

AdditionalDatabases.md

File metadata and controls

121 lines (104 loc) · 8.6 KB
Raw

SQL and analytical comparison targets

The owner authorized TimescaleDB, SQL Server, MySQL, ClickHouse and InfluxDB as benchmark targets on 2026-10-24. They extend the current session cohort to sixteen systems and 32 one/three-node rows. The benchmark projects alone own the drivers, native operation adapters and containers. KeyLoad's server, SDK and storage do not adopt these engines or their drivers.

Implemented operation contracts

Target Native storage and transport Current session operations
TimescaleDB A real hypertable through Npgsql; deterministic per-document timestamps and complete text payloads Load, numeric index, point reads, compound predicates, update/delete, mixed CRUD and ordered final readback
SQL Server A unique database, binary-collated primary key and parameterized Microsoft.Data.SqlClient commands The complete document session; delayed durability must be disabled
MySQL A unique database, InnoDB table, binary-collated primary key and parameterized MySqlConnector commands The complete document session; innodb_flush_log_at_trx_commit=1 and sync_binlog=1 are checked
ClickHouse A unique database, MergeTree and bounded native HTTP queries The complete document session; acknowledged asynchronous inserts wait for actual flush, inserts request filesystem synchronization, and native update/delete mutations wait for completion
InfluxDB A unique database/table, file object store, native line protocol and parameterized SQL over HTTP Load, point reads, compound predicates and full final readback. Tags retain document identity; fields retain the exact body and number. A deterministic timestamp identifies an update to the same point. Core has no row delete operation, so mixed CRUD is unavailable. Index construction is unavailable because Core manages its indexes during ingestion.

These document sessions do not qualify time-window aggregates, analytical scans, vector search or any other model family. The existing separate TimescaleDB TimeSeries targets and their open scale requirements under ADR-050/059 remain distinct. A hypertable document measurement cannot substitute for those results.

Each session still measures its initial million records and at least 100,000 operations in each applicable lane. Complete ordered readback verifies actual identities and bodies. InfluxDB's final readback verifies the unchanged corpus after its supported operations; it does not invent a mixed mutation result.

InfluxDB complete readback executes one ordered SQL query and consumes its native streaming JSONL response. It does not repeat the query for every 256 records. The framework pipe reader retains bounded buffers; MaxResponseBytes bounds each JSON row, and the frozen corpus identity capacity bounds the total row count. The existing operation deadline bounds response headers and each read's idle wait; caller cancellation also stops buffered delivery. The independent ordered oracle still checks every identity, full body and digest. The native million-record control also exercises an update, sixteen independent concurrent point reads after the complete readback, cancellation during readback and a healthy write and read afterward. The vendor documents JSONL as its streaming HTTP query format.

The typed deployment policy assigns InfluxDB's query execution pool 70% of the existing available memory, following the vendor's mixed-workload recommendation. This changes the engine's internal pool ceiling, not the runner's CPU or RAM. The native million-record check retains failures from sorting or deduplication; query memory errors remain failures.

InfluxDB snapshots each completed WAL file through its native wal-files-per-snapshot setting. The default 600-file interval otherwise leaves this short million-record load in the mutable buffer: the pinned engine materializes that buffer while planning each query, so sixteen point reads can simultaneously copy the complete corpus. Native snapshots make Parquet available during the load and bound this planning pressure. The line protocol, no_sync=false acknowledgement, independent clients, operation deadlines and complete result checks stay the same.

Deployment and qualification limits

The added native deployments currently implement one node. TimescaleDB, SQL Server, MySQL and ClickHouse three-node rows are explicitly unavailable because this benchmark does not yet implement or qualify their replicated deployment and acknowledgement contracts. This is an implementation gap, not a claim that these databases lack replication. InfluxDB 3 Core is a single-node product; Enterprise clustering is outside this comparison. Separate standalone servers never count as replicas.

Aspire owns the official pinned container images, anonymous session volumes, readiness and joined cleanup. Native SQL readiness opens a real provider connection and executes SELECT 1; HTTP targets use their server health endpoints. The complete Benchmarks workflow uses the Ubuntu 22.04 runner family so preparation and every worker share one declared OS baseline. Its published kernel is 6.8; MongoDB's native startup guard rejects kernels 6.19 through 7.0.13 because of its TCMalloc dependency. Preparation and worker logs record the actual host and Docker kernels. The pinned MongoDB image and its startup guard are unchanged. Actual kernel, CPU and memory evidence still determines resource equivalence; a common runner label alone does not qualify figures. See the runner image inventory and MongoDB's kernel requirement. SQL Server uses Developer edition for nonproduction benchmarks and requires the Linux x86-64 runner. Its local execution on ARM is not qualified.

AdditionalDatabaseTests exercises real native operations, complete-body updates, missing reads, independently owned clients, cancellation followed by healthy reads, supported deletes, compound predicates and ordered readback through a separate Aspire lifecycle. The Linux one-node workers run it before starting the fresh measured session. These bounded correctness cases provide no performance figures. Complete Debug build, actual native tests, Linux million-record sessions, resource observations and authenticated aggregate/site admission remain separate open gates until their original results exist. Local development verification on 2026-10-24 passed the complete Debug solution build with warnings as errors, formatter verification, governance, nine parser/admission scenarios (131 altered inputs rejected) and four genuine one-node adapter cases: TimescaleDB, MySQL, ClickHouse and InfluxDB. SQL Server execution remains pending on Linux x86-64; these small correctness cases do not qualify million-record performance or replication.

Driver adoption

The owner explicitly approved benchmark-only Microsoft.Data.SqlClient 7.1.1 and MySqlConnector 2.6.2 after review of versions, original MIT licenses, parameterized ADO.NET APIs, asynchronous cancellation and disposal. Both versions are centrally pinned and referenced only by KeyLoad.Comparisons. Native pools and command deadlines use the existing validated comparison execution policy. ClickHouse and InfluxDB use the existing bounded HTTP transport; TimescaleDB reuses Npgsql.

Publisher sources:

The benchmark project explicitly references System.Security.Cryptography.Pkcs at its approved central pin after restore selected vulnerable transitive version 9.0.18 and failed NU1902/NU1903. This explicit reference keeps the selected SqlClient dependency graph on the fixed Microsoft MIT package. Audit and warning gates stay enabled.