Skip to content

Latest commit

 

History

History
85 lines (69 loc) · 4.39 KB

contenders.md

File metadata and controls

85 lines (69 loc) · 4.39 KB

Contender registry

Active means included in standard runs for useful ongoing comparison. Candidate means evaluated in a spike before a retention decision. Eliminated means excluded from routine smoke, local and CI execution. This is a documentation registry, not runtime configuration.

Contender Status Coverage
Scriptella batch 1000 Active PostgreSQL, MariaDB
Handwritten JDBC batch 1000 Active PostgreSQL, MariaDB
Pandas + SQLAlchemy Active PostgreSQL, MariaDB; retained after the v5 candidate spike
PostgreSQL COPY Active native reference PostgreSQL only; not a portable ETL competitor
MariaDB LOAD DATA LOCAL INFILE Active native reference MariaDB only; not a portable ETL competitor
Scriptella / JDBC unbatched Retired Last published v4; early v5 spike preserved separately
csvsql chunk 1000 Eliminated Last standard version: v4

csvsql: eliminated for v5

The csvkit 2.2.0 direct importer used inferred types, SQLAlchemy 2.0.46, psycopg2-binary 2.9.11 and PyMySQL 1.1.2. The workload was the shared deterministic million-row fixture and pre-created schema on PostgreSQL 17 and MariaDB 11.

Database Scriptella batch 1000 csvsql chunk 1000 csvsql / Scriptella
PostgreSQL 7.775 s 378.555 s 48.7x
MariaDB 6.694 s 370.594 s 55.4x

Last v4 CI evidence records the environment, raw measurements and full dependencies. At approximately 50× slower (48.7× PostgreSQL, 55.4× MariaDB) than Scriptella batching, this workflow costs substantial routine execution time without enough ongoing analytical value. This conclusion applies to the documented workflow and settings, not every possible csvsql configuration.

Invocation:

csvsql --db '<SQLAlchemy URL>' --tables csv_import --no-create --insert \
  --chunk-size 1000 <fixture>

For reproduction, check out benchmark revision eecd095b71b251e3e9e280fd45eb9da8fabf7b34 in a separate worktree and use its Dockerfile, complete requirements.txt, Compose configuration and benchmark guide. That revision retains the small runner-based implementation; v5 carries no csvkit/agate runtime dependencies. Historical v4 data and docs are unchanged. To reinstate, restore that revision's launch branch, dependency pins, case and metadata definitions into a new benchmark version, validate both databases, and document why it provides a useful comparison. Do not relabel v4 results.

Unbatched Scriptella/JDBC: retired from final v5

Batching is the established baseline for this million-row workload. Unbatched imports are substantially slower and add little to the continuing comparison. No default or optional local/CI execution includes them; run-all and smoke-all no longer exist, and the Java importer rejects batch size zero for Scriptella/JDBC. Batch size zero is reserved for native reference dispatch.

Last published evidence is v4 CI. The early eleven-case v5 spike (including its interrupted second run) is retained privately with exact source patches and raw measurements. It is incompatible with final eight-case v5. Last published durations (seconds; contextual, not pooled with final v5):

Database Scriptella unbatched / batch 1000 JDBC unbatched / batch 1000
PostgreSQL 50.363 / 7.775 46.308 / 6.320
MariaDB 57.374 / 6.694 54.498 / 4.465

The public v4 revision eecd095b71b251e3e9e280fd45eb9da8fabf7b34 retains the old implementations, requirements and invocation. To reproduce them, use that revision in a separate worktree and follow its benchmark guide. Scriptella 1.6, pgJDBC 42.7.13, MariaDB Connector/J 3.5.7 and Commons CSV 1.14.1 were used. To reinstate them, restore the old case definitions/importer paths into a new compatibility identifier and validate them; there is no runtime toggle in final v5.

Pandas: retained after the v5 spike

The ordinary workflow is slower than Scriptella batching in the local spike, but remains useful as a familiar, portable DataFrame-based reference. Its small implementation, explicit conversions and chunked input make it inexpensive to maintain. Retention is an analytical-value decision, not a claim of equivalent performance. The curated final v5 results record the comparison; earlier spike investigations remain private. The published article is unchanged.