Compared to alternatives

Pramen is not a Flink/Spark replacement and not a warehouse. It competes in a narrower band: object storage → governed enrichment → operational PostgreSQL, crash-safe and budgeted. This page has two layers:

1. Orientation — when each alternative usually wins (qualitative).
2. Scoreboard — dated measurements; every number links a report.

Full prose: docs/compare/orientation.md.
Generated tables: docs/benchmarks/compare-scoreboard.md.

When to choose what

If your job is…	Prefer	Why
Enrich Parquet/NDJSON into Aurora/Postgres with budgets, reuse, batch pricing	Pramen	One binary, ledger + COPY delivery contract
Data already in Databricks/Snowflake/BigQuery and results stay there	Warehouse AI SQL	No extra hop
Broad connectors + light online AI in a streaming topology	Redpanda Connect	Catalog and ops model
In-process analytics, no governed LLM + PG contract	DuckDB / DataFusion	Less machinery
Stateful distributed streaming / huge joins	Flink / Spark / …	Different problem class

Honest caveat: warehouse vendors are investing heavily. Pramen’s wedge is strongest where residency, destination, model neutrality, or cost economics sit outside a single platform — see architecture §2.

Scoreboard (measured + harness-ready)

Offline legs regenerate on relevant merges (mise run compare-scoreboard). Competitor AI harnesses live under compare/ and stay harness_ready until a dated report lands.

PostgreSQL load path

From the v1 bench report (Apple M3 laptop — relative evidence):

System	Rows out/s	Notes
Pramen → PostgreSQL	434k–581k	~7× less CPU than DuckDB→PG on the same server
DuckDB → PostgreSQL	403k–620k	Wall-time tie (server-dominated); ~45 MiB RSS
DataFusion direct (no sink)	~4M	Engine ceiling
psql \copy CSV	117k	S1.3: Pramen binary COPY 3.1× faster

Semantic reuse (offline mock)

From RQ2:

Scenario	Result
Crash/replay	100% reuse; 0 tokens on replay
Batch crash reconcile	0 rebill
Duplicate-heavy (200/20)	90% savings vs naive

Not measured yet (harnesses ready)

Scenario	Harness
Redpanda Connect AI processor	compare/redpanda-connect/
DocETL extraction	compare/docetl/
Warehouse AI SQL	deferred (qualitative only)

Keeping numbers honest

Documented for contributors and agents in AGENTS.md:

• No public numeric claim without a report link.
• Offline scoreboard: regenerate when load path / ledger / bench / compare harness changes; CI --check fails on drift.
• Cloud competitor legs: env-gated, budget-capped — not PR-blocking.