Benchmark

Real measurements, with their conditions. They describe how Canot behaves, not how a platform behaves: throughput depends on the network, the machine and the source warehouse.

TPC-H SF10 from Snowflake to a local disk (2026-10-17)

Source	Snowflake, trial account, TPC-H SF10 sample dataset, XSMALL warehouse
Destination	local folder (Apache Iceberg, 2.8 GB of Parquet files)
Volume	8 tables, 86,586,082 rows, about 10.6 GB read (largest table: 59,986,052 rows, 7.3 GB)
Result	8 tables out of 8 matching at the full_aggregates level (row counts and per-column statistics recomputed in the source engine; content fingerprint of the rows as Canot read them against the copy; see proof levels); 2 queries out of 2 with the same result
Total duration	31 min 36 s (copy, comparison and replay included)
Copy throughput	about 7 MB/s, limited by reads from the trial warehouse
Memory	2.4 GB peak on the machine running the exercise (MacBook Air), for 10.6 GB of data: the comparison reads the copy in batches
Consumption	about 0.41 Snowflake credits
Version	Canot 0.1.0 + batch verification fixes, DuckDB 1.5.6, pyiceberg 0.12

Known limits: query replay loads the tables the queries touch into memory (--replay-max-gb option, 2 GB by default); here, only queries on small tables were replayed. No test beyond 10 GB.