A cold tier 23% smaller than Parquet — still queryable, still recoverable
AT-1 re-encodes a Parquet table 23% smaller than Parquet-zstd, byte-for-byte value-exact, and keeps it block-addressable for predicate/projection queries — a drop-in cold tier, not a re-platforming.
Who it's for: For data-engineering teams on Snowflake / Databricks / Iceberg with large, rarely-scanned Parquet partitions.
How it fits: Tier cold partitions from Parquet to AT-1 (a typed Parquet→AT-1→Parquet round-trip reconstructs the exact column values); hot partitions stay in the warehouse, cold ones shrink in the lake.
Measured on a real NYC-taxi table (2,964,624 rows × 19 cols): AT-1 36.3 MB vs Parquet-zstd 47.0 MB — 23% smaller, byte-for-byte lossless, and smaller than Parquet-gzip's 44.5 MB too.
Tables full of totals and rolled-up counters shrink further with the derived-column eliminator — it stores the rule (total = a + b) instead of the column, byte-exact. Real NetFlow: 8.27% smaller whole-file.
Put a number on your data.
Enter your storage bill; see the saving and a one-page proposal in 30 seconds.