The GitHub Actions job "Build CI with different platforms/configs" on hive.git/HIVE-30105 has succeeded. Run started by GitHub user deniskuzZ (triggered by deniskuzZ).
Head commit for run: ebaea41992f35915b1915d82f7eaecfb3f00bd2a / Denys Kuzmenko <[email protected]> HIVE-30105: Parquet: decode primitive columns a batch at a time - ParquetVectorUpdater/ParquetVectorUpdaterFactory replace the per-value virtual-dispatch reader: definition levels and values decode a batch per call for PLAIN and dictionary pages (int, bigint, float, double, decimal, date, boolean, char, varchar, string, binary) - A dictionary is flattened once per column chunk; a string row becomes a reference into it instead of a copy - VectorizedPrimitiveColumnReader drops BaseVectorizedColumnReader and implements VectorizedColumnReader directly; HIVE-30059's ParquetRowGroupDecoder (which constructs it) needed no change, since it already typed its array as the VectorizedColumnReader interface - decimal64 fast path; single-string-key fast path (dictionary-aware) - BIT_PACKED definition-level values keep the per-value path; timestamps keep it too (HIVE-30116) - a batch of equal values is repeating only when their bits are equal: 0.0 and -0.0 compare equal as doubles, and the per-value reader read a PLAIN batch of both as one repeated value, every row taking the first one's sign - tests: TestVectorizedColumnReader and TestVectorizedDictionaryEncodingColumnReader read every type through both paths; TestVectorizedRleValuesReader, TestDictionaryValues; parquet_signed_zero.q: 2,000 rows alternating 0.0 and -0.0, DOUBLE and FLOAT, PLAIN and dictionary pages, vectorized equal to row mode (master: vectorized PLAIN reads 2,000 x -0.0); parquet_vectorization_0/not.q.out: last digits of variance and stddev (summation order) - Impact (local, 20M store_sales-shaped rows): decode 72.2 -> 15.7 ns/row (4.6x), full reader 89.9 -> 28.3 ns/row (3.2x) Report URL: https://github.com/apache/hive/actions/runs/43081519735 With regards, GitHub Actions via GitBox --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
