apache / parquet-cpp / 802

push

travis-ci

PARQUET-676: Fix incorrect MaxBufferSize for small bit widths

This function was implemented incorrectly in the original source code from Impala. The bug never presented itself because Impala allocates a much larger buffer than is required for the data page rather than using the `MaxBufferSize` value.

In a worst case scenario, the RleEncoder can result in a concatenation of short RLE runs of 8 values (the minimum for RLE). For example:

0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 0 0 0 0 0 0 0 0 ...

In encoded form, each RLE run occupies 2 bytes (for bit width 1). Thus 1024 values with this structure and bit width 1, at least 256 bytes are required to fully encode. If these were encoded as all literal runs, you can encode up to 504 values in a literal run (plus the indicator byte). Thus, in the same case (bit width 1) only

(504 + 504 + 16)
64 + 64 + 3 = 131

bytes are required. This should also fix PARQUET-698.

Author: Wes McKinney <wes.mckinney@twosigma.com>

Closes #150 from wesm/PARQUET-676 and squashes the following commits:

6ca44a8 [Wes McKinney] clang-format
e48b0bf [Wes McKinney] Fix maximum buffer sizing for hybrid RLE encoding case where short RLE runs use more space than mostly literal runs

20 of 20 new or added lines in 3 files covered. (100.0%)

8402 of 8649 relevant lines covered (97.14%)

11159.2 hits per line