Repository navigation
Conversation
|
@rymarm, thanks for finding this issue. I believe your fix is correct. The I would feel more confident if we had a unit test. That this bug exists suggests that, despite the rather excessive number of tests I wrote way back when, I somehow omitted this case. Perhaps a reason is that the bit vector is special: it uses code different than the other vectors, and so needs its own tests. Tests for the fixed-width vectors are in
|
DRILL-8511: Overflow appeared when the batch reached rows limit
Description
The size-aware scan framework fails to end the batch.
Framework tries to reallocate the vector on batch end, due to a hidden, minor bug in
BitColumnWriter- which in general is not notable, but in a specific case, when the initial vector allocation size limit is exceeded and a reader reaches the batch row size limit.BitColumnWriteruses instead of a write index a value count and this causes unexpected vector reallocation (look at the changes).Documentation
No changes required.
Testing
Manual tests