Author here. This came out of reproducing a Databend issue about the same
class of bug, then I got curious where Snowflake's own limit sits since it's
not in the docs.
Two things I want to be upfront about:
1. It's a black-box measurement — inferred from partitions_scanned, not from
source. The exact byte boundary interacts with how you encode the tail
(left-padded serials shift the first significant digit), so read it as
"~40, order of magnitude," not a hard constant.
2. Almost certainly version-dependent (tested 10.25.101).
If anyone has a cleaner probe that nails the exact byte count, or works at
Snowflake and can say whether ~40 is intentional, I'd love to know.
1 comments