Repository navigation
Simplify / speed up implementation of character_length to unicode points #3049
Description
Is your feature request related to a problem or challenge? Please describe what you are trying to do.
It looks like postgresql and spark have a different, (probably much faster) implementation of calculating the string length, which doesn't depend on calculating the grapheme clusters, but on utf8 code points.
As an example: select length('ä')
PostgreSQL | Spark | DataFusion
| Postgres | Spark | DataFusion |
|---|---|---|
| 2 | 2 | 1 |
Describe the solution you'd like
Probably .chars().count() is a faster solution and matches the implementation of PostgreSQL (and Spark).
It can be applied to other string functions as well.
Describe alternatives you've considered
Accepting our own implementation being slower but "superior" to other solutions.
Additional context
This came up as being quite slow when profiling this benchmark: https://github.com/DataPsycho/data-pipelines-in-rust/tree/main/amazon_review_pipeline
We're using the grapheme option in other places as well, maybe here we could use the faster
Activity
andygrove commentedon Aug 6, 2022
This is something that we could potentially add a configuration option for as well, defaulting to the fast version but allowing the user to select a slower and more accurate version.
Dandandan commentedon Aug 6, 2022
In general, I think it would be better to choose one version of the algorithm rather than have the result depend on the configuration.
As we aim to be mostly compatible with PostgreSQL, it makes sense to me to choose counting the code points instead of grapheme clusters.