You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
The zipped structural encoding is similar to the mini-block structural encoding in that we must first calculate the repetition and definition levels. However, the zipped structural encoding should be used when the data is wide (e.g. vector embeddings, non-trivial strings). In this case the number of values per mini block will be small (or less than 1 even) and the overhead of mini block encoding will be considerable.
Instead, we can take the repetition and definition information and convert it into a control word (1-2 bytes) that is attached at the beginning of each value (this is similar to the zip operation in python/rust). The limit between narrow and wide will need to be figured out but it is probably somewhere greater than 16 bytes per value and less than or equal to 128 bytes per value.
Zipped encodings will need to be dense and cannot be opaque. If a data type is wide then we should not apply any compressive encodings that will make it opaque (e.g. don't apply delta encoding to vector embeddings).
The text was updated successfully, but these errors were encountered:
The zipped structural encoding is similar to the mini-block structural encoding in that we must first calculate the repetition and definition levels. However, the zipped structural encoding should be used when the data is wide (e.g. vector embeddings, non-trivial strings). In this case the number of values per mini block will be small (or less than 1 even) and the overhead of mini block encoding will be considerable.
Instead, we can take the repetition and definition information and convert it into a control word (1-2 bytes) that is attached at the beginning of each value (this is similar to the zip operation in python/rust). The limit between narrow and wide will need to be figured out but it is probably somewhere greater than 16 bytes per value and less than or equal to 128 bytes per value.
Zipped encodings will need to be dense and cannot be opaque. If a data type is wide then we should not apply any compressive encodings that will make it opaque (e.g. don't apply delta encoding to vector embeddings).
The text was updated successfully, but these errors were encountered: