WIP fsst+ #4132
Draft
WIP fsst+ #4132
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Signed-off-by: Andrew Duffy <[email protected]>
|
Does fsst guarantee that no code is a prefix of another? I was always unsure about that part and lcp would be hard if it wasn’t the case |
|
Should we maybe put this in the fsst repo? |
|
The algorithm isn't really FSST-dependent AFAICT, it treats the input as opaque byte strings. The handling code seems to do this step after FSST-encoding all of the values. So I actually think this makes more sense in Vortex than in the core fsst crate |
yan-alex
reviewed
Aug 15, 2025
| } | ||
|
|
||
| /// Maximum shared prefix length. | ||
| pub const MAX_PREFIX: usize = 128; |
There was a problem hiding this comment.
The maximum prefix size should be 255 here :)
128 is the batch size, i.e. how many rows at a time the chunk_by_similarity() function would process.
But nice start, you guys are fast
5 tasks
a10y
added a commit
that referenced
this pull request
Oct 2, 2025
[FineWeb](https://huggingface.co/datasets/HuggingFaceFW/fineweb) is a string-focused ML training dataset provided by HuggingFace. This PR adds some data from the fineweb sample as a new benchmark. I hand-crafted some queries, also the benchmarks are quite fast, I'm only pulling a 2GB subset of the data so we might want to pull the full ~25GB sample. FineWeb contains data extracts from CommonCrawl, and has a number of string fields. We don't really have many string-filtering-heavy benchmarks and I think this can be a useful one to evaluate things like #4548 or #4132 --------- Signed-off-by: Andrew Duffy <[email protected]>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Was glancing through FSST+, the meat is the cleaving algorithm in https://github.com/cwida/fsst_plus/blob/main/src/cleaving/cleaving.h#L75-L159 which dictates the optimal point to split the sorted blocks by maximum prefix.
Far from complete. Also the test case is just for development ease, in actuality it seems that the cleaving happens after FSST encoding, not before. So you split the codes by prefix rather than the source strings. This probably makes sense given the LCP is faster to compute over compressed strings than full strings.
An alternative to this is the very simplistic splitting in FSSTView from LiquidCache: https://github.com/XiangpengHao/liquid-cache/blob/main/dev/design/00-fsst-view.md