Repository navigation
Evaluate icebug-format as the persisted CSR layout #77
Description
Idea
icebug-format stores a graph as flat Parquet files of offsets (indptr) and neighbours (indices), designed so a graph database can mount it directly. It's discussed in lance-format/lance-graph#159. If the persisted CSR file in #72 followed it, LadybugDB might be able to read lance-csr's CSR, and the other way round.
Open question
Check whether icebug-format assumes dense node IDs. lance-csr's CSR stores raw node keys so that node-table writes never invalidate it; a layout that requires dense IDs would conflict with that.
Activity
prrao87 commentedon Oct 30, 2026
Decision: don't adopt icebug-format as the CSR cache layout
The open question has an answer: icebug-format assumes dense node ids. At
be7ca65, the first column of each node table is the primary key, nodes are sorted by it and mapped to dense 0-based CSR indices, andindicesstores those indices (cli.py).lance-csr's CSR stores raw target keys instead, so a node-table write never invalidates the adjacency: only the node index is rebuilt. An icebug file's targets are positions in one particular sorted node table, so any node insert or delete would shift them and force a full CSR rebuild. That breaks the property #72's cache relies on.
i64, sparse or negative allowed)indicesSo #72 keeps its own raw-key layout. Interchange with LadybugDB stays possible as a separate export step that maps keys to dense positions against one pinned node-table version. We can open that as its own issue if someone needs it.
Closing this evaluation; #72 and #80 stay open.