The current DGL random walk implementations require a CSR graph format. For my use case, the CSR format requires a much larger memory footprint than its COO format:
For 68B edges (5 edge types), DGL(COO) is 1099GB cpu memory, and DGL(CSR) is 2470GB cpu memory.
Do you think it’d be possible to implement a random walk implementation that uses the COO format and is also just as performant as the CSR-based implementation? The high CPU memory overhead of CSR is causing complications with serving.