Problem
Temporal sampling currently propagates the sampled edge timestamp to the next hop. Some node-time workloads instead require every hop to use the original per-seed cutoff, without imposing temporal ordering between consecutive edges.
For example:
seed cutoff = 10
hop 1: root -> A at time 3
hop 2: A -> B at time 8
Both edges should be eligible because both timestamps are at or before the original cutoff. With monotonically decreasing traversal, the first edge changes the next-hop bound to 3, so the edge at time 8 is excluded.
#5573 proposes per-seed time windows with an end time that remains fixed during traversal. That may overlap with this requirement, but it is unclear whether the changing start time would still enforce cross-hop ordering.
Target behavior
Provide a temporal sampling policy where:
- each seed has its own cutoff;
- the same original cutoff is applied at every hop;
- sampled edge timestamps do not replace that cutoff;
- separate seed labels retain separate cutoffs, including when they reference the same source vertex.
For the example above, a two-hop call should return both edges at times 3 and 8.
This policy should be available for homogeneous and heterogeneous temporal neighbor sampling. It is independent of the neighbor selection strategy: eligible edges may subsequently be sampled uniformly or selected using latest-k.
Problem
Temporal sampling currently propagates the sampled edge timestamp to the next hop. Some node-time workloads instead require every hop to use the original per-seed cutoff, without imposing temporal ordering between consecutive edges.
For example:
Both edges should be eligible because both timestamps are at or before the original cutoff. With monotonically decreasing traversal, the first edge changes the next-hop bound to 3, so the edge at time 8 is excluded.
#5573 proposes per-seed time windows with an end time that remains fixed during traversal. That may overlap with this requirement, but it is unclear whether the changing start time would still enforce cross-hop ordering.
Target behavior
Provide a temporal sampling policy where:
For the example above, a two-hop call should return both edges at times 3 and 8.
This policy should be available for homogeneous and heterogeneous temporal neighbor sampling. It is independent of the neighbor selection strategy: eligible edges may subsequently be sampled uniformly or selected using latest-
k.