Loop issue multiple lag across M4300 stack

I would suggest that you cable the switches up whereby each leaf connects to each spine redundantly, that would give you 4 x 10 Gbps capacity from one leaf to another. Then distribute your LAGs redundantly over the leaf switches and turn on local preference mode.

To clarify, I believe the following reference diagram in the Netgear documentation to be un-optimal. We run this in production on a number of stacks without issues for 4+ years:

My understanding of the leaf-spine (Spine & Leaf) CLOS network design is to ensure each point on the spine and leaf switches has multiple paths to another in the same layer. Spine switches should connect to each leaf, never another spine. Each leaf should also never connect to another leaf, only to the spine switches.

We implemented a stack of 3 x M4300-8F8X (spine) with 4 x M4300-96X (leaf) switches.We setup distributed LAGs from the X96 leaf switches to servers and enabled local preference mode. This results in the 96X forwarding received traffic via the destination's LAG port on the same switch traffic arrived on and stops traffic unnecessarily loading stack uplinks. This works as expected, although Netgear documentation is extremely poor (something along the lines of "enabling local preference mode enables local preference").

If each 96X has 6 x 10G stack members (2 to each 8X8F) there should be 6 equal distance hops between each 96X leaf switch. If server A sends data to server B and both only have 2 x uplinks running LACP (each uplink in a unique 96X) a single steam of data would have 6 possible choices.