marius-team / marius

Large scale graph learning on a single machine.
https://marius-project.org
Apache License 2.0
160 stars 45 forks source link

Fix bugs that appear with small or empty mini batch neighborhoods #147

Closed rogerwaleffe closed 10 months ago

rogerwaleffe commented 10 months ago

There are a few bugs in the main branch that cause the system to crash when small batches have a few or zero total neighbors. These bugs often appear when running disk-based training, especially as the in-memory subgraph gets sparser with small buffer capacities relative to the number of partitions. This PR fixes these issues.

rogerwaleffe commented 10 months ago

This PR does not affect the build/install which we worked to improve and should be stable (using the docker file instructions) as of #131. We should work to make the "Build and Test" check more stable so that it doesn't fail every new PR as dependencies etc. change, but that is beyond the scope of this PR.