The data size aware balancing introduced in v6.0.3+ versions can potentially schedule migrations between shards that are above the optimal data size threshold.

The chunk-based balancing was avoiding that by only scheduling migrations to shards with a sufficiently low number of chunks.

Example

There are 400GB to spread across 4 shards (ideally 100GB per shard) currently distributed in the following way:

• Shard0: 130GB
• Shard1: 120GB
• Shard2: 110GB
• Shard3: 40GB

Next balancing round (currently):

• Shard0 donates to shard3
• Shard1 donates to shard2 (this is the bug -> there is no real need to donate from shard1 to shard2 because they both need to reach 100GB)

Expected next balancing round after fixing this:

• Shard0 donates to shard3

(Problem/solution very similar to SERVER-75481)