Sharding
Splitting a network so each node handles only part of the work, instead of every node processing every transaction.
Execution sharding divides transaction processing across shards, which multiplies throughput but requires cross-shard messaging and breaks the guarantee that any two contracts can interact within a single atomic transaction. Data sharding divides only the duty of publishing and sampling data, leaving execution to rollups, and this is the direction Ethereum took: blobs and data availability sampling are the building blocks of danksharding. Near and Zilliqa implement execution sharding of their own. The central difficulty in any sharded design is keeping each shard secure on its own, since a shard defended by a small subset of validators is easier to corrupt than the whole network.
In practice
Data availability sampling lets a node confirm that a large block was fully published by checking a few random pieces rather than downloading all of it.
The common misunderstanding
Ethereum did not abandon sharding in favor of rollups so much as change what gets sharded, moving from splitting execution to splitting data availability.