The problem with sharding is that if you want miners to work separately in each shard (as you must), then the hashrate of the network is divided up by N shards, meaning a double spend attack just got N times easier within each shard.
I just read a bit about Peter Todd's "treechains" (which are described in the text linked by Pmalek) and this proposal doesn't seem to suffer from the problem as miners contribute all to a "timestamp chain" or "main chain", while transactions are collected in sub-chains. However I'll have to read further as I've still not fully understood the proposal. It seems, however, that it needs much deeper changes to Bitcoin's code than "sidechain" proposals.
Merged mined sidechains/shards may also be a solution for that issue, however, the 2 way peg problem needed is still unsolved (with the drivechain proposal as a "contender"), and it may not be really considered "sharding".