Don't pass blocks around, pass CIDs instead

[daviddias]

As noted in: #76

The decision engine buffers the blocks in memory when it informs the message queue that there is one more block to send, this means that if thousands of transfers are being performed, every block will stay in memory till the message gets sent. A better approach would be to only read from disk when the message is going to be sent.

With this, we might loose some perf by having to go read from disk, but that is why I created this issue ipfs/js-ipfs-repo#110 to add a LRU cache on the repo.

[dignifiedquire]

With the perf PR reads from the store are much more efficient amd will only happen when preparing for message sending. I don't believe that the suggested approach here would help regarding performance in any way.

[daviddias]

With the perf PR reads from the store

Ok, this might be new, because it wasn't like that before perf PR, the full blocks were being passed and put in the "MessageQueue" of each peer, waiting for async to pick it up.

I don't believe that the suggested approach here would help regarding performance in any way.

You don't think that putting a LRU cache for reads in IPFS-Repo can help perf? Could you elaborate? Think about all disk reads or even worse, when the store is actually behind a network (i.e. S3)

[dignifiedquire]

I don't know about the LRU cache we would need to benchmark this. I meant changing the way we pass blocks vs cids around at the moment.

[dignifiedquire]

I might be missing something, would be great if you could point to some code in the latest version of the PR where you think this would help.